# AI Agent Access Control: every page as markdown AI Agent Access Control maps the identity, default permissions and boundaries of AI agent platforms in AWS, Azure and Google Cloud, with a 30-check checklist. --- # AI agent access control: what an agent can reach in your cloud, and where to draw the line Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. An AI agent in your cloud can reach whatever the identity it runs as can reach: an IAM execution role in AWS, a service account in Google Cloud, a managed identity in Azure. The Agent Access Map covers 8 agent platforms; the checklist has 30 checks. The agent, then the identity it runs as, then the line (policies, guardrails, network endpoints), then what it reaches. ## The Agent Access Map, read 10 Nov 2026 | Agent platform | Cloud | Runs as | By default | Boundaries that apply | | --- | --- | --- | --- | --- | | [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic)](/agent-platforms/bedrock-agents/) | AWS | AWS documents that the agent runs as an IAM service role that Amazon Bedrock assumes. | The documented service role policy covers invoking named foundation models, reading the OpenAPI schema objects in S3, and querying attached knowledge bases, with optional statements for guardrails, collaborators and code interpretation. | IAM identity-based policy on the agent service role; Trust policy conditions (aws:SourceAccount, aws:SourceArn); Resource-based policy on the action group Lambda function; SCPs and permissions boundaries; VPC endpoint policy | | [Amazon Bedrock AgentCore (Runtime, Gateway, Identity)](/agent-platforms/bedrock-agentcore/) | AWS | Runtime runs the agent under an IAM execution role that AgentCore assumes. | AWS documents that the console can create the execution role for you if the console role has the listed IAM create permissions, and that Gateway can create a service role and attach permissions each time you add a target. | IAM policy on the execution role and gateway service role; Resource-based policies on AgentCore resources; SCPs and permissions boundaries; IAM deny of InvokeAgentRuntimeForUser; AgentCore Policy at Gateway; VPC security groups | | [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder)](/agent-platforms/agent-runtime-gcp/) | Google Cloud | Google documents three choices: a per-agent IAM agent identity (recommended, a principal tied to the individual agent); the Google-managed AI Platform Reasoning Engine Service Agent used by default; or a custom service account. | With agent identity, an agent starts with the Agent Default Access role (basic project-wide logging and model calling) and the Agent Context Editor role (own sessions, memories and sandboxes only). | IAM allow policies on the agent identity; IAM deny policy; Principal Access Boundary policy; Context-Aware Access (mTLS binding); VPC Service Controls; Organization policy for cross-project service accounts | | [Gemini Enterprise (agents hosted on Agent Runtime)](/agent-platforms/gemini-enterprise/) | Google Cloud | Gemini Enterprise registers an ADK agent that runs on Agent Runtime, so the agent runs under the Agent Runtime identity described in that entry. | Permissions Gemini Enterprise itself grants to a registered agent: none found on 2026-11-10. | Gemini Enterprise IAM roles; Agent sharing state; VPC Service Controls; Agent Runtime IAM, deny, PAB | | [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service)](/agent-platforms/foundry-agent-service/) | Azure | Each agent gets a Microsoft Entra agent identity, a special service principal created under an agent blueprint. | Microsoft documents that a newly created agent identity starts without resource permissions and that Foundry does not automatically assign RBAC roles to it; you or an administrator grant access through the target service. | Azure RBAC on the agent identity; Foundry RBAC roles and agent-scope assignments; Azure Policy; Virtual network isolation; Hosted agent sandbox | | [Self-hosted agent on AWS Lambda or Amazon EKS](/agent-platforms/self-hosted-aws/) | AWS | A Lambda function runs as its execution role. | When you create a function in the Lambda console, AWS documents that it creates an execution role with minimal permissions (basic logging to CloudWatch Logs). | IAM policy on the execution role or pod role; Pod Identity trust policy conditions; SCPs and permissions boundaries; IMDS restriction | | [Self-hosted agent on Google Cloud Run or GKE](/agent-platforms/self-hosted-gcp/) | Google Cloud | Cloud Run runs the service as a service account (the Compute Engine default service account if you do not choose one). | Google documents that, depending on organization policy, the Compute Engine default service account might automatically be granted the Editor role on the project, and that the constraint iam.automaticIamGrantsForDefaultServiceAccounts is enforced by default for organizations created after May 3, 2024. | IAM roles on the service account; Organization policy iam.automaticIamGrantsForDefaultServiceAccounts; IAM deny policies; Principal Access Boundary policies | | [Self-hosted agent on Azure Container Apps or AKS](/agent-platforms/self-hosted-azure/) | Azure | A container app can have a system-assigned managed identity (tied to the app) or user-assigned identities (standalone resources). | Microsoft documents that for some resources you must configure role assignments for the app's managed identity, otherwise calls to services such as Key Vault and SQL Database are rejected even with a valid token. | Azure RBAC role assignments; Azure Policy; Deny assignments; Kubernetes network policy on AKS | From provider documentation read 10 Nov 2026. Each platform page carries the full entry with its sources. - [The full map, field by field](/map/) - [The Agent Boundary Checklist: 30 checks](/checklist/) - [What an AI agent can access in AWS](/clouds/aws/) - [What an AI agent can access in Google Cloud](/clouds/gcp/) - [What an AI agent can access in Azure](/clouds/azure/) ## Tools scored on AI access coverage | Rank | Tool | Total | Removes (20) | Safe (15) | Class (10) | Fixed (10) | Id+Net+AI (35) | Native (5) | Audit (5) | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | 1 | [Act Security](/tools/act/) | 5.3 | 3 | 7 | 7 | 4 | 6 | 6 | 3 | | 2 | [Sonrai Security](/tools/sonrai/) | 4.8 | 5 | 5 | 7 | 4 | 4 | 7 | 4 | | 3 | [Wiz](/tools/wiz/) | 4.2 | 3 | 4 | 1 | 4 | 6 | 2 | 6 | | 4 | [Google Cloud IAM recommender](/tools/google-iam-recommender/) | 4.0 | 4 | 6 | 3 | 2 | 4 | 5 | 2 | | 5= | [Apono](/tools/apono/) | 3.9 | 5 | 3 | 3 | 3 | 4 | 5 | 4 | | 5= | [Tenable Cloud Security](/tools/tenable/) | 3.9 | 5 | 3 | 0 | 4 | 5 | 0 | 5 | | 7 | [Britive](/tools/britive/) | 3.6 | 5 | 1 | 3 | 3 | 4 | 5 | 4 | | 8= | [CrowdStrike](/tools/crowdstrike/) | 3.3 | 3 | 4 | 0 | 4 | 4 | 0 | 6 | | 8= | [P0 Security](/tools/p0/) | 3.3 | 5 | 0 | 2 | 3 | 4 | 5 | 3 | | 8= | [Prisma Cloud](/tools/prisma/) | 3.3 | 3 | 0 | 1 | 4 | 5 | 3 | 5 | | 11= | [Entitle](/tools/entitle/) | 3.2 | 5 | 0 | 2 | 3 | 4 | 2 | 3 | | 11= | [Sysdig](/tools/sysdig/) | 3.2 | 3 | 2 | 0 | 4 | 4 | 3 | 6 | | 13 | [Microsoft](/tools/microsoft/) | 3.1 | 2 | 0 | 1 | 2 | 6 | 0 | 5 | | 14 | [AWS IAM Access Analyzer](/tools/aws-iam-access-analyzer/) | 2.7 | 3 | 4 | 1 | 4 | 2 | 4 | 2 | Identity, network and AI access together 35, removes access 20, safe to enforce 15, stays fixed 10, one change for a class of risk 10, uses the cloud's own controls 5, evidence for auditors 5. Scores 0 to 10 from the evidence read 10 Nov 2026. Act is first at 5.3, ahead of Sonrai Security at 4.8. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## What would change this ranking Sonrai Security would move ahead of Act Security if its score on identity, network and AI rose by 2 points (from 4 to 6) with nothing else changing. A score changes only when a sourced fact changes, and every change is listed on the corrections page. ## The tools that cover identity, network and AI access best ### Rank 1: Act Security (5.3) Access hardening platform, Safe to enforce 7 Highest score: safe to enforce, 7 of 10. Lowest: removes access, 3. [Read the profile](/tools/act/) ### Rank 2: Sonrai Security (4.8) Cloud permissions firewall, One change, a class of risk 7 Highest score: one change, a class of risk, 7 of 10. Lowest: identity, network and AI, 4. [Read the profile](/tools/sonrai/) ### Rank 3: Wiz (4.2) CNAPP with entitlement management, Identity, network and AI 6 Highest score: identity, network and AI, 6 of 10. Lowest: one change, a class of risk, 1. [Read the profile](/tools/wiz/) ### Rank 4: Google Cloud IAM recommender (4.0) Provider tool, Safe to enforce 6 Highest score: safe to enforce, 6 of 10. Lowest: stays fixed, 2. [Read the profile](/tools/google-iam-recommender/) ## Where each is strongest - **Act Security:** Act removes broad classes of access by simulating guardrail and ring-fence policies against a digital twin and delivering them as pull requests to the customer's IaC using cloud-native controls. - **Sonrai Security:** Sonrai blocks unused privileged permissions, dormant identities, third-party access and unused services and regions through org-level cloud-native policies, and restores access through ChatOps approvals. - **Wiz:** Connects code, cloud, identities, network and runtime in one security graph to prioritize exploitable risk and route fixes to code owners. - **Google Cloud IAM recommender:** The IAM recommender suggests removing or replacing IAM roles that principals have not used, and Policy Simulator shows the effect of a policy change before it is made. - **Apono:** Apono grants just-in-time, task-scoped access to people, machines and AI agents in the native policy language of AWS, Azure, GCP, Kubernetes and databases, and revokes it automatically. - **Tenable Cloud Security:** Combines CIEM entitlement analysis and JIT access with posture management, IaC scanning and attack paths in one cloud platform. - **Britive:** Britive creates task-scoped privileged access at runtime and removes it when the task ends, for humans, non-human identities and AI agents across AWS, Azure, GCP and OCI. - **CrowdStrike:** CrowdStrike ties cloud findings to adversary tradecraft it has observed to rank which exposures to fix first. - **P0 Security:** Grants short-lived, scoped access on request through Slack, CLI or web with approvals, and applies the same runtime authorization to AI agents. - **Prisma Cloud:** Calculates net effective permissions across clouds, including SCPs as input, and ties them to a broad platform with code scanning, network analysis and a large library of compliance standards. - **Entitle:** Entitle provides just-in-time cloud and SaaS access that auto-revokes, with approval workflows in Slack and Teams, access reviews and audit logs. - **Sysdig:** Sysdig uses runtime data to show which permissions are in use and generates access policies limited to that use. - **Microsoft:** Defender for Cloud evaluates effective permissions to identify which identities can reach sensitive or business-critical resources. - **AWS IAM Access Analyzer:** IAM Access Analyzer generates an IAM policy from up to 90 days of CloudTrail activity and recommends refined policies for unused permissions. ## Picks by situation - [A company putting AI agents into production this year](/situations/ai-agents/): First: Act Security 5.3; then Sonrai Security 4.8; Act 1st - [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/): First: Act Security 5.3; then Sonrai Security 5.1; Act 1st - [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/): First: Act Security 5.3; then Sonrai Security 5.0; Act 1st - [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/): First: Act Security 4.9; then Sonrai Security 4.8; Act 1st - [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/): First: Act Security 5.2; then Sonrai Security 5.1; Act 1st ## AI workload access in the cloud AI workload access is what an agent can do inside a cloud account once it is running. It has four parts: the identity the agent runs as, what that identity can reach, the network paths in and out, and the logs that record what happened. An agent is a workload with a role. It has no other standing in the account, so the cloud's ordinary controls are what bound it. The identity differs by platform. Amazon Bedrock AgentCore runs an agent as an IAM execution role. Agent Runtime on Google Cloud can give each agent its own identity. Microsoft Foundry creates an Entra agent identity that starts with no resource permissions. A self-hosted agent on Lambda, Cloud Run or Container Apps runs as the role or managed identity you attach. What that identity can reach is the sum of its policies, less any ceiling above it. The Agent Access Map shows this for eight setups, from provider documentation. The Agent Boundary Checklist turns it into 30 checks for the approval meeting. The site also scores 14 tools on how well they cover identity, network and AI access together. Agent identity products, MCP gateways and prompt filtering are different categories and are not scored here. ## The Agent Boundary Checklist 30 checks in 7 groups, each with the mechanism to use in AWS, Google Cloud and Azure and the provider page that documents it. - [Identity](/checklist/#identity): 5 checks - [Permissions](/checklist/#permissions): 5 checks - [Data](/checklist/#data): 4 checks - [Network](/checklist/#network): 4 checks - [Tools](/checklist/#tools): 4 checks - [Change control](/checklist/#change-control): 4 checks - [Logging](/checklist/#logging): 4 checks ## Guides and glossary - [The identity an AI agent runs as in AWS, Google Cloud and Azure](/guides/the-identity-an-agent-runs-as/) - [Why an AI agent's permissions are not its user's permissions](/guides/why-an-agents-permissions-are-not-its-users/) - [Network reachability for AI agents on AWS, Google Cloud and Azure](/guides/network-reachability-for-agents/) - [Logging what an AI agent did: audit, data and model logs](/guides/logging-what-an-agent-did/) - [AI agent permissions: how to cut an agent's access](/guides/least-privilege-for-ai-agents/) - [Amazon Bedrock agent permissions: roles, trust and what the console creates](/guides/bedrock-agent-permissions/) - [Glossary, 29 terms](/glossary/) - [30 questions answered](/faq/) ## Sources - docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-ident: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 10 Nov 2026) - cloud.google.com/gemini/enterprise/docs/agents-overview: https://cloud.google.com/gemini/enterprise/docs/agents-overview (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 10 Nov 2026) - docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html: https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/ --- # Seven criteria, fixed weights, and coverage of identity, network and AI access carrying the most Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. The site scores 14 tools from 0 to 10 on seven criteria whose weights were fixed on 10 Nov 2026 before any tool was scored. Identity, network and AI access together carries 35 of 100 points, removes access 20. A total is rounded to one decimal, and every score has a published reason. **What these weights produce.** Act is first at 5.3, ahead of Sonrai Security at 4.8. The seven criteria and weights are identity, network and AI access together (35), removes access (20), safe to enforce (15), one change for a class of risk (10), stays fixed (10), native controls (5) and evidence for auditors (5). They were written down on 10 Nov 2026, before any tool was scored, and none has changed since. Coverage carries the most because an agent's reach is the sum of identity, network and AI-service access. Each tool gets an integer from 0 to 10 per criterion, from facts read on the vendor's pages, documentation and, where one exists, an analyst report. A total is the sum of score times weight, divided by the sum of the weights, rounded to one decimal. Equal rounded totals share a rank, and the unrounded total orders them. Five situation pages re-rank the same scores on the mean of these weights and a situation's own weights. The map and the checklist are not scored. They are built from cloud provider documentation read on 10 Nov 2026, and each cell or check carries a quote and a source URL. A capability not found in the pages read is shown as none found on 2026-11-10, which is a statement about the pages and not about the product. ## The seven criteria and the weights this site uses | Criterion | What is scored | Weight | | --- | --- | --- | | Removes access, not only reports it | Does the product end with an enforced change that takes access away, or with a finding for someone else to act on | 20 | | Safe to enforce | Is each change simulated against the real environment before it ships, with the impact shown | 15 | | One change, a class of risk | Boundaries and guardrails that remove many paths at once, as opposed to permission-by-permission clean-up | 10 | | Stays fixed | Prevents drift: violations checked before production, exceptions handled in the team's own tools, with proof | 10 | | Identity, network and AI access together | Does the model of reachability cover all three, including AI workloads and agents | 35 | | Uses the cloud's own controls | Enforcement through native mechanisms the customer already owns, with nothing new in the request path | 5 | | Evidence for auditors | Mapped controls and continuous evidence for the named frameworks | 5 | Weights sum to 100, fixed on 10 Nov 2026 and committed before any vendor was scored. These weights produce the main ranking; situation pages use the mean described below. ## The scale and the rules, in brief Each criterion is scored 0 to 10 from the vendor's evidence: 0 means none found on the read date, 7 is the ceiling for anything resting on the vendor's own material, and 8 to 10 need a source the vendor does not control. Scoring was blind, from evidence with product names removed. The total is the weighted mean on the weights above, to one decimal; equal totals share a rank and are listed by the unrounded total. - Every vendor claim is printed as the vendor's claim, so no score rests on anything but the text of its own pages and named reports. - Microsoft is scored on Defender for Cloud because Entra Permissions Management was retired on 2025-11-01. - The tools are scored on what they show or change about agent reach in a cloud account, and not on prompt or output inspection. ## Situation weights Each situation page ranks on the mean of this site's weights and the situation's own weights, both fixed here before scoring. | Situation | Removes | Safe | Class | Fixed | Id+Net+AI | Native | Audit | | --- | --- | --- | --- | --- | --- | --- | --- | | An AWS-heavy organization with hundreds of accounts and years of accumulated access | 22.5 | 17.5 | 17.5 | 12.5 | 20 | 5 | 5 | | A multi-cloud enterprise with separate teams per cloud | 20 | 15 | 12.5 | 15 | 25 | 10 | 2.5 | | A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53 | 20 | 15 | 10 | 15 | 20 | 5 | 15 | | A company putting AI agents into production this year | 20 | 15 | 10 | 10 | 35 | 5 | 5 | | A small cloud security team buried under posture findings it cannot clear | 25 | 15 | 17.5 | 12.5 | 20 | 5 | 5 | Published weights: the mean of this site's weights and the situation's own weights, both fixed before scoring. ## The ranking these weights produce | Rank | Platform | Total | Removes (20) | Safe (15) | Class (10) | Fixed (10) | Id+Net+AI (35) | Native (5) | Audit (5) | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | 1 | [Act Security](/tools/act/) | 5.3 | 3 | 7 | 7 | 4 | 6 | 6 | 3 | | 2 | [Sonrai Security](/tools/sonrai/) | 4.8 | 5 | 5 | 7 | 4 | 4 | 7 | 4 | | 3 | [Wiz](/tools/wiz/) | 4.2 | 3 | 4 | 1 | 4 | 6 | 2 | 6 | | 4 | [Google Cloud IAM recommender](/tools/google-iam-recommender/) | 4.0 | 4 | 6 | 3 | 2 | 4 | 5 | 2 | | 5= | [Apono](/tools/apono/) | 3.9 | 5 | 3 | 3 | 3 | 4 | 5 | 4 | | 5= | [Tenable Cloud Security](/tools/tenable/) | 3.9 | 5 | 3 | 0 | 4 | 5 | 0 | 5 | | 7 | [Britive](/tools/britive/) | 3.6 | 5 | 1 | 3 | 3 | 4 | 5 | 4 | | 8= | [CrowdStrike](/tools/crowdstrike/) | 3.3 | 3 | 4 | 0 | 4 | 4 | 0 | 6 | | 8= | [P0 Security](/tools/p0/) | 3.3 | 5 | 0 | 2 | 3 | 4 | 5 | 3 | | 8= | [Prisma Cloud](/tools/prisma/) | 3.3 | 3 | 0 | 1 | 4 | 5 | 3 | 5 | | 11= | [Entitle](/tools/entitle/) | 3.2 | 5 | 0 | 2 | 3 | 4 | 2 | 3 | | 11= | [Sysdig](/tools/sysdig/) | 3.2 | 3 | 2 | 0 | 4 | 4 | 3 | 6 | | 13 | [Microsoft](/tools/microsoft/) | 3.1 | 2 | 0 | 1 | 2 | 6 | 0 | 5 | | 14 | [AWS IAM Access Analyzer](/tools/aws-iam-access-analyzer/) | 2.7 | 3 | 4 | 1 | 4 | 2 | 4 | 2 | Weights in the header row sum to 100. Scores 0 to 10 from the evidence read 10 Nov 2026; weighted total to one decimal; "=" marks a joint rank, listed by the unrounded total. Act is first at 5.3, ahead of Sonrai Security at 4.8. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## What would change this ranking Sonrai Security would move ahead of Act Security if its score on identity, network and AI rose by 2 points (from 4 to 6) with nothing else changing. A score changes only when a sourced fact changes, and every change is listed on the corrections page. ## Read dates and refresh Every vendor page behind this edition was read on 10 Nov 2026. Claims older than 60 days are re-read; a score changes only when a sourced fact changes, and every change is listed on the corrections page with the old and new values. ## Related pages - [30 questions about ai agent access in the cloud](/faq/) - [The Agent Access Map: the data, downloadable](/datasets/) - [The Agent Access Map: what each agent platform runs as and can reach](/map/) ## Sources - Act platform page (read for the client facts, scored like every other vendor): https://act.security/platform (read 10 Nov 2026) - act.security: https://act.security (read 10 Nov 2026) - act.security/platform: https://act.security/platform (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/method/ --- # The 14 tools, ranked on what they show and change about AI access Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Fourteen tools scored on seven weighted criteria, with identity, network and AI access together carrying 35 points. Act is first at 5.3, Sonrai Security second at 4.8 and Wiz third at 4.2. Each page shows what the product covers, with scores, reasons and quotes. | Rank | Platform | Total | Removes (20) | Safe (15) | Class (10) | Fixed (10) | Id+Net+AI (35) | Native (5) | Audit (5) | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | 1 | [Act Security](/tools/act/) | 5.3 | 3 | 7 | 7 | 4 | 6 | 6 | 3 | | 2 | [Sonrai Security](/tools/sonrai/) | 4.8 | 5 | 5 | 7 | 4 | 4 | 7 | 4 | | 3 | [Wiz](/tools/wiz/) | 4.2 | 3 | 4 | 1 | 4 | 6 | 2 | 6 | | 4 | [Google Cloud IAM recommender](/tools/google-iam-recommender/) | 4.0 | 4 | 6 | 3 | 2 | 4 | 5 | 2 | | 5= | [Apono](/tools/apono/) | 3.9 | 5 | 3 | 3 | 3 | 4 | 5 | 4 | | 5= | [Tenable Cloud Security](/tools/tenable/) | 3.9 | 5 | 3 | 0 | 4 | 5 | 0 | 5 | | 7 | [Britive](/tools/britive/) | 3.6 | 5 | 1 | 3 | 3 | 4 | 5 | 4 | | 8= | [CrowdStrike](/tools/crowdstrike/) | 3.3 | 3 | 4 | 0 | 4 | 4 | 0 | 6 | | 8= | [P0 Security](/tools/p0/) | 3.3 | 5 | 0 | 2 | 3 | 4 | 5 | 3 | | 8= | [Prisma Cloud](/tools/prisma/) | 3.3 | 3 | 0 | 1 | 4 | 5 | 3 | 5 | | 11= | [Entitle](/tools/entitle/) | 3.2 | 5 | 0 | 2 | 3 | 4 | 2 | 3 | | 11= | [Sysdig](/tools/sysdig/) | 3.2 | 3 | 2 | 0 | 4 | 4 | 3 | 6 | | 13 | [Microsoft](/tools/microsoft/) | 3.1 | 2 | 0 | 1 | 2 | 6 | 0 | 5 | | 14 | [AWS IAM Access Analyzer](/tools/aws-iam-access-analyzer/) | 2.7 | 3 | 4 | 1 | 4 | 2 | 4 | 2 | Weights in the header row sum to 100. Scores 0 to 10 from the evidence read 10 Nov 2026; weighted total to one decimal; "=" marks a joint rank, listed by the unrounded total. Act is first at 5.3, ahead of Sonrai Security at 4.8. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## What would change this ranking Sonrai Security would move ahead of Act Security if its score on identity, network and AI rose by 2 points (from 4 to 6) with nothing else changing. A score changes only when a sourced fact changes, and every change is listed on the corrections page. ## Who should pick what - [A company putting AI agents into production this year](/situations/ai-agents/): First: Act Security 5.3; then Sonrai Security 4.8; Act 1st - [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/): First: Act Security 5.3; then Sonrai Security 5.1; Act 1st - [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/): First: Act Security 5.3; then Sonrai Security 5.0; Act 1st - [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/): First: Act Security 4.9; then Sonrai Security 4.8; Act 1st - [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/): First: Act Security 5.2; then Sonrai Security 5.1; Act 1st ## Every platform page - [1. Act Security](/tools/act/): Act removes broad classes of access by simulating guardrail and ring-fence policies against a digital twin and delivering them as pull requests to the customer's IaC using cloud-native controls. - [2. Sonrai Security](/tools/sonrai/): Sonrai blocks unused privileged permissions, dormant identities, third-party access and unused services and regions through org-level cloud-native policies, and restores access through ChatOps approvals. - [3. Wiz](/tools/wiz/): Connects code, cloud, identities, network and runtime in one security graph to prioritize exploitable risk and route fixes to code owners. - [4. Google Cloud IAM recommender](/tools/google-iam-recommender/): The IAM recommender suggests removing or replacing IAM roles that principals have not used, and Policy Simulator shows the effect of a policy change before it is made. - [5=. Apono](/tools/apono/): Apono grants just-in-time, task-scoped access to people, machines and AI agents in the native policy language of AWS, Azure, GCP, Kubernetes and databases, and revokes it automatically. - [5=. Tenable Cloud Security](/tools/tenable/): Combines CIEM entitlement analysis and JIT access with posture management, IaC scanning and attack paths in one cloud platform. - [7. Britive](/tools/britive/): Britive creates task-scoped privileged access at runtime and removes it when the task ends, for humans, non-human identities and AI agents across AWS, Azure, GCP and OCI. - [8=. CrowdStrike](/tools/crowdstrike/): CrowdStrike ties cloud findings to adversary tradecraft it has observed to rank which exposures to fix first. - [8=. P0 Security](/tools/p0/): Grants short-lived, scoped access on request through Slack, CLI or web with approvals, and applies the same runtime authorization to AI agents. - [8=. Prisma Cloud](/tools/prisma/): Calculates net effective permissions across clouds, including SCPs as input, and ties them to a broad platform with code scanning, network analysis and a large library of compliance standards. - [11=. Entitle](/tools/entitle/): Entitle provides just-in-time cloud and SaaS access that auto-revokes, with approval workflows in Slack and Teams, access reviews and audit logs. - [11=. Sysdig](/tools/sysdig/): Sysdig uses runtime data to show which permissions are in use and generates access policies limited to that use. - [13. Microsoft](/tools/microsoft/): Defender for Cloud evaluates effective permissions to identify which identities can reach sensitive or business-critical resources. - [14. AWS IAM Access Analyzer](/tools/aws-iam-access-analyzer/): IAM Access Analyzer generates an IAM policy from up to 90 days of CloudTrail activity and recommends refined policies for unused permissions. ## Related pages - [Picks by situation](/situations/) - [Guides](/guides/) - [Glossary](/glossary/) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/tools/ --- # Act Security on AI agent access: 1st at 5.3 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Act Security is 1st of 14 on AI access coverage at 5.3 of 10. It scores 6 on identity, network and AI access together, the criterion weighted 35 here, and 3 on removing access. Every score on this page was read from the evidence on 10 Nov 2026 and links to its source. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. Identity, network and AI access together is the criterion that carries 35 of the 100 points here, and it is where Act scores highest: 6 of 10, level with Wiz and Microsoft. Act says it models identity, network and AI access together, and that its AI boundary covers endpoint AI agents and cloud AI agents. SACR's report says Act's Access Map combines identity, network, resource and organizational controls and separates configured, effective and observed access. Act's own blog says Act inventories workloads and AI services, including unknown ones, and works on MCP servers and agent gateways as well as cloud resources. The limits are plain. Act is read-only: it delivers each change as a pull request or ticket and your team applies it, and its free IAM Collect tool downloads policies from AWS accounts only. SACR says customer evidence for Act's AI workload model is not yet as mature as for its broader access control. ## What Act covers of an agent's access ### Identity, network and AI access together: 6 of 10, weight 35 One model combines identity, network, resource and organization controls with configured, effective and observed access, plus endpoint and cloud AI agents. Customer evidence for AI is less mature. - "Identity, network, and AI access are analyzed and enforced as a single model, closing gaps that siloed tools cannot structurally reach." (Act's own material, [read 10 Nov 2026](https://act.security/platform?utm_source=aiagentaccesscontrol.com&utm_medium=referral&utm_campaign=platform)) - "Contain endpoint AI agents, govern the infrastructure they run on, and ring-fence cloud AI agents so they can't reach beyond their intended scope." (Act's own material, [read 10 Nov 2026](https://act.security/platform?utm_source=aiagentaccesscontrol.com&utm_medium=referral&utm_campaign=platform)) ### Removes access, not only reports it: 3 of 10, weight 20 Changes are delivered as infrastructure-as-code pull requests that the customer applies. The product is read-only, and direct enforcement is still under discussion. - "Simulate each policy change first to validate its impact before it reaches production, then apply it through your own cloud-native controls." (Act's own material, [read 10 Nov 2026](https://act.security/?utm_source=aiagentaccesscontrol.com&utm_medium=referral&utm_campaign=platform)) - "Native enforcement : Deploy directly as pull requests to your existing Infrastructure-as-Code (Terraform, CloudFormation) using the cloud-native controls you already own." (Act's own material, [read 10 Nov 2026](https://act.security/platform?utm_source=aiagentaccesscontrol.com&utm_medium=referral&utm_campaign=platform)) ### Safe to enforce: 7 of 10, weight 15 Each change is tested against a digital twin built from cloud configuration data and run in audit mode against audit-log events before enforcement. This is on the vendor's own material only, and rare dependencies may go untested. - "Simulation: Test changes against a digital twin to see their exact impact before enforcing, ensuring nothing breaks." (Act's own material, [read 10 Nov 2026](https://act.security/platform?utm_source=aiagentaccesscontrol.com&utm_medium=referral&utm_campaign=platform)) - "Audit mode : Run changes in audit mode as long as needed to confirm no legitimate access is revoked." (Act's own material, [read 10 Nov 2026](https://act.security/platform?utm_source=aiagentaccesscontrol.com&utm_medium=referral&utm_campaign=platform)) ### Uses the cloud's own controls: 6 of 10, weight 5 Act enforces through named native controls (security groups, NSGs, SCPs, RCPs, private endpoints) delivered via infrastructure as code, with no new infrastructure. For AI agents it also enforces on the agent and the gateway. - "Act enforces boundaries through your cloud's own native controls, using Infrastructure-as-Code. No agents and no new infrastructure to maintain." (Act's own material, [read 10 Nov 2026](https://act.security/platform?utm_source=aiagentaccesscontrol.com&utm_medium=referral&utm_campaign=platform)) - "Act’s differentiation is therefore less about a unique enforcement mechanism, as it relies on the cloud’s native controls" (softwareanalyst.io, counted as vendor material, [read 10 Nov 2026](https://softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement)) ## All seven scores | Criterion | Weight here | Score | Weighted points | | --- | --- | --- | --- | | Removes access, not only reports it | 20 | 3 | 0.60 | | Safe to enforce | 15 | 7 | 1.05 | | One change, a class of risk | 10 | 7 | 0.70 | | Stays fixed | 10 | 4 | 0.40 | | Identity, network and AI access together | 35 | 6 | 2.10 | | Uses the cloud's own controls | 5 | 6 | 0.30 | | Evidence for auditors | 5 | 3 | 0.15 | Weighted total 5.3 (5.30 before rounding) on this site's weights. ## Act in the five situations | Situation | Position | Score | Leader | | --- | --- | --- | --- | | [A company putting AI agents into production this year](/situations/ai-agents/) | 1st | 5.3 | Act Security 5.3 | | [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/) | 1st | 5.3 | Act Security 5.3 | | [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/) | 1st | 5.3 | Act Security 5.3 | | [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/) | 1st | 4.9 | Act Security 4.9 | | [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/) | 1st | 5.2 | Act Security 5.2 | ## Strongest, in one sentence Act removes broad classes of access by simulating guardrail and ring-fence policies against a digital twin and delivering them as pull requests to the customer's IaC using cloud-native controls. ([source](https://act.security/platform?utm_source=aiagentaccesscontrol.com&utm_medium=referral&utm_campaign=platform)) ## Related pages - [Sonrai Security for AI agent access: 2nd of 14 at 4.8](/tools/sonrai/) - [Wiz for AI agent access: 3rd of 14 at 4.2](/tools/wiz/) - [Google Cloud IAM recommender for AI agent access: 4th of 14 at 4.0](/tools/google-iam-recommender/) ## Sources - act.security: https://act.security (read 10 Nov 2026) - act.security/platform: https://act.security/platform (read 10 Nov 2026) - softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement: https://softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement (read 10 Nov 2026) - act.security/resources/ai-agent-security-access-chain: https://act.security/resources/ai-agent-security-access-chain (read 10 Nov 2026) - act.security/resources/zero-trust-ai-agents: https://act.security/resources/zero-trust-ai-agents (read 10 Nov 2026) - raw.githubusercontent.com/act-security/terraform-provider-act/main/README.md: https://raw.githubusercontent.com/act-security/terraform-provider-act/main/README.md (read 10 Nov 2026) - act.security/resources/network-and-identity-each-see-half-your-cloud-your-attack: https://act.security/resources/network-and-identity-each-see-half-your-cloud-your-attacker-sees-both (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/tools/act/ --- # Sonrai Security on AI agent access: 2nd at 4.8 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Sonrai Security is 2nd of 14 on AI access coverage at 4.8 of 10. It scores 4 on identity, network and AI access together, the criterion weighted 35 here, and 5 on removing access. Sonrai blocks unused privileged permissions, dormant identities, third-party access and unused services and regions through org-level cloud-native policies, and restores access through ChatOps approvals. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. Sonrai scores 4 of 10 on identity, network and AI access together. Sonrai's usage analysis covers humans, machines, roles and AI agents, and for each agent identity Sonrai compares the permissions granted with the permissions used. Sonrai says it enforces least privilege for agent identities at the IAM layer with AWS SCPs, IAM conditions and RCPs, so that agent roles are scoped and out-of-scope actions are blocked at the cloud API level. The pages read describe no network reachability model, which is why Sonrai's coverage stops at 4 while Act's reaches 6. What Sonrai adds to an agent review is the ceiling: because Sonrai's organization-level policies block once deployed, the limit on what an agent role can call does not wait for a pull request to be merged. Sonrai is second on this site's main view, at 4.8 against Act's 5.3. ## What Sonrai covers of an agent's access ### Identity, network and AI access together: 4 of 10, weight 35 Identities, including AI agent identities, are enforced at the IAM layer. No network reachability model is described. - "Sonrai monitors how permissions are actually used across humans, machines, roles, and AI agents" (Sonrai's own material, [read 10 Nov 2026](https://sonraisecurity.com/use-cases/least-privilege)) - "Sonrai identifies each agent identity's permissions, compares them against actual usage, and enforces least-privilege policies using cloud-native IAM controls." (Sonrai's own material, [read 10 Nov 2026](https://sonraisecurity.com/use-cases/ai-agent-security)) ### Removes access, not only reports it: 5 of 10, weight 20 Once the customer deploys it, organization-level policies block unused privileges and quarantine dormant identities automatically. Removal of unused permissions is only suggested. - "The Cloud Permissions Firewall continuously analyzes real permission usage across all identities and automatically blocks unused privileges using cloud-native, org-level policies." (Sonrai's own material, [read 10 Nov 2026](https://sonraisecurity.com/platform)) - "Unused permissions are restricted, dormant identities are quarantined, and risky services or regions are blocked" (Sonrai's own material, [read 10 Nov 2026](https://sonraisecurity.com/platform)) ### Safe to enforce: 5 of 10, weight 15 Policies are built from observed usage, in-use permissions are exempted, and the exemption list is shown before deployment. The change is reversible, but it is not described as a simulation of the change. - "Policies are built from your usage data. Anything in use is exempted before it’s written. You see the exemption list first." (Sonrai's own material, [read 10 Nov 2026](https://sonraisecurity.com)) - "All reversible and logged." (Sonrai's own material, [read 10 Nov 2026](https://sonraisecurity.com)) ### Uses the cloud's own controls: 7 of 10, weight 5 Enforcement uses SCPs, RCPs, Azure RBAC and GCP deny policies, with no proxy, broker or connector. This rests on the vendor's own material only. - Enforcement uses SCPs, RCPs, Azure RBAC and GCP deny policies, and the vendor states there is no proxy or broker. (Sonrai's own material, [read 10 Nov 2026](https://sonraisecurity.com)) - Limit: "When the allocated time is over, access is automatically revoked." (Sonrai's own material, [read 10 Nov 2026](https://sonraisecurity.com)) ## All seven scores | Criterion | Weight here | Score | Weighted points | | --- | --- | --- | --- | | Removes access, not only reports it | 20 | 5 | 1.00 | | Safe to enforce | 15 | 5 | 0.75 | | One change, a class of risk | 10 | 7 | 0.70 | | Stays fixed | 10 | 4 | 0.40 | | Identity, network and AI access together | 35 | 4 | 1.40 | | Uses the cloud's own controls | 5 | 7 | 0.35 | | Evidence for auditors | 5 | 4 | 0.20 | Weighted total 4.8 (4.80 before rounding) on this site's weights. ## Sonrai in the five situations | Situation | Position | Score | Leader | | --- | --- | --- | --- | | [A company putting AI agents into production this year](/situations/ai-agents/) | 2nd | 4.8 | Act Security 5.3 | | [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/) | 2nd | 5.1 | Act Security 5.3 | | [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/) | 2nd | 5.0 | Act Security 5.3 | | [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/) | 2nd | 4.8 | Act Security 4.9 | | [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/) | 2nd | 5.1 | Act Security 5.2 | ## Strongest, in one sentence Sonrai blocks unused privileged permissions, dormant identities, third-party access and unused services and regions through org-level cloud-native policies, and restores access through ChatOps approvals. ([source](https://sonraisecurity.com/platform)) ## Related pages - [Wiz for AI agent access: 3rd of 14 at 4.2](/tools/wiz/) - [Google Cloud IAM recommender for AI agent access: 4th of 14 at 4.0](/tools/google-iam-recommender/) - [Apono for AI agent access: joint 5th of 14 at 3.9](/tools/apono/) ## Sources - sonraisecurity.com/platform: https://sonraisecurity.com/platform (read 10 Nov 2026) - sonraisecurity.com: https://sonraisecurity.com (read 10 Nov 2026) - sonraisecurity.com/use-cases/least-privilege: https://sonraisecurity.com/use-cases/least-privilege (read 10 Nov 2026) - sonraisecurity.com/use-cases/ai-agent-security: https://sonraisecurity.com/use-cases/ai-agent-security (read 10 Nov 2026) - sonraisecurity.com/use-cases/just-in-time-access: https://sonraisecurity.com/use-cases/just-in-time-access (read 10 Nov 2026) - sonraisecurity.com/pricing: https://sonraisecurity.com/pricing (read 10 Nov 2026) - sonraisecurity.com/use-cases/third-party-access: https://sonraisecurity.com/use-cases/third-party-access/ (read 10 Nov 2026) - sonraisecurity.com/firewall-slack-integration: https://sonraisecurity.com/firewall-slack-integration/ (read 10 Nov 2026) - sonraisecurity.com/comparisons/wiz: https://sonraisecurity.com/comparisons/wiz/ (read 10 Nov 2026) - sonraisecurity.com/resource/iso27002: https://sonraisecurity.com/resource/iso27002/ (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/tools/sonrai/ --- # Wiz on AI agent access: 3rd at 4.2 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Wiz is 3rd of 14 on AI access coverage at 4.2 of 10. It scores 6 on identity, network and AI access together, the criterion weighted 35 here, and 3 on removing access. Connects code, cloud, identities, network and runtime in one security graph to prioritize exploitable risk and route fixes to code owners. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. Wiz scores 6 of 10 on identity, network and AI access together, level with Act and Microsoft. Wiz models identities and network together with code, cloud and runtime, and it discovers AI models, agents and services. SACR says Wiz's Security Graph connects resources, identities, permissions, vulnerabilities, data, code and runtime signals and identifies attack paths. The reason it stops at 6 is that the AI coverage is inventory, not reachability. For an agent review, Wiz shows that an agent exists, what it is attached to and which attack paths pass through it. Wiz describes its strength as connecting code, cloud, identities, network and runtime in one security graph to prioritize exploitable risk and route fixes to code owners. Wiz is third on this site's main view at 4.2. Wiz now shows as part of Google Cloud on its own site, after an acquisition Google completed on 11 March 2026 according to Wikipedia. ## What Wiz covers of an agent's access ### Identity, network and AI access together: 6 of 10, weight 35 A graph connects identities, permissions, network exposure, workloads and data into reachable attack paths, plus AI model and agent discovery. AI coverage is inventory rather than reachability. - "We connect code, cloud, identities, network, and runtime to model:" (Wiz's own material, [read 10 Nov 2026](https://www.wiz.io)) - "Discover every technology running in your code and cloud including AI models, agents, and services across PaaS platforms, custom deployments, and SaaS AI applications using multiple detection methods." (Wiz's own material, [read 10 Nov 2026](https://www.wiz.io/platform)) ### Removes access, not only reports it: 3 of 10, weight 20 Wiz generates plans, policies and pull requests sent to code owners. A generated fix is not an applied control. - "Automatically generate actionable plans, assign ownership, and execute fixes to drive your teams toward zero critical risk." (Wiz's own material, [read 10 Nov 2026](https://www.wiz.io)) - Limit: "Wiz analyzes cloud entitlements and auto-generates least privilege policies across your cloud, to help teams visualize, detect, prioritize, and remediate identity (IAM) risks." (Wiz's own material, [read 10 Nov 2026](https://www.wiz.io/solutions/ciem)) ### Safe to enforce: 4 of 10, weight 15 Wiz evaluates historical activity to judge disruption before remediation. This is not a simulation of control changes, and infrequent dependencies are not covered. - "The Green Agent can use Security Graph context and historical activity to evaluate whether a proposed remediation could disrupt an underlying application." (softwareanalyst.io, counted as vendor material, [read 10 Nov 2026](https://softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement)) - Limit: "although historical usage cannot establish every infrequent or future dependency." (softwareanalyst.io, counted as vendor material, [read 10 Nov 2026](https://softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement)) ### Uses the cloud's own controls: 2 of 10, weight 5 Wiz abstracts requirements from provider technology. Whether its output is a native preventive control is left open. - Limit: "whether the resulting control prevents the prohibited state or remediates it after creation." (softwareanalyst.io, counted as vendor material, [read 10 Nov 2026](https://softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement)) - "Wiz described an approach centered on abstracting common security requirements from the underlying provider-specific technology." (softwareanalyst.io, counted as vendor material, [read 10 Nov 2026](https://softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement)) ## All seven scores | Criterion | Weight here | Score | Weighted points | | --- | --- | --- | --- | | Removes access, not only reports it | 20 | 3 | 0.60 | | Safe to enforce | 15 | 4 | 0.60 | | One change, a class of risk | 10 | 1 | 0.10 | | Stays fixed | 10 | 4 | 0.40 | | Identity, network and AI access together | 35 | 6 | 2.10 | | Uses the cloud's own controls | 5 | 2 | 0.10 | | Evidence for auditors | 5 | 6 | 0.30 | Weighted total 4.2 (4.20 before rounding) on this site's weights. ## Wiz in the five situations | Situation | Position | Score | Leader | | --- | --- | --- | --- | | [A company putting AI agents into production this year](/situations/ai-agents/) | 3rd | 4.2 | Act Security 5.3 | | [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/) | 5th | 3.7 | Act Security 5.3 | | [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/) | 5th | 3.8 | Act Security 5.3 | | [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/) | 3rd | 4.1 | Act Security 4.9 | | [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/) | joint 5th | 3.6 | Act Security 5.2 | ## Strongest, in one sentence Connects code, cloud, identities, network and runtime in one security graph to prioritize exploitable risk and route fixes to code owners. ([source](https://www.wiz.io)) ## Related pages - [Google Cloud IAM recommender for AI agent access: 4th of 14 at 4.0](/tools/google-iam-recommender/) - [Apono for AI agent access: joint 5th of 14 at 3.9](/tools/apono/) - [Tenable Cloud Security for AI agent access: joint 5th of 14 at 3.9](/tools/tenable/) ## Sources - wiz.io: https://www.wiz.io (read 10 Nov 2026) - wiz.io/solutions/ciem: https://www.wiz.io/solutions/ciem (read 10 Nov 2026) - wiz.io/platform: https://www.wiz.io/platform (read 10 Nov 2026) - wiz.io/platform/wiz-code: https://www.wiz.io/platform/wiz-code (read 10 Nov 2026) - wiz.io/solutions/compliance: https://www.wiz.io/solutions/compliance (read 10 Nov 2026) - wiz.io/platform/wiz-cloud: https://www.wiz.io/platform/wiz-cloud (read 10 Nov 2026) - wiz.io/blog/introducing-azure-least-privilege-enforce-least-privilege-access-for: https://www.wiz.io/blog/introducing-azure-least-privilege-enforce-least-privilege-access-for-azure (read 10 Nov 2026) - softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement: https://softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/tools/wiz/ --- # Google Cloud IAM recommender on AI agent access: 4th at 4.0 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Google Cloud IAM recommender is 4th of 14 on AI access coverage at 4.0 of 10. It scores 4 on identity, network and AI access together, the criterion weighted 35 here, and 4 on removing access. Every score on this page was read from the evidence on 10 Nov 2026 and links to its source. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. Google Cloud IAM recommender scores 4 of 10 on identity, network and AI access together. The IAM recommender's role recommendations are based on the permissions principals use, which is identity. A VPC Service Controls troubleshooter is listed among the access troubleshooters, which touches the network side. No page read describes AI workloads or agents. The IAM recommender suggests removing or replacing IAM roles that principals have not used, and Policy Simulator shows the effect of a policy change before it is made. The checklist uses the recommender to find roles on an agent's service account that the agent never exercises. The IAM recommender covers Google Cloud only, and the pages read do not tie it to agent identities. ## What IAM recommender covers of an agent's access ### Identity, network and AI access together: 4 of 10, weight 35 Identity role recommendations and a VPC Service Controls troubleshooter are provided. No AI workloads are covered. - "Role recommendations help you enforce the principle of least privilege by ensuring that principals have only the permissions that they actually need." (IAM recommender's own material, [read 10 Nov 2026](https://cloud.google.com/policy-intelligence/docs/overview)) - "Policy Troubleshooter for Identity and Access Management VPC Service Controls troubleshooter Policy Troubleshooter for Chrome Enterprise Premium" (IAM recommender's own material, [read 10 Nov 2026](https://cloud.google.com/policy-intelligence/docs/overview)) ### Removes access, not only reports it: 4 of 10, weight 20 A person checks and applies role recommendations in the product. Organization policy recommendations are suggestions only. - Limit: "The IAM recommender does not apply recommendations automatically. Instead, you must review your recommendations and decide whether to apply or dismiss them." (IAM recommender's own material, [read 10 Nov 2026](https://cloud.google.com/policy-intelligence/docs/role-recommendations-overview)) - "You can apply recommendations to replace or remove roles, or dismiss recommendations that you don't want to apply." (IAM recommender's own material, [read 10 Nov 2026](https://cloud.google.com/policy-intelligence/docs/role-recommendations-overview)) ### Safe to enforce: 6 of 10, weight 15 Simulators show how an allow-policy change affects principals' access, and list resources that would violate a proposed organization policy before enforcement. Each simulator covers one policy type, and only Google Cloud. - "Policy Simulator for Identity and Access Management allow policies lets you see how a change to an allow policy might affect a principal's access before you commit to making the change." (IAM recommender's own material, [read 10 Nov 2026](https://cloud.google.com/policy-intelligence/docs/iam-simulator-overview)) - "Policy Simulator provides a list of resources that violate the proposed policy before it is enforced, allowing you to reconfigure those resources, request exceptions, or change the scope of your organization policy" (IAM recommender's own material, [read 10 Nov 2026](https://cloud.google.com/policy-intelligence/docs/overview)) ### Uses the cloud's own controls: 5 of 10, weight 5 A recommendation is an operation that removes a member from native IAM policy bindings. It works in one cloud only. - "action : remove , path : /iamPolicy/bindings//members/" (IAM recommender's own material, [read 10 Nov 2026](https://cloud.google.com/policy-intelligence/docs/role-recommendations-overview)) ## All seven scores | Criterion | Weight here | Score | Weighted points | | --- | --- | --- | --- | | Removes access, not only reports it | 20 | 4 | 0.80 | | Safe to enforce | 15 | 6 | 0.90 | | One change, a class of risk | 10 | 3 | 0.30 | | Stays fixed | 10 | 2 | 0.20 | | Identity, network and AI access together | 35 | 4 | 1.40 | | Uses the cloud's own controls | 5 | 5 | 0.25 | | Evidence for auditors | 5 | 2 | 0.10 | Weighted total 4.0 (3.95 before rounding) on this site's weights. ## IAM recommender in the five situations | Situation | Position | Score | Leader | | --- | --- | --- | --- | | [A company putting AI agents into production this year](/situations/ai-agents/) | 4th | 4.0 | Act Security 5.3 | | [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/) | 3rd | 3.9 | Act Security 5.3 | | [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/) | joint 3rd | 3.9 | Act Security 5.3 | | [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/) | 6th | 3.7 | Act Security 4.9 | | [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/) | 4th | 3.8 | Act Security 5.2 | ## Strongest, in one sentence The IAM recommender suggests removing or replacing IAM roles that principals have not used, and Policy Simulator shows the effect of a policy change before it is made. ([source](https://cloud.google.com/policy-intelligence/docs/role-recommendations-overview)) ## Related pages - [Apono for AI agent access: joint 5th of 14 at 3.9](/tools/apono/) - [Tenable Cloud Security for AI agent access: joint 5th of 14 at 3.9](/tools/tenable/) - [Britive for AI agent access: 7th of 14 at 3.6](/tools/britive/) ## Sources - cloud.google.com/policy-intelligence/docs/role-recommendations-overview: https://cloud.google.com/policy-intelligence/docs/role-recommendations-overview (read 10 Nov 2026) - cloud.google.com/policy-intelligence/docs/organization-policy-recommendations-ov: https://cloud.google.com/policy-intelligence/docs/organization-policy-recommendations-overview (read 10 Nov 2026) - cloud.google.com/policy-intelligence/docs/iam-simulator-overview: https://cloud.google.com/policy-intelligence/docs/iam-simulator-overview (read 10 Nov 2026) - cloud.google.com/policy-intelligence/docs/overview: https://cloud.google.com/policy-intelligence/docs/overview (read 10 Nov 2026) - cloud.google.com/docs/terraform/policy-validation/validate-policies: https://cloud.google.com/docs/terraform/policy-validation/validate-policies (read 10 Nov 2026) - cloud.google.com/recommender/docs/using-api: https://cloud.google.com/recommender/docs/using-api (read 10 Nov 2026) - cloud.google.com/recommender/docs/bq-export/export-recommendations-to-bq: https://cloud.google.com/recommender/docs/bq-export/export-recommendations-to-bq (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/tools/google-iam-recommender/ --- # Apono on AI agent access: joint 5th at 3.9 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Apono is joint 5th of 14 on AI access coverage at 3.9 of 10. It scores 4 on identity, network and AI access together, the criterion weighted 35 here, and 5 on removing access. Every score on this page was read from the evidence on 10 Nov 2026 and links to its source. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. Apono scores 4 of 10 on identity, network and AI access together. The pages read say one Apono platform governs engineers, pipelines, copilots and autonomous agents, so identity and AI agents are covered in one place. Apono's Agent Privilege Guard compares an agent's declared intent with its actual actions at runtime. The pages read do not describe a network reachability model. For an agent review, Apono says it changes what an agent can reach by granting just-in-time, task-scoped access in the native policy language of AWS, Azure, GCP, Kubernetes and databases, and revoking it automatically. The agent's standing permissions can then stay small. A simulation of a change against the environment before it ships: none found on 2026-11-10, and 1Password announced on June 15, 2026 that it was acquiring Apono. ## What Apono covers of an agent's access ### Identity, network and AI access together: 4 of 10, weight 35 Identity and AI agents are governed together, with runtime intent checks. No network reachability model is described. - "Apono governs engineers, automation pipelines, copilots, and autonomous agents on one cloud-native platform" (Apono's own material, [read 10 Nov 2026](https://www.apono.io)) - "checks what an agent declares it will do against what it actually does at runtime" (Apono's own material, [read 10 Nov 2026](https://www.apono.io)) ### Removes access, not only reports it: 5 of 10, weight 20 Just-in-time access is created at request time and revoked automatically. Fixes for standing access (delete or right-size) are code that the user runs. - "Apono's permissions are just-in-time and also ephemeral. Access is automatically revoked when no longer needed." (Apono's own material, [read 10 Nov 2026](https://docs.apono.io/docs/why-choose-apono)) - "Apono creates IAM roles, permissions, and access policies on the fly at request time, scoped to the exact need and in the native policy language of your cloud platform." (Apono's own material, [read 10 Nov 2026](https://www.apono.io)) ### Safe to enforce: 3 of 10, weight 15 A preview of who gets what access is shown before access-flow changes are saved, which is a narrow slice. Remediation actions have no simulation, and some fixes are code the user runs. - "Previews the proposed change. Apono applies only changes that have been previewed. The preview summarizes who receives what access, for how long, and who approves it. It also warns if the change expands access." (Apono's own material, [read 10 Nov 2026](https://docs.apono.io/docs/audits-and-reports/apono-admin-mcp-server.md)) - Limit: "On the Quarantine Principal tab, click Remediate to limit access within the dedicated access flow." (Apono's own material, [read 10 Nov 2026](https://docs.apono.io/docs/getting-started/access-discovery/investigate-and-resolve-overprivileged-access.md)) ### Uses the cloud's own controls: 5 of 10, weight 5 Permissions are written in the native policy languages of AWS, Azure and GCP. A connector mediates grants, and whether a gateway sits in the request path is not established. - "Apono creates permissions at request time in the native policy language of AWS, Azure, GCP, Kubernetes, and your databases." (Apono's own material, [read 10 Nov 2026](https://www.apono.io)) - Limit: "Connectors are the components that mediate between Apono and your resources to sync data from cloud applications and grant and revoke access permissions." (Apono's own material, [read 10 Nov 2026](https://docs.apono.io/docs/getting-started)) ## All seven scores | Criterion | Weight here | Score | Weighted points | | --- | --- | --- | --- | | Removes access, not only reports it | 20 | 5 | 1.00 | | Safe to enforce | 15 | 3 | 0.45 | | One change, a class of risk | 10 | 3 | 0.30 | | Stays fixed | 10 | 3 | 0.30 | | Identity, network and AI access together | 35 | 4 | 1.40 | | Uses the cloud's own controls | 5 | 5 | 0.25 | | Evidence for auditors | 5 | 4 | 0.20 | Weighted total 3.9 (3.90 before rounding) on this site's weights. ## Apono in the five situations | Situation | Position | Score | Leader | | --- | --- | --- | --- | | [A company putting AI agents into production this year](/situations/ai-agents/) | joint 5th | 3.9 | Act Security 5.3 | | [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/) | 4th | 3.8 | Act Security 5.3 | | [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/) | joint 3rd | 3.9 | Act Security 5.3 | | [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/) | 4th | 3.9 | Act Security 4.9 | | [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/) | 3rd | 3.9 | Act Security 5.2 | ## Strongest, in one sentence Apono grants just-in-time, task-scoped access to people, machines and AI agents in the native policy language of AWS, Azure, GCP, Kubernetes and databases, and revokes it automatically. ([source](https://www.apono.io)) ## Related pages - [Tenable Cloud Security for AI agent access: joint 5th of 14 at 3.9](/tools/tenable/) - [Britive for AI agent access: 7th of 14 at 3.6](/tools/britive/) - [CrowdStrike for AI agent access: joint 8th of 14 at 3.3](/tools/crowdstrike/) ## Sources - docs.apono.io/docs/why-choose-apono: https://docs.apono.io/docs/why-choose-apono (read 10 Nov 2026) - apono.io: https://www.apono.io (read 10 Nov 2026) - docs.apono.io/docs/getting-started: https://docs.apono.io/docs/getting-started (read 10 Nov 2026) - docs.apono.io/docs/audits-and-reports/apono-admin-mcp-server.md: https://docs.apono.io/docs/audits-and-reports/apono-admin-mcp-server.md (read 10 Nov 2026) - docs.apono.io/docs/getting-started/access-discovery/investigate-and-resolve-over: https://docs.apono.io/docs/getting-started/access-discovery/investigate-and-resolve-overprivileged-access.md (read 10 Nov 2026) - docs.apono.io/docs/access-flows/manage-access-flows/right-sizing.md: https://docs.apono.io/docs/access-flows/manage-access-flows/right-sizing.md (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/tools/apono/ --- # Tenable Cloud Security on AI agent access: joint 5th at 3.9 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Tenable Cloud Security is joint 5th of 14 on AI access coverage at 3.9 of 10. It scores 5 on identity, network and AI access together, the criterion weighted 35 here, and 5 on removing access. Combines CIEM entitlement analysis and JIT access with posture management, IaC scanning and attack paths in one cloud platform. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. Tenable Cloud Security scores 5 of 10 on identity, network and AI access together. Tenable's entitlement analysis covers identities across AWS, Azure and GCP, and Tenable identifies overprivileged AI identities and exposed training buckets. Network coverage is narrower: remediation priority considers whether a resource is publicly accessible, which is a context flag and not a reachability model. Tenable Cloud Security (formerly Ermetic) combines CIEM entitlement analysis and JIT access with posture management, IaC scanning and attack paths in one cloud platform, and its product page is now titled Tenable One Cloud Exposure. Tenable closed its acquisition of Ermetic in 2023. For an agent review it shows which AI identities hold too much and which data they could expose. A class-wide boundary from Tenable: none found on 2026-11-10. ## What Tenable covers of an agent's access ### Identity, network and AI access together: 5 of 10, weight 35 Identity across AWS, Azure and GCP and overprivileged AI identities are covered. Network coverage is limited to public-accessibility context. - "continuously analyzes entitlements across AWS, Azure, and GCP." (Tenable's own material, [read 10 Nov 2026](https://www.tenable.com/products/tenable-cloud-security)) - "Identify overprivileged AI identities and exposed training buckets." (Tenable's own material, [read 10 Nov 2026](https://www.tenable.com/cloud-security)) ### Removes access, not only reports it: 5 of 10, weight 20 Time-limited just-in-time permissions replace standing privilege. Right-sizing policies are applied by the user, and remediation workflows do not say who applies them. - "Use the Tenable cloud infrastructure and entitlements management (CIEM) engine to identify excessive or toxic entitlements and enforce least-privilege policies." (Tenable's own material, [read 10 Nov 2026](https://www.tenable.com/products/tenable-cloud-security)) - "Automated remediation workflows help security teams right-size permissions at scale without disrupting development velocity." (Tenable's own material, [read 10 Nov 2026](https://www.tenable.com/products/tenable-cloud-security)) ### Safe to enforce: 3 of 10, weight 15 The product shows the diff between a generated policy and the current one. No simulation against the environment is described. - Tenable's blog says the product generates right-sized policies the user can apply and shows the exact difference between the generated policy and the current policy. The page does not describe a simulation against the environment. (Tenable's own material, [read 10 Nov 2026](https://www.tenable.com/blog/least-privilege-policy-automated-analysis-trumps-native-aws-tools)) ### Uses the cloud's own controls: 0 of 10, weight 5 None found on 2026-11-10 after checking 19 pages. None found on 10 Nov 2026 in the pages read for Tenable Cloud Security. ## All seven scores | Criterion | Weight here | Score | Weighted points | | --- | --- | --- | --- | | Removes access, not only reports it | 20 | 5 | 1.00 | | Safe to enforce | 15 | 3 | 0.45 | | One change, a class of risk | 10 | 0 | 0.00 | | Stays fixed | 10 | 4 | 0.40 | | Identity, network and AI access together | 35 | 5 | 1.75 | | Uses the cloud's own controls | 5 | 0 | 0.00 | | Evidence for auditors | 5 | 5 | 0.25 | Weighted total 3.9 (3.85 before rounding) on this site's weights. ## Tenable in the five situations | Situation | Position | Score | Leader | | --- | --- | --- | --- | | [A company putting AI agents into production this year](/situations/ai-agents/) | joint 5th | 3.9 | Act Security 5.3 | | [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/) | 7th | 3.4 | Act Security 5.3 | | [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/) | 7th | 3.4 | Act Security 5.3 | | [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/) | 5th | 3.8 | Act Security 4.9 | | [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/) | 7th | 3.5 | Act Security 5.2 | ## Strongest, in one sentence Combines CIEM entitlement analysis and JIT access with posture management, IaC scanning and attack paths in one cloud platform. ([source](https://www.tenable.com/cloud-security)) ## Related pages - [Britive for AI agent access: 7th of 14 at 3.6](/tools/britive/) - [CrowdStrike for AI agent access: joint 8th of 14 at 3.3](/tools/crowdstrike/) - [P0 Security for AI agent access: joint 8th of 14 at 3.3](/tools/p0/) ## Sources - tenable.com/products/tenable-cloud-security: https://www.tenable.com/products/tenable-cloud-security (read 10 Nov 2026) - tenable.com/cloud-security: https://www.tenable.com/cloud-security (read 10 Nov 2026) - tenable.com/blog/least-privilege-policy-automated-analysis-trumps-native-aws-too: https://www.tenable.com/blog/least-privilege-policy-automated-analysis-trumps-native-aws-tools (read 10 Nov 2026) - tenable.com/blog/the-future-of-cloud-access-management-how-tenable-cloud-securit: https://www.tenable.com/blog/the-future-of-cloud-access-management-how-tenable-cloud-security-redefines-just-in-time-access (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/tools/tenable/ --- # Britive on AI agent access: 7th at 3.6 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Britive is 7th of 14 on AI access coverage at 3.6 of 10. It scores 4 on identity, network and AI access together, the criterion weighted 35 here, and 5 on removing access. Britive creates task-scoped privileged access at runtime and removes it when the task ends, for humans, non-human identities and AI agents across AWS, Azure, GCP and OCI. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. Britive scores 4 of 10 on identity, network and AI access together. Britive's discovery maps human, agentic AI and non-human identities and the privileges they reach across cloud, SaaS, on-prem and AI workflows, which covers identity and AI. The pages read describe no network reachability model, so the picture of what an agent can reach stops at identities and the privileges attached to them. Britive says it creates task-scoped privileged access at runtime and removes it when the task ends, for humans, non-human identities and AI agents across AWS, Azure, GCP and OCI. That changes an agent's reach by keeping standing permissions low and handing out the rest per task. Britive also lists an MCP gateway in its platform, which is a separate product category and outside this site's scoring. ## What Britive covers of an agent's access ### Identity, network and AI access together: 4 of 10, weight 35 Britive maps human, agentic AI and non-human identities and their privileges. No network reachability model is described. - "Map every identity: human, agentic AI, and NHI, and every privilege it can reach, across cloud, SaaS, DaaS, on-prem, and AI workflows." (Britive's own material, [read 10 Nov 2026](https://www.britive.com)) ### Removes access, not only reports it: 5 of 10, weight 20 Just-in-time privilege is created and removed automatically when the task ends, named for AWS, Azure, GCP and OCI. - "Privileged access is created at the moment it's needed and removed when the task ends, automatically, across AWS, Azure, GCP, and OCI." (Britive's own material, [read 10 Nov 2026](https://www.britive.com)) - "Granular least-privilege provisioned at runtime, removed automatically." (Britive's own material, [read 10 Nov 2026](https://www.britive.com)) ### Safe to enforce: 1 of 10, weight 15 Only a policy Validate button and profile-data validation exist. No impact simulation is described. - "Enter the Name and the Policy code in JSON format in the Create Policy page. Click Validate to validate the policy details." (Britive's own material, [read 10 Nov 2026](https://docs.britive.com/docs/profilemgmt-managing-profiles.md)) - "Britive validates the profile policy data and prompts the user for confirmation before skipping the erroneous policies." (Britive's own material, [read 10 Nov 2026](https://docs.britive.com/docs/profilemgmt-managing-profiles.md)) ### Uses the cloud's own controls: 5 of 10, weight 5 The decision is enforced natively at the resource, agentless and proxyless, but the MCP gateway sits in the path of agent calls. - "The decision lands natively at the resource itself." (Britive's own material, [read 10 Nov 2026](https://www.britive.com)) - "Britive's agentless, proxyless architecture breaks that model." (Britive's own material, [read 10 Nov 2026](https://www.britive.com/company/britive-advantages)) ## All seven scores | Criterion | Weight here | Score | Weighted points | | --- | --- | --- | --- | | Removes access, not only reports it | 20 | 5 | 1.00 | | Safe to enforce | 15 | 1 | 0.15 | | One change, a class of risk | 10 | 3 | 0.30 | | Stays fixed | 10 | 3 | 0.30 | | Identity, network and AI access together | 35 | 4 | 1.40 | | Uses the cloud's own controls | 5 | 5 | 0.25 | | Evidence for auditors | 5 | 4 | 0.20 | Weighted total 3.6 (3.60 before rounding) on this site's weights. ## Britive in the five situations | Situation | Position | Score | Leader | | --- | --- | --- | --- | | [A company putting AI agents into production this year](/situations/ai-agents/) | 7th | 3.6 | Act Security 5.3 | | [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/) | 6th | 3.5 | Act Security 5.3 | | [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/) | 6th | 3.6 | Act Security 5.3 | | [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/) | 7th | 3.6 | Act Security 4.9 | | [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/) | joint 5th | 3.6 | Act Security 5.2 | ## Strongest, in one sentence Britive creates task-scoped privileged access at runtime and removes it when the task ends, for humans, non-human identities and AI agents across AWS, Azure, GCP and OCI. ([source](https://www.britive.com)) ## Related pages - [CrowdStrike for AI agent access: joint 8th of 14 at 3.3](/tools/crowdstrike/) - [P0 Security for AI agent access: joint 8th of 14 at 3.3](/tools/p0/) - [Prisma Cloud for AI agent access: joint 8th of 14 at 3.3](/tools/prisma/) ## Sources - britive.com: https://www.britive.com (read 10 Nov 2026) - britive.com/use-cases: https://www.britive.com/use-cases (read 10 Nov 2026) - britive.com/platform: https://www.britive.com/platform (read 10 Nov 2026) - docs.britive.com: https://docs.britive.com (read 10 Nov 2026) - britive.com/company/britive-advantages: https://www.britive.com/company/britive-advantages (read 10 Nov 2026) - docs.britive.com/docs/profilemgmt-managing-profiles.md: https://docs.britive.com/docs/profilemgmt-managing-profiles.md (read 10 Nov 2026) - docs.britive.com/docs/managed-permissions.md: https://docs.britive.com/docs/managed-permissions.md (read 10 Nov 2026) - docs.britive.com/docs/britive-profile-management-for-aws.md: https://docs.britive.com/docs/britive-profile-management-for-aws.md (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/tools/britive/ --- # CrowdStrike on AI agent access: joint 8th at 3.3 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. CrowdStrike is joint 8th of 14 on AI access coverage at 3.3 of 10. It scores 4 on identity, network and AI access together, the criterion weighted 35 here, and 3 on removing access. CrowdStrike ties cloud findings to adversary tradecraft it has observed to rank which exposures to fix first. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. CrowdStrike scores 4 of 10 on identity, network and AI access together. CrowdStrike's Falcon Cloud Security CIEM monitors excess permissions and escalations for each resource, user, group or role. Its attack paths connect misconfigurations, vulnerabilities and excess permissions, and network reachability is not named in the text read. AI-SPM gives agentless visibility across named AI services, and the CNAPP page says posture and runtime defense extend to models, workloads and AI agents. CrowdStrike says it ties cloud findings to adversary tradecraft it has observed, to rank which exposures to fix first. In an agent review that helps decide which over-permissioned agent role to fix before the others. The CIEM pages read describe remediation recommendations, so the finding goes to your team. A class-wide boundary from CrowdStrike: none found on 2026-11-10. ## What CrowdStrike covers of an agent's access ### Identity, network and AI access together: 4 of 10, weight 35 Identity entitlements and AI security posture management across named AI services are covered. Attack paths do not name network reachability. - "Monitor excess permissions, identify suspicious escalations, and conduct thorough reviews of permitted actions for each entity" (CrowdStrike's own material, [read 10 Nov 2026](https://www.crowdstrike.com/en-us/platform/cloud-security/ciem/)) - "Graph-based analysis connects misconfigurations, vulnerabilities, and excess permissions into potential attack paths." (CrowdStrike's own material, [read 10 Nov 2026](https://www.crowdstrike.com/en-us/platform/cloud-security/cspm/)) ### Removes access, not only reports it: 3 of 10, weight 20 CrowdStrike generates least-privilege and resource-based policies and says it removes unwanted access, but recommendations are what is described as output. - "Analyze cloud permissions and generate least-privilege access policies to prevent lateral movement." (CrowdStrike's own material, [read 10 Nov 2026](https://www.crowdstrike.com/en-us/platform/cloud-security/cnapp/)) - Limit: "generate remediation recommendations" (CrowdStrike's own material, [read 10 Nov 2026](https://www.crowdstrike.com/en-us/cybersecurity-101/cloud-security/cloud-infrastructure-entitlement-management-ciem/)) ### Safe to enforce: 4 of 10, weight 15 CrowdStrike states a dry-run simulation of remediation before deployment, without saying what it runs against. - "One-click remediation testing: Simulate remediation tactics to understand outcomes and ensure confidence by performing a dry run prior to deployment." (CrowdStrike's own material, [read 10 Nov 2026](https://www.crowdstrike.com/blog/crowdstrike-extends-ciem-capabilities/)) ### Uses the cloud's own controls: 0 of 10, weight 5 None found on 2026-11-10 after checking 17 pages. None found on 10 Nov 2026 in the pages read for CrowdStrike. ## All seven scores | Criterion | Weight here | Score | Weighted points | | --- | --- | --- | --- | | Removes access, not only reports it | 20 | 3 | 0.60 | | Safe to enforce | 15 | 4 | 0.60 | | One change, a class of risk | 10 | 0 | 0.00 | | Stays fixed | 10 | 4 | 0.40 | | Identity, network and AI access together | 35 | 4 | 1.40 | | Uses the cloud's own controls | 5 | 0 | 0.00 | | Evidence for auditors | 5 | 6 | 0.30 | Weighted total 3.3 (3.30 before rounding) on this site's weights. ## CrowdStrike in the five situations | Situation | Position | Score | Leader | | --- | --- | --- | --- | | [A company putting AI agents into production this year](/situations/ai-agents/) | joint 8th | 3.3 | Act Security 5.3 | | [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/) | 9th | 3.0 | Act Security 5.3 | | [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/) | joint 9th | 3.0 | Act Security 5.3 | | [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/) | 8th | 3.5 | Act Security 4.9 | | [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/) | joint 9th | 3.0 | Act Security 5.2 | ## Strongest, in one sentence CrowdStrike ties cloud findings to adversary tradecraft it has observed to rank which exposures to fix first. ([source](https://www.crowdstrike.com/en-us/platform/cloud-security/cspm/)) ## Related pages - [P0 Security for AI agent access: joint 8th of 14 at 3.3](/tools/p0/) - [Prisma Cloud for AI agent access: joint 8th of 14 at 3.3](/tools/prisma/) - [Entitle for AI agent access: joint 11th of 14 at 3.2](/tools/entitle/) ## Sources - crowdstrike.com/en-us/platform/cloud-security/cnapp: https://www.crowdstrike.com/en-us/platform/cloud-security/cnapp/ (read 10 Nov 2026) - crowdstrike.com/en-us/cybersecurity-101/cloud-security/cloud-infrastructure-enti: https://www.crowdstrike.com/en-us/cybersecurity-101/cloud-security/cloud-infrastructure-entitlement-management-ciem/ (read 10 Nov 2026) - crowdstrike.com/en-us/platform/cloud-security/ciem: https://www.crowdstrike.com/en-us/platform/cloud-security/ciem/ (read 10 Nov 2026) - crowdstrike.com/en-us/platform/cloud-security/cspm: https://www.crowdstrike.com/en-us/platform/cloud-security/cspm/ (read 10 Nov 2026) - crowdstrike.com/en-us/platform/cloud-security: https://www.crowdstrike.com/en-us/platform/cloud-security/ (read 10 Nov 2026) - crowdstrike.com/en-us/platform/cloud-security/ai-spm: https://www.crowdstrike.com/en-us/platform/cloud-security/ai-spm/ (read 10 Nov 2026) - crowdstrike.com/en-us/platform/cloud-security/cloud-compliance: https://www.crowdstrike.com/en-us/platform/cloud-security/cloud-compliance/ (read 10 Nov 2026) - crowdstrike.com/blog/crowdstrike-extends-ciem-capabilities: https://www.crowdstrike.com/blog/crowdstrike-extends-ciem-capabilities/ (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/tools/crowdstrike/ --- # P0 Security on AI agent access: joint 8th at 3.3 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. P0 Security is joint 8th of 14 on AI access coverage at 3.3 of 10. It scores 4 on identity, network and AI access together, the criterion weighted 35 here, and 5 on removing access. Grants short-lived, scoped access on request through Slack, CLI or web with approvals, and applies the same runtime authorization to AI agents. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. P0 Security scores 4 of 10 on identity, network and AI access together. P0's access inventory covers identities and access risks across several clouds and identity systems. For AI agents, P0 verifies the originator and the agent behind each MCP tool call and evaluates policy before each action. P0's connector documentation states that P0 does not fix network reachability: the customer supplies the network path. P0 grants short-lived, scoped access on request through Slack, the CLI or the web with approvals, and says it applies the same runtime authorization to AI agents. For an approval meeting, P0 answers who may get which access, for how long and with whose sign-off. A model of where an agent's traffic can go: none found on 2026-11-10. A simulation of a change before it ships: none found on 2026-11-10. ## What P0 covers of an agent's access ### Identity, network and AI access together: 4 of 10, weight 35 Identities across several clouds and MCP agent calls are covered. The documentation states that network reachability is not addressed. - "Discover production assets, identities, and access risks across AWS, Google Cloud, Kubernetes, Okta, Google Workspace, and Microsoft Azure (beta) with P0's access inventory." (P0's own material, [read 10 Nov 2026](https://docs.p0.dev/readme/access-inventory.md)) - "P0 verifies the originator and the agent behind every MCP tool call, evaluates policy before each action through the self-hosted P0 AI Gateway" (P0's own material, [read 10 Nov 2026](https://docs.p0.dev)) ### Removes access, not only reports it: 5 of 10, weight 20 Just-in-time access is provisioned and revoked automatically, with history. Posture findings get suggested shell commands only. - "Provision only the privilege required for the task, then remove it when the work is done." (P0's own material, [read 10 Nov 2026](https://www.p0.dev/platform)) - "Access requests, together with their justifications, approval, grant, and revocation histories are recorded by P0, and can be exported to your SIEM." (P0's own material, [read 10 Nov 2026](https://docs.p0.dev/readme/just-in-time-access.md)) ### Safe to enforce: 0 of 10, weight 15 None found on 2026-11-10 after checking 20 pages. None found on 10 Nov 2026 in the pages read for P0 Security. ### Uses the cloud's own controls: 5 of 10, weight 5 Default paths are agentless on native IAM, with session-bound roles in AWS and GCP. The AI Gateway sits in the agent data path, and connectors broker just-in-time access. - "P0's default access paths are agentless and driven entirely by your cloud provider's native IAM." (P0's own material, [read 10 Nov 2026](https://docs.p0.dev/readme/connectors.md)) - "when P0 has an IAM integration with the upstream service (AWS and GCP today), enforcement happens inside that service via a session-bound temporary user, role, or policy" (P0's own material, [read 10 Nov 2026](https://docs.p0.dev/readme/agentic-control-plane.md)) ## All seven scores | Criterion | Weight here | Score | Weighted points | | --- | --- | --- | --- | | Removes access, not only reports it | 20 | 5 | 1.00 | | Safe to enforce | 15 | 0 | 0.00 | | One change, a class of risk | 10 | 2 | 0.20 | | Stays fixed | 10 | 3 | 0.30 | | Identity, network and AI access together | 35 | 4 | 1.40 | | Uses the cloud's own controls | 5 | 5 | 0.25 | | Evidence for auditors | 5 | 3 | 0.15 | Weighted total 3.3 (3.30 before rounding) on this site's weights. ## P0 in the five situations | Situation | Position | Score | Leader | | --- | --- | --- | --- | | [A company putting AI agents into production this year](/situations/ai-agents/) | joint 8th | 3.3 | Act Security 5.3 | | [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/) | 8th | 3.1 | Act Security 5.3 | | [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/) | 8th | 3.3 | Act Security 5.3 | | [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/) | joint 10th | 3.2 | Act Security 4.9 | | [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/) | 8th | 3.2 | Act Security 5.2 | ## Strongest, in one sentence Grants short-lived, scoped access on request through Slack, CLI or web with approvals, and applies the same runtime authorization to AI agents. ([source](https://docs.p0.dev)) ## Related pages - [Prisma Cloud for AI agent access: joint 8th of 14 at 3.3](/tools/prisma/) - [Entitle for AI agent access: joint 11th of 14 at 3.2](/tools/entitle/) - [Sysdig for AI agent access: joint 11th of 14 at 3.2](/tools/sysdig/) ## Sources - p0.dev/platform: https://www.p0.dev/platform (read 10 Nov 2026) - docs.p0.dev/readme/just-in-time-access.md: https://docs.p0.dev/readme/just-in-time-access.md (read 10 Nov 2026) - docs.p0.dev/readme/posture.md: https://docs.p0.dev/readme/posture.md (read 10 Nov 2026) - docs.p0.dev/llms.txt: https://docs.p0.dev/llms.txt (read 10 Nov 2026) - docs.p0.dev/readme/access-inventory.md: https://docs.p0.dev/readme/access-inventory.md (read 10 Nov 2026) - docs.p0.dev: https://docs.p0.dev (read 10 Nov 2026) - docs.p0.dev/readme/connectors.md: https://docs.p0.dev/readme/connectors.md (read 10 Nov 2026) - docs.p0.dev/readme/agentic-control-plane.md: https://docs.p0.dev/readme/agentic-control-plane.md (read 10 Nov 2026) - p0.dev/customers: https://www.p0.dev/customers (read 10 Nov 2026) - docs.p0.dev/policy-studio/policy-studio/configure-your-first-access-policy.md: https://docs.p0.dev/policy-studio/policy-studio/configure-your-first-access-policy.md (read 10 Nov 2026) - p0.dev/blog/snowflake-least: https://p0.dev/blog/snowflake-least/ (read 10 Nov 2026) - docs.p0.dev/posture/finding-details.md: https://docs.p0.dev/posture/finding-details.md (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/tools/p0/ --- # Prisma Cloud on AI agent access: joint 8th at 3.3 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Prisma Cloud is joint 8th of 14 on AI access coverage at 3.3 of 10. It scores 5 on identity, network and AI access together, the criterion weighted 35 here, and 3 on removing access. Every score on this page was read from the evidence on 10 Nov 2026 and links to its source. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. Prisma Cloud scores 5 of 10 on identity, network and AI access together. Prisma Cloud calculates effective permissions across cloud providers, and its Cloud Network Analyzer gives network path visibility from source to destination. Palo Alto Networks' Cortex Cloud CIEM page covers identities from human users to AI agents, but AI agents are covered only by a general statement, which holds the score at 5. Palo Alto Networks says Prisma Cloud calculates net effective permissions across clouds, with SCPs as an input, so the permission an agent role really holds is shown after the organization ceiling is applied. The network analyzer is a separate feature, so identity and network are read in two places. Palo Alto Networks markets the next version as Cortex Cloud, and Prisma Cloud pages and docs were still live on 2026-11-10. ## What Prisma Cloud covers of an agent's access ### Identity, network and AI access together: 5 of 10, weight 35 Effective permissions across clouds are covered, with a separate network path analyzer. AI agents are covered only by a general statement. - "Prisma Cloud automatically calculates users' effective permissions across cloud service providers, detects overly permissive access, and suggests corrections to reach least privilege." (Prisma Cloud's own material, [read 10 Nov 2026](https://www.paloaltonetworks.com/prisma/cloud/cloud-infrastructure-entitlement-management)) - "Cloud Network Analyzer (CNA) by Prisma Cloud takes a multi-dimensional approach in identifying overly-exposed assets by providing end-to-end network path visibility" (docs.prismacloud.io, counted as vendor material, [read 10 Nov 2026](https://docs.prismacloud.io/content-collections/administration/network-security/cloud-network-analyzer.md)) ### Removes access, not only reports it: 3 of 10, weight 20 IAM fixes are CLI commands the user runs, or a script the user configures. There is no GCP auto-remediation. - Limit: "You can manually remediate the alerts by copying the AWS, Azure, or GCP CLI commands and then run them in your cloud environment or you can configure a custom python script to automate the remediation steps." (docs.prismacloud.io, counted as vendor material, [read 10 Nov 2026](https://docs.prismacloud.io/content-collections/administration/configure-iam-security/remediate-alerts-for-iam-security.md)) - Limit: "Automatic remediation for GCP IAM alerts is not supported." (docs.prismacloud.io, counted as vendor material, [read 10 Nov 2026](https://docs.prismacloud.io/content-collections/administration/configure-iam-security/remediate-alerts-for-iam-security.md)) ### Safe to enforce: 0 of 10, weight 15 None found on 2026-11-10 after checking 19 pages. None found on 10 Nov 2026 in the pages read for Prisma Cloud. ### Uses the cloud's own controls: 3 of 10, weight 5 Remediation runs as native CLI commands that the user executes, with no proxy. It is manual. - Limit: IAM remediation runs CLI commands (or a script) in the customer's own cloud environment; no proxy in the request path was described on pages read. (docs.prismacloud.io, counted as vendor material, [read 10 Nov 2026](https://docs.prismacloud.io/content-collections/administration/configure-iam-security/remediate-alerts-for-iam-security.md)) ## All seven scores | Criterion | Weight here | Score | Weighted points | | --- | --- | --- | --- | | Removes access, not only reports it | 20 | 3 | 0.60 | | Safe to enforce | 15 | 0 | 0.00 | | One change, a class of risk | 10 | 1 | 0.10 | | Stays fixed | 10 | 4 | 0.40 | | Identity, network and AI access together | 35 | 5 | 1.75 | | Uses the cloud's own controls | 5 | 3 | 0.15 | | Evidence for auditors | 5 | 5 | 0.25 | Weighted total 3.3 (3.25 before rounding) on this site's weights. ## Prisma Cloud in the five situations | Situation | Position | Score | Leader | | --- | --- | --- | --- | | [A company putting AI agents into production this year](/situations/ai-agents/) | joint 8th | 3.3 | Act Security 5.3 | | [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/) | joint 11th | 2.8 | Act Security 5.3 | | [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/) | joint 9th | 3.0 | Act Security 5.3 | | [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/) | joint 10th | 3.2 | Act Security 4.9 | | [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/) | joint 11th | 2.8 | Act Security 5.2 | ## Strongest, in one sentence Calculates net effective permissions across clouds, including SCPs as input, and ties them to a broad platform with code scanning, network analysis and a large library of compliance standards. ([source](https://docs.prismacloud.io/content-collections/administration/configure-iam-security/what-is-prisma-cloud-iam-security.md)) ## Related pages - [Entitle for AI agent access: joint 11th of 14 at 3.2](/tools/entitle/) - [Sysdig for AI agent access: joint 11th of 14 at 3.2](/tools/sysdig/) - [Microsoft for AI agent access: 13th of 14 at 3.1](/tools/microsoft/) ## Sources - docs.prismacloud.io/content-collections/administration/configure-iam-security/re: https://docs.prismacloud.io/content-collections/administration/configure-iam-security/remediate-alerts-for-iam-security.md (read 10 Nov 2026) - paloaltonetworks.com/cortex/cloud/cloud-infrastructure-entitlement-management: https://www.paloaltonetworks.com/cortex/cloud/cloud-infrastructure-entitlement-management (read 10 Nov 2026) - docs.prismacloud.io/content-collections/administration/configure-iam-security/wh: https://docs.prismacloud.io/content-collections/administration/configure-iam-security/what-is-prisma-cloud-iam-security.md (read 10 Nov 2026) - docs.prismacloud.io/content-collections/application-security/risk-management/mon: https://docs.prismacloud.io/content-collections/application-security/risk-management/monitor-and-manage-code-build/pull-request-scan.md (read 10 Nov 2026) - paloaltonetworks.com/prisma/cloud/cloud-infrastructure-entitlement-management: https://www.paloaltonetworks.com/prisma/cloud/cloud-infrastructure-entitlement-management (read 10 Nov 2026) - docs.prismacloud.io/content-collections/administration/network-security/cloud-ne: https://docs.prismacloud.io/content-collections/administration/network-security/cloud-network-analyzer.md (read 10 Nov 2026) - docs.prismacloud.io/content-collections/compliance/compliance-standards.md: https://docs.prismacloud.io/content-collections/compliance/compliance-standards.md (read 10 Nov 2026) - paloaltonetworks.com/customers/michaels-adopts-least-privilege-access-strategy-w: https://www.paloaltonetworks.com/customers/michaels-adopts-least-privilege-access-strategy-with-prisma-cloud-ciem (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/tools/prisma/ --- # Entitle on AI agent access: joint 11th at 3.2 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Entitle is joint 11th of 14 on AI access coverage at 3.2 of 10. It scores 4 on identity, network and AI access together, the criterion weighted 35 here, and 5 on removing access. Entitle provides just-in-time cloud and SaaS access that auto-revokes, with approval workflows in Slack and Teams, access reviews and audit logs. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. Entitle scores 4 of 10 on identity, network and AI access together. The product shows users, identities, integrations, resources and roles, which is identity access. BeyondTrust, which acquired Entitle, states that AI agent functions are kept from sensitive resources without human review. No network reachability model was found in the pages read on 2026-11-10. Entitle's strength, in its own words, is just-in-time cloud and SaaS access that auto-revokes, with approval workflows in Slack and Teams, access reviews and audit logs. For an agent, Entitle means the human check sits in front of a sensitive grant rather than inside the agent's role. A simulation of a change before it ships: none found on 2026-11-10. The definitive agreement with BeyondTrust was announced on April 16, 2024. ## What Entitle covers of an agent's access ### Identity, network and AI access together: 4 of 10, weight 35 Identity access is covered, along with a statement that AI agent functions need a human check for sensitive resources. No network model is described. - "ensuring Agentic AI functions can’t access sensitive resources without human review." (Entitle's own material, [read 10 Nov 2026](https://www.beyondtrust.com/products/entitle)) ### Removes access, not only reports it: 5 of 10, weight 20 Elevated access is revoked automatically after a set duration, a ticket or a rotation, and approvers can take bulk actions on identities. - "Enable controlled, self-serving privileged access that is auto revoked after a set duration, ticket resolution, or on-call rotation." (entitle.io, counted as vendor material, [read 10 Nov 2026](https://www.entitle.io)) - "Reports are automatically generated so you can review entitlements and take bulk actions on identities across your IaaS and SaaS environments." (Entitle's own material, [read 10 Nov 2026](https://www.beyondtrust.com/products/entitle)) ### Safe to enforce: 0 of 10, weight 15 None found on 2026-11-10 after checking 15 pages. None found on 10 Nov 2026 in the pages read for Entitle. ### Uses the cloud's own controls: 2 of 10, weight 5 Access is provisioned through the vendor's own VPC agent, which has outbound access, and an API sidecar. - "First, you install an agent on your virtual private cloud (VPC) using Terraform and Helms Charts." (Entitle's own material, [read 10 Nov 2026](https://www.beyondtrust.com/products/entitle)) - Limit: "The Entitle agent requires outbound network access to function." (Entitle's own material, [read 10 Nov 2026](https://docs.beyondtrust.com/entitle/docs/entitle-agent-network-access-requirements)) ## All seven scores | Criterion | Weight here | Score | Weighted points | | --- | --- | --- | --- | | Removes access, not only reports it | 20 | 5 | 1.00 | | Safe to enforce | 15 | 0 | 0.00 | | One change, a class of risk | 10 | 2 | 0.20 | | Stays fixed | 10 | 3 | 0.30 | | Identity, network and AI access together | 35 | 4 | 1.40 | | Uses the cloud's own controls | 5 | 2 | 0.10 | | Evidence for auditors | 5 | 3 | 0.15 | Weighted total 3.2 (3.15 before rounding) on this site's weights. ## Entitle in the five situations | Situation | Position | Score | Leader | | --- | --- | --- | --- | | [A company putting AI agents into production this year](/situations/ai-agents/) | joint 11th | 3.2 | Act Security 5.3 | | [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/) | 10th | 2.9 | Act Security 5.3 | | [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/) | joint 9th | 3.0 | Act Security 5.3 | | [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/) | 12th | 3.0 | Act Security 4.9 | | [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/) | joint 9th | 3.0 | Act Security 5.2 | ## Strongest, in one sentence Entitle provides just-in-time cloud and SaaS access that auto-revokes, with approval workflows in Slack and Teams, access reviews and audit logs. ([source](https://www.entitle.io)) ## Related pages - [Sysdig for AI agent access: joint 11th of 14 at 3.2](/tools/sysdig/) - [Microsoft for AI agent access: 13th of 14 at 3.1](/tools/microsoft/) - [AWS IAM Access Analyzer for AI agent access: 14th of 14 at 2.7](/tools/aws-iam-access-analyzer/) ## Sources - entitle.io: https://www.entitle.io (read 10 Nov 2026) - beyondtrust.com/products/entitle: https://www.beyondtrust.com/products/entitle (read 10 Nov 2026) - docs.beyondtrust.com/entitle/docs/entitle-agent-network-access-requirements: https://docs.beyondtrust.com/entitle/docs/entitle-agent-network-access-requirements (read 10 Nov 2026) - entitle.io/security: https://www.entitle.io/security (read 10 Nov 2026) - docs.beyondtrust.com/entitle/docs/rules: https://docs.beyondtrust.com/entitle/docs/rules (read 10 Nov 2026) - docs.beyondtrust.com/entitle/docs/birthright-policies: https://docs.beyondtrust.com/entitle/docs/birthright-policies (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/tools/entitle/ --- # Sysdig on AI agent access: joint 11th at 3.2 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Sysdig is joint 11th of 14 on AI access coverage at 3.2 of 10. It scores 4 on identity, network and AI access together, the criterion weighted 35 here, and 3 on removing access. Sysdig uses runtime data to show which permissions are in use and generates access policies limited to that use. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. Sysdig scores 4 of 10 on identity, network and AI access together. Sysdig's CIEM builds a profile of permission usage from cloud audit logs. Its attack path analysis covers vulnerabilities, misconfigurations and excessive permissions, and network reachability is not named in the text read. Sysdig AI Defense covers AI agents on endpoints and in the cloud, with an inventory of agents, MCP servers, skills and plugins. Sysdig says it uses runtime data to show which permissions are in use and generates access policies limited to that use. For an agent, that gives a measured starting point for the narrow role the checklist asks for, drawn from what the agent actually called. The CIEM page does not say Sysdig applies them to the account. A class-wide boundary from Sysdig: none found on 2026-11-10. ## What Sysdig covers of an agent's access ### Identity, network and AI access together: 4 of 10, weight 35 Identity usage comes from audit logs, and an AI agent inventory is included. Attack paths do not name network reachability. - "Sysdig Secure analyzes cloud audit logs of all executed commands in your accounts and correlates them with users, roles, and policies." (Sysdig's own material, [read 10 Nov 2026](https://www.sysdig.com/solutions/cloud-infrastructure-entitlement-management-ciem)) - "Visualize how attackers could move across your environment to exploit vulnerabilities, misconfigurations, and excessive permissions." (Sysdig's own material, [read 10 Nov 2026](https://www.sysdig.com/products/secure)) ### Removes access, not only reports it: 3 of 10, weight 20 Sysdig generates least-privilege policies and infrastructure-as-code fix pull requests. The user validates and applies them. - Limit: "receive suggested least-privilege versions of existing policies, making cleanup fast and low risk." (Sysdig's own material, [read 10 Nov 2026](https://www.sysdig.com/solutions/cloud-infrastructure-entitlement-management-ciem)) - "Sysdig analyzes access patterns to identify unused permissions and automatically generate access policies that restrict privileges to precisely what users need to perform their jobs effectively." (Sysdig's own material, [read 10 Nov 2026](https://www.sysdig.com/solutions/cloud-infrastructure-entitlement-management-ciem)) ### Safe to enforce: 2 of 10, weight 15 Recommendations are claimed to be non-disruptive, and AI agent policies start in alert mode. No simulation step is described. - "With Sysdig’s careful recommendations, teams can confidently tighten IAM policies without disrupting legitimate activity." (Sysdig's own material, [read 10 Nov 2026](https://www.sysdig.com/solutions/cloud-infrastructure-entitlement-management-ciem)) - "Policies begin in alert mode and move to enforcement when a team is ready." (Sysdig's own material, [read 10 Nov 2026](https://www.sysdig.com/press-releases/sysdig-launches-sysdig-ai-defense-bringing-runtime-security-to-every-action-ai-agents-take)) ### Uses the cloud's own controls: 3 of 10, weight 5 For cloud, the output is a generated IAM policy that the user applies. Agent enforcement uses the product's own kernel-level host software. - "It checks every action against the organization’s policy, blocks the ones that break it before they run, and kills the process if one gets past." (Sysdig's own material, [read 10 Nov 2026](https://www.sysdig.com/press-releases/sysdig-launches-sysdig-ai-defense-bringing-runtime-security-to-every-action-ai-agents-take)) - "Optimize permissions with automatically generated policies and total visibility of in-use permissions" (Sysdig's own material, [read 10 Nov 2026](https://www.sysdig.com/solutions/cloud-infrastructure-entitlement-management-ciem)) ## All seven scores | Criterion | Weight here | Score | Weighted points | | --- | --- | --- | --- | | Removes access, not only reports it | 20 | 3 | 0.60 | | Safe to enforce | 15 | 2 | 0.30 | | One change, a class of risk | 10 | 0 | 0.00 | | Stays fixed | 10 | 4 | 0.40 | | Identity, network and AI access together | 35 | 4 | 1.40 | | Uses the cloud's own controls | 5 | 3 | 0.15 | | Evidence for auditors | 5 | 6 | 0.30 | Weighted total 3.2 (3.15 before rounding) on this site's weights. ## Sysdig in the five situations | Situation | Position | Score | Leader | | --- | --- | --- | --- | | [A company putting AI agents into production this year](/situations/ai-agents/) | joint 11th | 3.2 | Act Security 5.3 | | [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/) | joint 11th | 2.8 | Act Security 5.3 | | [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/) | joint 9th | 3.0 | Act Security 5.3 | | [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/) | 9th | 3.4 | Act Security 4.9 | | [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/) | joint 11th | 2.8 | Act Security 5.2 | ## Strongest, in one sentence Sysdig uses runtime data to show which permissions are in use and generates access policies limited to that use. ([source](https://www.sysdig.com/solutions/cloud-infrastructure-entitlement-management-ciem)) ## Related pages - [Microsoft for AI agent access: 13th of 14 at 3.1](/tools/microsoft/) - [AWS IAM Access Analyzer for AI agent access: 14th of 14 at 2.7](/tools/aws-iam-access-analyzer/) - [Act Security for AI agent access: 1st of 14 at 5.3](/tools/act/) ## Sources - sysdig.com/solutions/cloud-infrastructure-entitlement-management-ciem: https://www.sysdig.com/solutions/cloud-infrastructure-entitlement-management-ciem (read 10 Nov 2026) - sysdig.com/about: https://www.sysdig.com/about (read 10 Nov 2026) - sysdig.com/press-releases/sysdig-launches-sysdig-ai-defense-bringing-runtime-sec: https://www.sysdig.com/press-releases/sysdig-launches-sysdig-ai-defense-bringing-runtime-security-to-every-action-ai-agents-take (read 10 Nov 2026) - sysdig.com/solutions/infrastructure-as-code-security: https://www.sysdig.com/solutions/infrastructure-as-code-security (read 10 Nov 2026) - sysdig.com/products/secure: https://www.sysdig.com/products/secure (read 10 Nov 2026) - sysdig.com/solutions/cloud-security-posture-management-cspm: https://www.sysdig.com/solutions/cloud-security-posture-management-cspm (read 10 Nov 2026) - sysdig.com/blog/ciem-security-sysdig-secure: https://www.sysdig.com/blog/ciem-security-sysdig-secure (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/tools/sysdig/ --- # Microsoft on AI agent access: 13th at 3.1 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Microsoft is 13th of 14 on AI access coverage at 3.1 of 10. It scores 6 on identity, network and AI access together, the criterion weighted 35 here, and 2 on removing access. Defender for Cloud evaluates effective permissions to identify which identities can reach sensitive or business-critical resources. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. Microsoft scores 6 of 10 on identity, network and AI access together, level with Act and Wiz. Defender for Cloud's CIEM covers Entra ID, AWS IAM and Google Cloud IAM identities. Microsoft's cloud security graph includes network connections and internet exposure, and its attack path analysis covers AI agents among other resources. Microsoft states the AI coverage only briefly, which is why the score is not higher. This profile is for Defender for Cloud, because Microsoft's page says Entra Permissions Management stopped being available for purchase on April 1, 2025 and was to be retired on November 1, 2025. Defender for Cloud evaluates effective permissions to identify which identities can reach sensitive or business-critical resources. That is a view of reach, and Microsoft describes the CIEM recommendations as guidance for remediation. ## What Microsoft covers of an agent's access ### Identity, network and AI access together: 6 of 10, weight 35 Identity across three clouds, network connections and internet exposure in the security graph, and AI agents in attack paths are covered. AI coverage is stated only briefly. - "Google Cloud IAM users, groups, and service accounts" (Microsoft's own material, [read 10 Nov 2026](https://learn.microsoft.com/en-us/azure/defender-for-cloud/permissions-management)) - "lateral movement possibilities, internet exposure, permissions, network connections, vulnerabilities, and more." (Microsoft's own material, [read 10 Nov 2026](https://learn.microsoft.com/en-us/azure/defender-for-cloud/concept-attack-path)) ### Removes access, not only reports it: 2 of 10, weight 20 Microsoft is scored on cloud infrastructure entitlement management (CIEM) in Defender for Cloud, because Entra Permissions Management was retired on 2025-11-01 and has not been sold since 2025-04-01. Defender for Cloud CIEM gives recommendations described as guidance for remediating, and from 2026-08-06 no longer publishes unused-action details for AWS. - Limit: "Right-sizing permissions for overprovisioned identities based on actual usage" (Microsoft's own material, [read 10 Nov 2026](https://learn.microsoft.com/en-us/azure/defender-for-cloud/permissions-management)) - Limit: "The built-in recommendations provide guidance for remediating inactive identities, overprovisioned permissions, and insecure identity settings." (Microsoft's own material, [read 10 Nov 2026](https://learn.microsoft.com/en-us/azure/defender-for-cloud/permissions-management)) ### Safe to enforce: 0 of 10, weight 15 Microsoft is scored on cloud infrastructure entitlement management (CIEM) in Defender for Cloud, because Entra Permissions Management was retired on 2025-11-01 and has not been sold since 2025-04-01. None found on 2026-11-10 for Defender for Cloud. - Limit: "Select Mode , and then, if you want recommendations to be generated and applied manually, select On-Demand ." (Microsoft's own material, [read 10 Nov 2026](https://learn.microsoft.com/en-us/entra/permissions-management/how-to-create-rule)) ### Uses the cloud's own controls: 0 of 10, weight 5 Microsoft is scored on cloud infrastructure entitlement management (CIEM) in Defender for Cloud, because Entra Permissions Management was retired on 2025-11-01 and has not been sold since 2025-04-01. None found on 2026-11-10 for Defender for Cloud. None found on 10 Nov 2026 in the pages read for Microsoft. ## All seven scores | Criterion | Weight here | Score | Weighted points | | --- | --- | --- | --- | | Removes access, not only reports it | 20 | 2 | 0.40 | | Safe to enforce | 15 | 0 | 0.00 | | One change, a class of risk | 10 | 1 | 0.10 | | Stays fixed | 10 | 2 | 0.20 | | Identity, network and AI access together | 35 | 6 | 2.10 | | Uses the cloud's own controls | 5 | 0 | 0.00 | | Evidence for auditors | 5 | 5 | 0.25 | Weighted total 3.1 (3.05 before rounding) on this site's weights. ## Microsoft in the five situations | Situation | Position | Score | Leader | | --- | --- | --- | --- | | [A company putting AI agents into production this year](/situations/ai-agents/) | 13th | 3.1 | Act Security 5.3 | | [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/) | 14th | 2.3 | Act Security 5.3 | | [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/) | 14th | 2.5 | Act Security 5.3 | | [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/) | joint 13th | 2.8 | Act Security 4.9 | | [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/) | 14th | 2.4 | Act Security 5.2 | ## Strongest, in one sentence Defender for Cloud evaluates effective permissions to identify which identities can reach sensitive or business-critical resources. ([source](https://learn.microsoft.com/en-us/azure/defender-for-cloud/permissions-management)) ## Related pages - [AWS IAM Access Analyzer for AI agent access: 14th of 14 at 2.7](/tools/aws-iam-access-analyzer/) - [Act Security for AI agent access: 1st of 14 at 5.3](/tools/act/) - [Sonrai Security for AI agent access: 2nd of 14 at 4.8](/tools/sonrai/) ## Sources - learn.microsoft.com/en-us/azure/defender-for-cloud/permissions-management: https://learn.microsoft.com/en-us/azure/defender-for-cloud/permissions-management (read 10 Nov 2026) - learn.microsoft.com/en-us/entra/permissions-management/faqs: https://learn.microsoft.com/en-us/entra/permissions-management/faqs (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/defender-for-cloud/defender-for-cloud-introducti: https://learn.microsoft.com/en-us/azure/defender-for-cloud/defender-for-cloud-introduction (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/defender-for-cloud/concept-attack-path: https://learn.microsoft.com/en-us/azure/defender-for-cloud/concept-attack-path (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/defender-for-cloud/regulatory-compliance-dashboa: https://learn.microsoft.com/en-us/azure/defender-for-cloud/regulatory-compliance-dashboard (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/defender-for-cloud/enable-permissions-management: https://learn.microsoft.com/en-us/azure/defender-for-cloud/enable-permissions-management (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/governance/policy/overview: https://learn.microsoft.com/en-us/azure/governance/policy/overview (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/tools/microsoft/ --- # AWS IAM Access Analyzer on AI agent access: 14th at 2.7 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. AWS IAM Access Analyzer is 14th of 14 on AI access coverage at 2.7 of 10. It scores 2 on identity, network and AI access together, the criterion weighted 35 here, and 3 on removing access. IAM Access Analyzer generates an IAM policy from up to 90 days of CloudTrail activity and recommends refined policies for unused permissions. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. AWS IAM Access Analyzer scores 2 of 10 on identity, network and AI access together, the lowest on the page. Access Analyzer's unused access analyzers report unused roles, access keys and passwords, which is identity. The external and internal access analyzers analyze only resource-based policies. No network reachability model or AI workload model is described in the pages read. IAM Access Analyzer is still useful in an agent review, because Access Analyzer does one checklist item well. It generates an IAM policy from up to 90 days of CloudTrail activity and recommends refined policies for unused permissions, so an agent role can be trimmed to what it called. IAM Access Analyzer is an AWS service, so it says nothing about Google Cloud or Azure agents. ## What Access Analyzer covers of an agent's access ### Identity, network and AI access together: 2 of 10, weight 35 Only identity unused access and resource-based policies are covered. No network or AI model is covered. - "The findings highlight unused roles, unused access keys for IAM users, and unused passwords for IAM users." (Access Analyzer's own material, [read 10 Nov 2026](https://docs.aws.amazon.com/IAM/latest/UserGuide/what-is-access-analyzer.html)) - Limit: "It only analyzes resource-based policies." (Access Analyzer's own material, [read 10 Nov 2026](https://docs.aws.amazon.com/IAM/latest/UserGuide/access-analyzer-resources.html)) ### Removes access, not only reports it: 3 of 10, weight 20 Refined policies and delete links are generated, and the user attaches or deletes them. - "For unused roles, access keys, or passwords, IAM Access Analyzer provides quick links in the console to help you delete them." (Access Analyzer's own material, [read 10 Nov 2026](https://aws.amazon.com/iam/access-analyzer/)) - Limit: "For unused permissions, IAM Access Analyzer reviews your existing policies and recommends a refined policy that is tailored to your access activity." (Access Analyzer's own material, [read 10 Nov 2026](https://aws.amazon.com/iam/access-analyzer/)) ### Safe to enforce: 4 of 10, weight 15 The tool previews a policy's effect on public and cross-account access and checks new access against the existing version. It does not simulate the impact of removal, and generated policies omit data events and PassRole. - "policy affects public and cross-account access to your resource before deploying resource" (Access Analyzer's own material, [read 10 Nov 2026](https://docs.aws.amazon.com/IAM/latest/UserGuide/what-is-access-analyzer.html)) - "You can check whether the updated policy grants new access compared to the existing version." (Access Analyzer's own material, [read 10 Nov 2026](https://docs.aws.amazon.com/IAM/latest/UserGuide/access-analyzer-custom-policy-checks.html)) ### Uses the cloud's own controls: 4 of 10, weight 5 It generates native IAM policies from CloudTrail and is not in the request path. It works in one cloud only, and the user applies the result. - "IAM Access Analyzer generates IAM policies based on access activity in your AWS CloudTrail logs." (Access Analyzer's own material, [read 10 Nov 2026](https://docs.aws.amazon.com/IAM/latest/UserGuide/what-is-access-analyzer.html)) ## All seven scores | Criterion | Weight here | Score | Weighted points | | --- | --- | --- | --- | | Removes access, not only reports it | 20 | 3 | 0.60 | | Safe to enforce | 15 | 4 | 0.60 | | One change, a class of risk | 10 | 1 | 0.10 | | Stays fixed | 10 | 4 | 0.40 | | Identity, network and AI access together | 35 | 2 | 0.70 | | Uses the cloud's own controls | 5 | 4 | 0.20 | | Evidence for auditors | 5 | 2 | 0.10 | Weighted total 2.7 (2.70 before rounding) on this site's weights. ## Access Analyzer in the five situations | Situation | Position | Score | Leader | | --- | --- | --- | --- | | [A company putting AI agents into production this year](/situations/ai-agents/) | 14th | 2.7 | Act Security 5.3 | | [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/) | joint 11th | 2.8 | Act Security 5.3 | | [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/) | 13th | 2.9 | Act Security 5.3 | | [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/) | joint 13th | 2.8 | Act Security 4.9 | | [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/) | 13th | 2.7 | Act Security 5.2 | ## Strongest, in one sentence IAM Access Analyzer generates an IAM policy from up to 90 days of CloudTrail activity and recommends refined policies for unused permissions. ([source](https://docs.aws.amazon.com/IAM/latest/UserGuide/access-analyzer-policy-generation.html)) ## Related pages - [Act Security for AI agent access: 1st of 14 at 5.3](/tools/act/) - [Sonrai Security for AI agent access: 2nd of 14 at 4.8](/tools/sonrai/) - [Wiz for AI agent access: 3rd of 14 at 4.2](/tools/wiz/) ## Sources - aws.amazon.com/iam/access-analyzer: https://aws.amazon.com/iam/access-analyzer/ (read 10 Nov 2026) - docs.aws.amazon.com/IAM/latest/UserGuide/what-is-access-analyzer.html: https://docs.aws.amazon.com/IAM/latest/UserGuide/what-is-access-analyzer.html (read 10 Nov 2026) - docs.aws.amazon.com/IAM/latest/UserGuide/access-analyzer-custom-policy-checks.ht: https://docs.aws.amazon.com/IAM/latest/UserGuide/access-analyzer-custom-policy-checks.html (read 10 Nov 2026) - docs.aws.amazon.com/IAM/latest/UserGuide/access-analyzer-policy-generation.html: https://docs.aws.amazon.com/IAM/latest/UserGuide/access-analyzer-policy-generation.html (read 10 Nov 2026) - docs.aws.amazon.com/IAM/latest/UserGuide/access-analyzer-resources.html: https://docs.aws.amazon.com/IAM/latest/UserGuide/access-analyzer-resources.html (read 10 Nov 2026) - aws.amazon.com/iam/features/analyze-access: https://aws.amazon.com/iam/features/analyze-access/ (read 10 Nov 2026) - docs.aws.amazon.com/securityhub/latest/userguide/iam-controls.html: https://docs.aws.amazon.com/securityhub/latest/userguide/iam-controls.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/tools/aws-iam-access-analyzer/ --- # Five situations, five re-weighted rankings Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Five situations re-rank the same scores: AWS at scale, multi-cloud, regulated, AI agents and small team. Act is first on all five. Sonrai Security is second on each, by 0.1 on the regulated and small-team pages and by 0.5 on AI agents. Each page prints its weights. - [A company putting AI agents into production this year](/situations/ai-agents/): First: Act Security 5.3; then Sonrai Security 4.8; Act 1st - [An AWS-heavy organization with hundreds of accounts and years of accumulated access](/situations/aws-scale/): First: Act Security 5.3; then Sonrai Security 5.1; Act 1st - [A multi-cloud enterprise with separate teams per cloud](/situations/multi-cloud/): First: Act Security 5.3; then Sonrai Security 5.0; Act 1st - [A regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53](/situations/regulated/): First: Act Security 4.9; then Sonrai Security 4.8; Act 1st - [A small cloud security team buried under posture findings it cannot clear](/situations/small-team/): First: Act Security 5.2; then Sonrai Security 5.1; Act 1st **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [Guides](/guides/) - [Glossary](/glossary/) - [News](/news/) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/situations/ --- # For a company putting AI agents into production this year, Act is first at 5.3 and Sonrai Security is second at 4.8 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. For a company putting AI agents into production this year, the top three on this situation's weights are Act Security (5.3), Sonrai Security (4.8) and Wiz (4.2). An agent's reach is the sum of identity, network and AI-service access; a model that misses one misses the path. Act is first at 5.3, ahead of Sonrai Security at 4.8. An agent's reach is the sum of identity, network and AI-service access, and a model that misses one misses the path. This page uses the main weights: identity, network and AI coverage 35, removes 20, safe to enforce 15, one change for a class of risk 10, stays fixed 10, native controls 5 and evidence for auditors 5. Act totals 5.300 and Sonrai Security 4.800, the widest gap on any situation page. Act scores 6 on coverage against Sonrai's 4, and its customer evidence for AI boundaries is less mature than for its core access control. Sonrai still leads on removes (5 against 3). Wiz is third at 4.2 with a coverage score of 6, held back because its AI coverage is inventory. ## The top three for this situation ### Rank 1: Act Security (5.3) Access hardening platform, Safe to enforce 7 Highest score: safe to enforce, 7 of 10. Lowest: removes access, 3. [Read the profile](/tools/act/) ### Rank 2: Sonrai Security (4.8) Cloud permissions firewall, One change, a class of risk 7 Highest score: one change, a class of risk, 7 of 10. Lowest: identity, network and AI, 4. [Read the profile](/tools/sonrai/) ### Rank 3: Wiz (4.2) CNAPP with entitlement management, Identity, network and AI 6 Highest score: identity, network and AI, 6 of 10. Lowest: one change, a class of risk, 1. [Read the profile](/tools/wiz/) ## The full order under these weights | Rank | Platform | Total | Removes (20) | Safe (15) | Class (10) | Fixed (10) | Id+Net+AI (35) | Native (5) | Audit (5) | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | 1 | [Act Security](/tools/act/) | 5.3 | 3 | 7 | 7 | 4 | 6 | 6 | 3 | | 2 | [Sonrai Security](/tools/sonrai/) | 4.8 | 5 | 5 | 7 | 4 | 4 | 7 | 4 | | 3 | [Wiz](/tools/wiz/) | 4.2 | 3 | 4 | 1 | 4 | 6 | 2 | 6 | | 4 | [Google Cloud IAM recommender](/tools/google-iam-recommender/) | 4.0 | 4 | 6 | 3 | 2 | 4 | 5 | 2 | | 5= | [Apono](/tools/apono/) | 3.9 | 5 | 3 | 3 | 3 | 4 | 5 | 4 | | 5= | [Tenable Cloud Security](/tools/tenable/) | 3.9 | 5 | 3 | 0 | 4 | 5 | 0 | 5 | | 7 | [Britive](/tools/britive/) | 3.6 | 5 | 1 | 3 | 3 | 4 | 5 | 4 | | 8= | [CrowdStrike](/tools/crowdstrike/) | 3.3 | 3 | 4 | 0 | 4 | 4 | 0 | 6 | | 8= | [P0 Security](/tools/p0/) | 3.3 | 5 | 0 | 2 | 3 | 4 | 5 | 3 | | 8= | [Prisma Cloud](/tools/prisma/) | 3.3 | 3 | 0 | 1 | 4 | 5 | 3 | 5 | | 11= | [Entitle](/tools/entitle/) | 3.2 | 5 | 0 | 2 | 3 | 4 | 2 | 3 | | 11= | [Sysdig](/tools/sysdig/) | 3.2 | 3 | 2 | 0 | 4 | 4 | 3 | 6 | | 13 | [Microsoft](/tools/microsoft/) | 3.1 | 2 | 0 | 1 | 2 | 6 | 0 | 5 | | 14 | [AWS IAM Access Analyzer](/tools/aws-iam-access-analyzer/) | 2.7 | 3 | 4 | 1 | 4 | 2 | 4 | 2 | Published weights for this situation, the mean of this site's weights and the situation's own (id+net+ai 35, removes 20, safe 15, class 10, fixed 10, native 5, audit 5), fixed before scoring. ## What would change this ranking Sonrai Security would move ahead of Act Security if its score on identity, network and AI rose by 2 points (from 4 to 6) with nothing else changing. A score changes only when a sourced fact changes, and every change is listed on the corrections page. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [For an AWS estate with hundreds of accounts, Act is first at 5.3 and Sonrai Security is second at 5.1](/situations/aws-scale/) - [For an agent estate across AWS, Azure and Google Cloud, Act is first at 5.3 and Sonrai Security is second at 5.0](/situations/multi-cloud/) - [For a regulated company that must prove its agent boundaries, Act is first at 4.9 and Sonrai Security is a close second at 4.8](/situations/regulated/) ## Sources - act.security: https://act.security (read 10 Nov 2026) - act.security/platform: https://act.security/platform (read 10 Nov 2026) - softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement: https://softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement (read 10 Nov 2026) - sonraisecurity.com/platform: https://sonraisecurity.com/platform (read 10 Nov 2026) - sonraisecurity.com: https://sonraisecurity.com (read 10 Nov 2026) - sonraisecurity.com/use-cases/least-privilege: https://sonraisecurity.com/use-cases/least-privilege (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/situations/ai-agents/ --- # For an AWS estate with hundreds of accounts, Act is first at 5.3 and Sonrai Security is second at 5.1 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. For an AWS-heavy organization with hundreds of accounts and years of accumulated access, the top three on this situation's weights are Act Security (5.3), Sonrai Security (5.1) and Google Cloud IAM recommender (3.9). Hundreds of accounts make permission-by-permission clean-up impossible; one change must remove a class of paths, safely. Act is first at 5.3, ahead of Sonrai Security at 5.1. Hundreds of accounts make permission-by-permission clean-up impossible, so one change has to remove a class of paths, safely. This page weights removes 22.5, safe to enforce 17.5, one change for a class of risk 17.5, stays fixed 12.5 and identity, network and AI access together 20. Agents multiply the roles in each account, which is why the class weight matters here. Act totals 5.275 before rounding and Sonrai Security 5.075. Sonrai leads on removes (5 against 3) because its organization-level policies block once deployed. Act leads on safe to enforce (7 against 5) and on identity, network and AI coverage (6 against 4), and the two are level on class (7 each). Google Cloud IAM recommender is third at 3.9. ## The top three for this situation ### Rank 1: Act Security (5.3) Access hardening platform, Safe to enforce 7 Highest score: safe to enforce, 7 of 10. Lowest: removes access, 3. [Read the profile](/tools/act/) ### Rank 2: Sonrai Security (5.1) Cloud permissions firewall, One change, a class of risk 7 Highest score: one change, a class of risk, 7 of 10. Lowest: identity, network and AI, 4. [Read the profile](/tools/sonrai/) ### Rank 3: Google Cloud IAM recommender (3.9) Provider tool, Safe to enforce 6 Highest score: safe to enforce, 6 of 10. Lowest: stays fixed, 2. [Read the profile](/tools/google-iam-recommender/) ## The full order under these weights | Rank | Platform | Total | Removes (22.5) | Safe (17.5) | Class (17.5) | Fixed (12.5) | Id+Net+AI (20) | Native (5) | Audit (5) | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | 1 | [Act Security](/tools/act/) | 5.3 | 3 | 7 | 7 | 4 | 6 | 6 | 3 | | 2 | [Sonrai Security](/tools/sonrai/) | 5.1 | 5 | 5 | 7 | 4 | 4 | 7 | 4 | | 3 | [Google Cloud IAM recommender](/tools/google-iam-recommender/) | 3.9 | 4 | 6 | 3 | 2 | 4 | 5 | 2 | | 4 | [Apono](/tools/apono/) | 3.8 | 5 | 3 | 3 | 3 | 4 | 5 | 4 | | 5 | [Wiz](/tools/wiz/) | 3.7 | 3 | 4 | 1 | 4 | 6 | 2 | 6 | | 6 | [Britive](/tools/britive/) | 3.5 | 5 | 1 | 3 | 3 | 4 | 5 | 4 | | 7 | [Tenable Cloud Security](/tools/tenable/) | 3.4 | 5 | 3 | 0 | 4 | 5 | 0 | 5 | | 8 | [P0 Security](/tools/p0/) | 3.1 | 5 | 0 | 2 | 3 | 4 | 5 | 3 | | 9 | [CrowdStrike](/tools/crowdstrike/) | 3.0 | 3 | 4 | 0 | 4 | 4 | 0 | 6 | | 10 | [Entitle](/tools/entitle/) | 2.9 | 5 | 0 | 2 | 3 | 4 | 2 | 3 | | 11= | [Sysdig](/tools/sysdig/) | 2.8 | 3 | 2 | 0 | 4 | 4 | 3 | 6 | | 11= | [AWS IAM Access Analyzer](/tools/aws-iam-access-analyzer/) | 2.8 | 3 | 4 | 1 | 4 | 2 | 4 | 2 | | 11= | [Prisma Cloud](/tools/prisma/) | 2.8 | 3 | 0 | 1 | 4 | 5 | 3 | 5 | | 14 | [Microsoft](/tools/microsoft/) | 2.3 | 2 | 0 | 1 | 2 | 6 | 0 | 5 | Published weights for this situation, the mean of this site's weights and the situation's own (removes 25, safe 20, class 25, fixed 15, id+net+ai 5, native 5, audit 5), fixed before scoring. ## What would change this ranking Sonrai Security would move ahead of Act Security if its score on removes access rose by 1 point (from 5 to 6) with nothing else changing. A score changes only when a sourced fact changes, and every change is listed on the corrections page. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [For an agent estate across AWS, Azure and Google Cloud, Act is first at 5.3 and Sonrai Security is second at 5.0](/situations/multi-cloud/) - [For a regulated company that must prove its agent boundaries, Act is first at 4.9 and Sonrai Security is a close second at 4.8](/situations/regulated/) - [For a small team that cannot clear its findings, Act is first at 5.2 and Sonrai Security is second at 5.1](/situations/small-team/) ## Sources - act.security: https://act.security (read 10 Nov 2026) - act.security/platform: https://act.security/platform (read 10 Nov 2026) - softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement: https://softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement (read 10 Nov 2026) - sonraisecurity.com/platform: https://sonraisecurity.com/platform (read 10 Nov 2026) - sonraisecurity.com: https://sonraisecurity.com (read 10 Nov 2026) - sonraisecurity.com/use-cases/least-privilege: https://sonraisecurity.com/use-cases/least-privilege (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/situations/aws-scale/ --- # For an agent estate across AWS, Azure and Google Cloud, Act is first at 5.3 and Sonrai Security is second at 5.0 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. For a multi-cloud enterprise with separate teams per cloud, the top three on this situation's weights are Act Security (5.3), Sonrai Security (5.0) and Google Cloud IAM recommender (3.9). Separate teams per cloud need changes enforced through each cloud's own controls and exceptions handled in each team's tools. Act is first at 5.3, ahead of Sonrai Security at 5.0. Separate teams per cloud need changes enforced through each cloud's own controls and exceptions handled in each team's tools. This page weights identity, network and AI coverage 25, removes 20, stays fixed 15, safe to enforce 15, one change for a class of risk 12.5, native controls 10 and evidence for auditors 2.5. Agents run as a different identity type in each cloud, so a model that sees all three matters. Act totals 5.300 and Sonrai Security 5.025, so Act is first by 0.275. Sonrai leads on removes (5 against 3) and native controls (7 against 6), worth about 0.50 together at these weights. Act's leads on identity, network and AI coverage (6 against 4) and on safe to enforce (7 against 5) are worth 0.80. Google Cloud IAM recommender is third at 3.9. ## The top three for this situation ### Rank 1: Act Security (5.3) Access hardening platform, Safe to enforce 7 Highest score: safe to enforce, 7 of 10. Lowest: removes access, 3. [Read the profile](/tools/act/) ### Rank 2: Sonrai Security (5.0) Cloud permissions firewall, One change, a class of risk 7 Highest score: one change, a class of risk, 7 of 10. Lowest: identity, network and AI, 4. [Read the profile](/tools/sonrai/) ### Rank 3 (joint): Google Cloud IAM recommender (3.9) Provider tool, Safe to enforce 6 Highest score: safe to enforce, 6 of 10. Lowest: stays fixed, 2. [Read the profile](/tools/google-iam-recommender/) ## The full order under these weights | Rank | Platform | Total | Removes (20) | Safe (15) | Class (12.5) | Fixed (15) | Id+Net+AI (25) | Native (10) | Audit (2.5) | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | 1 | [Act Security](/tools/act/) | 5.3 | 3 | 7 | 7 | 4 | 6 | 6 | 3 | | 2 | [Sonrai Security](/tools/sonrai/) | 5.0 | 5 | 5 | 7 | 4 | 4 | 7 | 4 | | 3= | [Google Cloud IAM recommender](/tools/google-iam-recommender/) | 3.9 | 4 | 6 | 3 | 2 | 4 | 5 | 2 | | 3= | [Apono](/tools/apono/) | 3.9 | 5 | 3 | 3 | 3 | 4 | 5 | 4 | | 5 | [Wiz](/tools/wiz/) | 3.8 | 3 | 4 | 1 | 4 | 6 | 2 | 6 | | 6 | [Britive](/tools/britive/) | 3.6 | 5 | 1 | 3 | 3 | 4 | 5 | 4 | | 7 | [Tenable Cloud Security](/tools/tenable/) | 3.4 | 5 | 3 | 0 | 4 | 5 | 0 | 5 | | 8 | [P0 Security](/tools/p0/) | 3.3 | 5 | 0 | 2 | 3 | 4 | 5 | 3 | | 9= | [Prisma Cloud](/tools/prisma/) | 3.0 | 3 | 0 | 1 | 4 | 5 | 3 | 5 | | 9= | [Entitle](/tools/entitle/) | 3.0 | 5 | 0 | 2 | 3 | 4 | 2 | 3 | | 9= | [CrowdStrike](/tools/crowdstrike/) | 3.0 | 3 | 4 | 0 | 4 | 4 | 0 | 6 | | 9= | [Sysdig](/tools/sysdig/) | 3.0 | 3 | 2 | 0 | 4 | 4 | 3 | 6 | | 13 | [AWS IAM Access Analyzer](/tools/aws-iam-access-analyzer/) | 2.9 | 3 | 4 | 1 | 4 | 2 | 4 | 2 | | 14 | [Microsoft](/tools/microsoft/) | 2.5 | 2 | 0 | 1 | 2 | 6 | 0 | 5 | Published weights for this situation, the mean of this site's weights and the situation's own (removes 20, safe 15, class 15, fixed 20, id+net+ai 15, native 15, audit 0), fixed before scoring. ## What would change this ranking Sonrai Security would move ahead of Act Security if its score on identity, network and AI rose by 2 points (from 4 to 6) with nothing else changing. A score changes only when a sourced fact changes, and every change is listed on the corrections page. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [For a regulated company that must prove its agent boundaries, Act is first at 4.9 and Sonrai Security is a close second at 4.8](/situations/regulated/) - [For a small team that cannot clear its findings, Act is first at 5.2 and Sonrai Security is second at 5.1](/situations/small-team/) - [For a company putting AI agents into production this year, Act is first at 5.3 and Sonrai Security is second at 4.8](/situations/ai-agents/) ## Sources - act.security: https://act.security (read 10 Nov 2026) - act.security/platform: https://act.security/platform (read 10 Nov 2026) - softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement: https://softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement (read 10 Nov 2026) - sonraisecurity.com/platform: https://sonraisecurity.com/platform (read 10 Nov 2026) - sonraisecurity.com: https://sonraisecurity.com (read 10 Nov 2026) - sonraisecurity.com/use-cases/least-privilege: https://sonraisecurity.com/use-cases/least-privilege (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/situations/multi-cloud/ --- # For a regulated company that must prove its agent boundaries, Act is first at 4.9 and Sonrai Security is a close second at 4.8 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. For a regulated company that must prove access boundaries for PCI DSS, HIPAA or NIST 800-53, the top three on this situation's weights are Act Security (4.9), Sonrai Security (4.8) and Wiz (4.1). An auditor asks for the boundary and the continuous evidence that it held. Act is first at 4.9, ahead of Sonrai Security at 4.8. An auditor asks for the boundary and the continuous evidence that it held. This page weights removes 20, identity, network and AI coverage 20, safe to enforce 15, stays fixed 15 and evidence for auditors 15. The situation names PCI DSS, HIPAA and NIST 800-53. For an agent, the evidence is the identity, the ceiling, and the log that ties each action to that identity. The lead is narrow. Act totals 4.900 and Sonrai Security 4.800, a gap of 0.10. Sonrai leads on removes (5 against 3), evidence for auditors (4 against 3) and native controls (7 against 6). Act leads on safe to enforce (7 against 5) and coverage (6 against 4). Wiz is third at 4.1, with 6 on evidence for auditors, the highest on the page and level with Sysdig and CrowdStrike. ## The top three for this situation ### Rank 1: Act Security (4.9) Access hardening platform, Safe to enforce 7 Highest score: safe to enforce, 7 of 10. Lowest: removes access, 3. [Read the profile](/tools/act/) ### Rank 2: Sonrai Security (4.8) Cloud permissions firewall, One change, a class of risk 7 Highest score: one change, a class of risk, 7 of 10. Lowest: identity, network and AI, 4. [Read the profile](/tools/sonrai/) ### Rank 3: Wiz (4.1) CNAPP with entitlement management, Identity, network and AI 6 Highest score: identity, network and AI, 6 of 10. Lowest: one change, a class of risk, 1. [Read the profile](/tools/wiz/) ## The full order under these weights | Rank | Platform | Total | Removes (20) | Safe (15) | Class (10) | Fixed (15) | Id+Net+AI (20) | Native (5) | Audit (15) | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | 1 | [Act Security](/tools/act/) | 4.9 | 3 | 7 | 7 | 4 | 6 | 6 | 3 | | 2 | [Sonrai Security](/tools/sonrai/) | 4.8 | 5 | 5 | 7 | 4 | 4 | 7 | 4 | | 3 | [Wiz](/tools/wiz/) | 4.1 | 3 | 4 | 1 | 4 | 6 | 2 | 6 | | 4 | [Apono](/tools/apono/) | 3.9 | 5 | 3 | 3 | 3 | 4 | 5 | 4 | | 5 | [Tenable Cloud Security](/tools/tenable/) | 3.8 | 5 | 3 | 0 | 4 | 5 | 0 | 5 | | 6 | [Google Cloud IAM recommender](/tools/google-iam-recommender/) | 3.7 | 4 | 6 | 3 | 2 | 4 | 5 | 2 | | 7 | [Britive](/tools/britive/) | 3.6 | 5 | 1 | 3 | 3 | 4 | 5 | 4 | | 8 | [CrowdStrike](/tools/crowdstrike/) | 3.5 | 3 | 4 | 0 | 4 | 4 | 0 | 6 | | 9 | [Sysdig](/tools/sysdig/) | 3.4 | 3 | 2 | 0 | 4 | 4 | 3 | 6 | | 10= | [Prisma Cloud](/tools/prisma/) | 3.2 | 3 | 0 | 1 | 4 | 5 | 3 | 5 | | 10= | [P0 Security](/tools/p0/) | 3.2 | 5 | 0 | 2 | 3 | 4 | 5 | 3 | | 12 | [Entitle](/tools/entitle/) | 3.0 | 5 | 0 | 2 | 3 | 4 | 2 | 3 | | 13= | [AWS IAM Access Analyzer](/tools/aws-iam-access-analyzer/) | 2.8 | 3 | 4 | 1 | 4 | 2 | 4 | 2 | | 13= | [Microsoft](/tools/microsoft/) | 2.8 | 2 | 0 | 1 | 2 | 6 | 0 | 5 | Published weights for this situation, the mean of this site's weights and the situation's own (audit 25, fixed 20, removes 20, safe 15, class 10, id+net+ai 5, native 5), fixed before scoring. ## What would change this ranking Sonrai Security would move ahead of Act Security if its score on removes access rose by 1 point (from 5 to 6) with nothing else changing. A score changes only when a sourced fact changes, and every change is listed on the corrections page. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [For a small team that cannot clear its findings, Act is first at 5.2 and Sonrai Security is second at 5.1](/situations/small-team/) - [For a company putting AI agents into production this year, Act is first at 5.3 and Sonrai Security is second at 4.8](/situations/ai-agents/) - [For an AWS estate with hundreds of accounts, Act is first at 5.3 and Sonrai Security is second at 5.1](/situations/aws-scale/) ## Sources - act.security: https://act.security (read 10 Nov 2026) - act.security/platform: https://act.security/platform (read 10 Nov 2026) - softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement: https://softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement (read 10 Nov 2026) - sonraisecurity.com/platform: https://sonraisecurity.com/platform (read 10 Nov 2026) - sonraisecurity.com: https://sonraisecurity.com (read 10 Nov 2026) - sonraisecurity.com/use-cases/least-privilege: https://sonraisecurity.com/use-cases/least-privilege (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/situations/regulated/ --- # For a small team that cannot clear its findings, Act is first at 5.2 and Sonrai Security is second at 5.1 Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. For a small cloud security team buried under posture findings it cannot clear, the top three on this situation's weights are Act Security (5.2), Sonrai Security (5.1) and Apono (3.9). A team that cannot clear findings needs the product to make the change, many paths at a time, and keep it fixed. Act is first at 5.2, ahead of Sonrai Security at 5.1. A team that cannot clear findings needs the product to make the change, many paths at a time, and keep it fixed. This page weights removes 25, identity, network and AI coverage 20, one change for a class of risk 17.5, safe to enforce 15 and stays fixed 12.5. Removes weighs more here than anywhere else on the site. The gap is 0.10, the same as on the regulated page. Act totals 5.175 and Sonrai Security 5.075. Sonrai applies changes itself, which is worth 0.50 on removes (5 against 3), and Act is read-only and delivers each change as a pull request or ticket for your team to apply. Act's leads on safe to enforce and on coverage are worth 0.70 together. Apono is third at 3.9. ## The top three for this situation ### Rank 1: Act Security (5.2) Access hardening platform, Safe to enforce 7 Highest score: safe to enforce, 7 of 10. Lowest: removes access, 3. [Read the profile](/tools/act/) ### Rank 2: Sonrai Security (5.1) Cloud permissions firewall, One change, a class of risk 7 Highest score: one change, a class of risk, 7 of 10. Lowest: identity, network and AI, 4. [Read the profile](/tools/sonrai/) ### Rank 3: Apono (3.9) Just-in-time access, Removes access 5 Highest score: removes access, 5 of 10. Lowest: safe to enforce, 3. [Read the profile](/tools/apono/) ## The full order under these weights | Rank | Platform | Total | Removes (25) | Safe (15) | Class (17.5) | Fixed (12.5) | Id+Net+AI (20) | Native (5) | Audit (5) | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | 1 | [Act Security](/tools/act/) | 5.2 | 3 | 7 | 7 | 4 | 6 | 6 | 3 | | 2 | [Sonrai Security](/tools/sonrai/) | 5.1 | 5 | 5 | 7 | 4 | 4 | 7 | 4 | | 3 | [Apono](/tools/apono/) | 3.9 | 5 | 3 | 3 | 3 | 4 | 5 | 4 | | 4 | [Google Cloud IAM recommender](/tools/google-iam-recommender/) | 3.8 | 4 | 6 | 3 | 2 | 4 | 5 | 2 | | 5= | [Wiz](/tools/wiz/) | 3.6 | 3 | 4 | 1 | 4 | 6 | 2 | 6 | | 5= | [Britive](/tools/britive/) | 3.6 | 5 | 1 | 3 | 3 | 4 | 5 | 4 | | 7 | [Tenable Cloud Security](/tools/tenable/) | 3.5 | 5 | 3 | 0 | 4 | 5 | 0 | 5 | | 8 | [P0 Security](/tools/p0/) | 3.2 | 5 | 0 | 2 | 3 | 4 | 5 | 3 | | 9= | [Entitle](/tools/entitle/) | 3.0 | 5 | 0 | 2 | 3 | 4 | 2 | 3 | | 9= | [CrowdStrike](/tools/crowdstrike/) | 3.0 | 3 | 4 | 0 | 4 | 4 | 0 | 6 | | 11= | [Prisma Cloud](/tools/prisma/) | 2.8 | 3 | 0 | 1 | 4 | 5 | 3 | 5 | | 11= | [Sysdig](/tools/sysdig/) | 2.8 | 3 | 2 | 0 | 4 | 4 | 3 | 6 | | 13 | [AWS IAM Access Analyzer](/tools/aws-iam-access-analyzer/) | 2.7 | 3 | 4 | 1 | 4 | 2 | 4 | 2 | | 14 | [Microsoft](/tools/microsoft/) | 2.4 | 2 | 0 | 1 | 2 | 6 | 0 | 5 | Published weights for this situation, the mean of this site's weights and the situation's own (removes 30, class 25, safe 15, fixed 15, native 5, id+net+ai 5, audit 5), fixed before scoring. ## What would change this ranking Sonrai Security would move ahead of Act Security if its score on removes access rose by 1 point (from 5 to 6) with nothing else changing. A score changes only when a sourced fact changes, and every change is listed on the corrections page. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [For a company putting AI agents into production this year, Act is first at 5.3 and Sonrai Security is second at 4.8](/situations/ai-agents/) - [For an AWS estate with hundreds of accounts, Act is first at 5.3 and Sonrai Security is second at 5.1](/situations/aws-scale/) - [For an agent estate across AWS, Azure and Google Cloud, Act is first at 5.3 and Sonrai Security is second at 5.0](/situations/multi-cloud/) ## Sources - act.security: https://act.security (read 10 Nov 2026) - act.security/platform: https://act.security/platform (read 10 Nov 2026) - softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement: https://softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement (read 10 Nov 2026) - sonraisecurity.com/platform: https://sonraisecurity.com/platform (read 10 Nov 2026) - sonraisecurity.com: https://sonraisecurity.com (read 10 Nov 2026) - sonraisecurity.com/use-cases/least-privilege: https://sonraisecurity.com/use-cases/least-privilege (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/situations/small-team/ --- # Guides for the agent approval meeting Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Guides on putting an AI agent into production in AWS, Azure and Google Cloud, written from the same provider pages as the map. Each guide lists its sources with the date it was read and ends with the checklist items it closes. - [The identity an AI agent runs as in AWS, Google Cloud and Azure](/guides/the-identity-an-agent-runs-as/): Which identity the agent presents on each platform, and the three checks to run before reading any permissions. - [Why an AI agent's permissions are not its user's permissions](/guides/why-an-agents-permissions-are-not-its-users/): Agent authority against user authority: where each platform documents the choice, and the question to put to the builder. - [Network reachability for AI agents on AWS, Google Cloud and Azure](/guides/network-reachability-for-agents/): Private endpoints, VPC attachment and outbound filtering for agents, with each provider's documented defaults. - [Logging what an AI agent did: audit, data and model logs](/guides/logging-what-an-agent-did/): Control-plane, data and agent-trace logs on each cloud: what is on by default, what is not, and what names the agent. - [AI agent permissions: how to cut an agent's access](/guides/least-privilege-for-ai-agents/): Write the policy from real tool calls, set a ceiling, trim with usage data, and test before enforcing. - [Amazon Bedrock agent permissions: roles, trust and what the console creates](/guides/bedrock-agent-permissions/): Service role, execution role, trust policy and what the console creates for Agents Classic and AgentCore. - [Vertex AI agent service account: Agent Runtime identities in Google Cloud](/guides/vertex-ai-agent-service-account/): Agent Runtime formerly Vertex AI Agent Engine: agent identity, service agent and service account compared. - [What can an AI agent access in AWS: the four layers to read](/guides/what-can-an-ai-agent-access-in-aws/): The four layers to read to answer what an agent can reach in AWS, in the order to read them. - [Approving an AI agent for production: running the review meeting](/guides/approving-an-agent-for-production/): How to run the review meeting with the Agent Boundary Checklist: pre-read, agenda, scoring and the decisions to record. - [AI agent permissions in Azure: Foundry Agent Service, identities and RBAC](/guides/ai-agent-permissions-in-azure/): Foundry Agent Service identities, managed identities and the RBAC roles that decide what an Azure agent can reach. - [AI agent authorization in AWS, Azure and Google Cloud](/ai-agent-authorization/): The three inputs to an authorization decision for an agent, in each cloud, with the provider's own wording. - [AI agent guardrails: SCPs, Azure Policy and org policy](/ai-agent-guardrails/): The access-layer guardrails for an agent in each cloud, and the one layer this site leaves out. - [AI access control in cloud accounts: roles, policy, network](/ai-access-control/): Access control for AI workloads in cloud accounts, layer by layer, without the model-side controls. - [Agent authorization patterns: five ways to scope an agent](/agent-authorization-patterns/): Five patterns, each with the mechanism and the provider page that documents it. - [Amazon Q Business default permissions: roles it creates](/default-permissions/amazon-q-business/): The roles the Amazon Q Business console creates, what each reaches, and how to narrow them. - [Copilot Studio default permissions: tool authentication](/default-permissions/copilot-studio/): Which credential a Copilot Studio agent's tools use, what that reaches, and the admin controls that narrow it. - [Azure OpenAI Assistants default permissions and access](/default-permissions/azure-openai-assistants/): What an Azure OpenAI assistant can reach by default, who can call it, and how to narrow it before migrating. - [Conversational Agents (Dialogflow CX) default permissions](/default-permissions/conversational-agents/): Who can reach a Dialogflow CX agent by default, and which identity its tools and webhooks use. ## Related pages - [Glossary](/glossary/) - [News](/news/) - [Agent platforms in AWS, Google Cloud and Azure](/agent-platforms/) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/guides/ --- # An agent has no access of its own: it runs as a role, a service account or an agent identity, and that sets its reach Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. An agent has no access of its own. Every call it makes is judged against the identity it runs as: an IAM role on AWS, a service account or agent identity on Google Cloud, an Entra agent identity or managed identity on Azure. This guide names that identity for each platform and what it starts with. ## The cloud judges the identity, not the agent A cloud account does not know that a piece of code is an AI agent. It sees a call signed by a role, a service account or a service principal, and it checks that identity's permissions. So the first question in a review is which identity the agent presents, because every later answer follows from it: what the agent can read, what it can delete, what the logs say. The same agent code can run as a tightly scoped identity in one deployment and as a broad shared one in another. Nothing in the prompt or the tool list shows which. You find out by reading the platform's identity settings, and [the map](/map/) records the answer for each platform covered here. Read the identity first, then the permissions, then everything else. ## On AWS the agent runs as an IAM role that a service assumes For [Amazon Bedrock Agents Classic](/agent-platforms/bedrock-agents/), AWS documents that the agent runs as an IAM service role. You can let Bedrock create the role or supply your own, and the trust policy names the bedrock.amazonaws.com service principal. AWS also documents that you can no longer sign up for Agents Classic. On [Amazon Bedrock AgentCore](/agent-platforms/bedrock-agentcore/), Runtime needs an IAM execution role that AgentCore assumes to run the agent, and the console can create that role when the console role holds the listed IAM create permissions. A self-hosted agent on [AWS Lambda or EKS](/agent-platforms/self-hosted-aws/) runs as whatever the compute gives it. A Lambda function runs as its execution role, and AWS documents that creating a function in the console creates a role with minimal permissions, basic logging to CloudWatch Logs. On EKS, a Pod Identity association maps a role to a service account in a namespace in a cluster, so two agents in one cluster can hold different roles. If access to the instance metadata service is not restricted, AWS documents that a pod's containers can also reach the node's IAM role. ## On Google Cloud the choice is an agent identity, a service agent or a service account Google documents three choices for an agent on [Agent Runtime](/agent-platforms/agent-runtime-gcp/): a per-agent IAM agent identity, which Google recommends; the Google-managed AI Platform Reasoning Engine Service Agent, used by default; or a custom service account. Google describes agent identity as tied to the lifecycle of the agent, and service accounts as shared across the agents you deploy with them. In a review the difference is plain: with a shared account, one agent's grants are every agent's grants. Google creates and manages service agents, and they are not listed by default on the IAM page of the console. A self-hosted agent on [Cloud Run or GKE](/agent-platforms/self-hosted-gcp/) uses a service account. Cloud Run runs the service as a service account and falls back to the Compute Engine default service account if you choose none. Google documents that, depending on organization policy, that default account might automatically be granted the Editor role on the project. On GKE, Workload Identity Federation for GKE assigns a distinct IAM identity to each application in the cluster, so an agent can be named in a policy on its own. ## On Azure the agent runs as an Entra agent identity or a managed identity For [Foundry Agent Service](/agent-platforms/foundry-agent-service/), Microsoft documents that each agent gets a Microsoft Entra agent identity, a special service principal created under an agent blueprint. Hosted agents get a dedicated identity at deploy time. A project-wide managed identity also exists, and Microsoft documents that legacy agents sharing it have no per-agent separation. Microsoft also documents that a newly created agent identity starts without resource permissions and that Foundry does not automatically assign RBAC roles to it, so its starting access is nothing until someone grants more. A self-hosted agent on [Container Apps or AKS](/agent-platforms/self-hosted-azure/) uses a managed identity. A system-assigned identity is tied to the container app and deleted with it, while a user-assigned identity is a standalone resource. On AKS, Microsoft Entra Workload ID uses a projected Kubernetes service account token and OIDC federation to give pods an Entra identity. Microsoft notes that for some resources you must still configure role assignments for the identity, or calls to services such as Key Vault are rejected even with a valid token. ## Three checks settle the identity before the permissions review Run three checks first. Each agent should have its own identity ([check](/checklist/own-identity-per-agent/)). No agent should run as a default or pre-granted account ([check](/checklist/no-default-accounts/)). The identity should be removed when the agent is retired ([check](/checklist/retire-identity-with-agent/)). The checklist records two gaps in the documentation we read: a way to tie the identity's lifetime to the agent on AWS, and a default-account warning for Azure, were both none found on 2026-11-10. Record those as open questions for the platform team, not as passes. All 30 checks are on [the checklist](/checklist/). ## Related pages - [Why an AI agent's permissions are not its user's permissions](/guides/why-an-agents-permissions-are-not-its-users/) - [Network reachability for AI agents on AWS, Google Cloud and Azure](/guides/network-reachability-for-agents/) - [Logging what an AI agent did: audit, data and model logs](/guides/logging-what-an-agent-did/) - [All of them](/guides/) ## Sources - AWS: create a service role for Amazon Bedrock Agents: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html (read 10 Nov 2026) - AWS: IAM permissions for AgentCore Runtime: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html (read 10 Nov 2026) - AWS: Lambda execution role: https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html (read 10 Nov 2026) - Google Cloud: agent identity on Agent Runtime: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 10 Nov 2026) - Google Cloud: Cloud Run service identity: https://cloud.google.com/run/docs/securing/service-identity (read 10 Nov 2026) - Microsoft: agent identity in Foundry Agent Service: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/guides/the-identity-an-agent-runs-as/ --- # An agent acts with its own authority unless the platform is set up to carry the user's, so users can reach more than they hold Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. A user who talks to an agent does not lend it their access. The agent calls tools with its own identity unless the platform exchanges tokens to act for the user. This guide shows how AgentCore, Agent Runtime and Foundry separate the two, and what to ask in a review. ## Two kinds of authority sit behind one chat window A person signs in, asks an agent for something, and the agent calls a tool. Whose permissions does the tool call carry? There are two answers. In the first, the tool runs with the agent's own authority, so every user of the agent can reach whatever the agent can reach, whatever their own rights. In the second, the tool runs with the user's authority, so the agent can do only what that user may. Both are legitimate, and most agents mix them tool by tool. The mistake is not choosing. A review has to ask which one each tool uses, and the answer has to be written down. ## AgentCore Gateway offers a different identity for each target On [Amazon Bedrock AgentCore](/agent-platforms/bedrock-agentcore/), AWS documents that outbound authorization is set per gateway target. A target can sign requests with the gateway service role, use the caller's IAM credentials, run the client credentials grant (machine to machine, acting on the application's behalf), run the authorization code grant (user-delegated, with consent), or use a token exchange. AWS describes the token exchange as swapping the inbound user's access token for a new, scoped token that carries both the user's identity and the agent's identity. The identity of a tool call therefore depends on which option each target uses, so one gateway can hold targets on both sides of the line. AWS also documents a sharp edge in the broad managed policy. BedrockAgentCoreFullAccess grants GetWorkloadAccessTokenForUserId, which issues workload access tokens from a caller-supplied user identifier string without verifying a token from an identity provider. AWS recommends a custom policy with only the permissions the application needs. The point for a review is that a policy that broad lets the caller choose whose identity to claim, so the people who can call that API deserve the same scrutiny as the agent. ## Google and Microsoft name the same split Google documents three identities for [Agent Runtime](/agent-platforms/agent-runtime-gcp/): the user identity, which sets caller context at ingress; the agent identity, which authenticates the workload inside the runtime; and a delegated identity that combines both when the agent acts for the user with scoped 3-legged OAuth. Calls to Google Cloud APIs made with agent identity are evaluated against the agent identity. Microsoft documents the same pair for [Foundry Agent Service](/agent-platforms/foundry-agent-service/): application-only access, where the agent acts under its own authority, and delegated access, where it acts for the signed-in user and needs both the agent's authorization and the user's access. Microsoft labels the delegated mode as preview. Both mean the review must ask for the agent's mode per tool, not per agent. ## Self-hosted agents get no user propagation from the platform For agents on Lambda, EKS, Cloud Run, GKE, Container Apps or AKS, the workload uses its own role, service account or managed identity. We found no documented mechanism in the Lambda, EKS, Cloud Run, GKE, Container Apps or AKS pages read on 2026-11-10 that carries an end user's permissions to the workload, so any such behavior is written in the builder's code. That is a reason to ask for it in writing: how does the code decide which user a call is for, and what stops it from using the workload's broader role instead? Compare the [self-hosted AWS](/agent-platforms/self-hosted-aws/), [Google Cloud](/agent-platforms/self-hosted-gcp/) and [Azure](/agent-platforms/self-hosted-azure/) entries. ## A worked case shows what changes between the two modes Take a support agent with one tool that reads customer records. If the tool runs on the agent's authority, every employee who can open the chat can read any record that the agent's role allows, whatever their own rights. The audit log may name only the agent's identity. If the tool runs on the user's authority, a support trainee gets the records the trainee could already open, and the log can show both identities on platforms that carry them. The second mode costs more to build, since a token has to be exchanged and verified. The first mode costs less to build and more to review, since the role must be narrow enough for the least trusted user. Neither mode fixes a broad role. On the agent's authority, the role is the limit. On the user's authority, the user's own access is the limit, and many users hold more than they need. The review should therefore record the mode for each tool and also say which role bounds it. Do not let the second mode stand in for the first: it moves the question from the agent's role to the user's role, and both need an owner. ## Put one question to the builder for every tool The checklist item is [decide for each tool whether it runs with the agent's authority or the user's](/checklist/decide-user-or-agent-authority/). Ask the builder for a list of tools with one of two answers beside each. For tools on the agent's authority, check that the agent's role is narrow enough for every user to hold it ([check](/checklist/least-privilege-policy/)). For tools on the user's authority, ask where the user's token comes from and who verifies it. A tool that needs neither answer should be deleted. Give each tool its own role where you can ([check](/checklist/tool-own-role/)), so that a misused tool reaches only what it was built for. ## Related pages - [Network reachability for AI agents on AWS, Google Cloud and Azure](/guides/network-reachability-for-agents/) - [Logging what an AI agent did: audit, data and model logs](/guides/logging-what-an-agent-did/) - [AI agent permissions: how to cut an agent's access](/guides/least-privilege-for-ai-agents/) - [All of them](/guides/) ## Sources - AWS: AgentCore Gateway outbound authorization: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html (read 10 Nov 2026) - AWS: IAM permissions for AgentCore Runtime: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html (read 10 Nov 2026) - Google Cloud: identity delegation architecture for agents: https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/identity-delegation-architecture (read 10 Nov 2026) - Microsoft: agent identity in Foundry Agent Service: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 10 Nov 2026) - AWS: Lambda execution role: https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/guides/why-an-agents-permissions-are-not-its-users/ --- # By default an agent can reach the internet on most platforms, and private networking is a setting you must choose Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. What an agent can reach over the network depends on where it runs and what you attach. This guide covers private endpoints, VPC attachment and egress filtering for Bedrock, AgentCore, Agent Runtime, Foundry and self-hosted agents, with the defaults each provider documents and the gaps we found. ## Reachability has two directions and the review needs both Network reach for an agent means two things. The first is which services the agent can call: model endpoints, cloud APIs, data stores, tool endpoints and the open internet. The second is who can call the agent or its tools. Identity decides what a call is allowed to do, but the network decides whether a stolen credential or a hijacked prompt can send data to a server of the attacker's choosing. A private path makes every outside connection a decision someone made, and an open path makes it a default nobody did. The defaults below differ by platform, so a review should read them platform by platform instead of assuming one. ## Inbound access to the agent and its tools is the other half Reachability also covers who may call the agent and its tools. A tool endpoint that accepts calls from any caller can be used without the agent at all. The checklist asks for a restriction on which principals may invoke each tool ([check](/checklist/restrict-tool-invokers/)). For a Bedrock action group that is a resource-based policy on the Lambda function allowing only the Bedrock service principal for one agent ARN. For Cloud Run it is granting the Invoker role only to the agent's service account. For Foundry it is assigning roles at agent scope, so access to one agent's endpoints does not extend to every agent in the project. These are identity controls, but they decide who can use a network path that already exists. ## AWS offers endpoints for the APIs and VPC attachment for the compute For [Amazon Bedrock Agents Classic](/agent-platforms/bedrock-agents/), AWS documents interface endpoints (AWS PrivateLink) for the bedrock-agent and bedrock-agent-runtime APIs, with an endpoint policy you can attach. AWS's VPC page lists model customization, batch inference and knowledge bases as the features it describes for VPC protection. Egress controls for the agent's own orchestration were none found on 2026-11-10. The network egress of an action group Lambda function follows that function's own VPC settings. On [AgentCore](/agent-platforms/bedrock-agentcore/), AWS documents VPC connectivity for Runtime and the built-in Code Interpreter and Browser tools. AgentCore creates network interfaces in your VPC through a service-linked role. A public subnet does not give internet access: AWS says internet access needs private subnets with a route to a NAT Gateway. The same rule applies to Lambda, which AWS documents as needing private subnets to reach private resources. For an agent on [Lambda or EKS](/agent-platforms/self-hosted-aws/), network controls for pods in EKS were none found in the pages read. ## Google Cloud puts the agent in a managed network until you attach yours Google documents that an agent on [Agent Runtime](/agent-platforms/agent-runtime-gcp/) runs in a Google-managed network without access to your VPC unless you configure a Private Service Connect interface. The interface gives private egress to your VPC and to on-premises resources, and Google says all traffic between the agent and your VPC stays inside Google's network. With VPC Service Controls, the interface is also required for internet access, and the project must be inside the perimeter before the agent is deployed, otherwise Google documents that the agent is not secured by the perimeter and continues to have public internet access. For [Cloud Run or GKE](/agent-platforms/self-hosted-gcp/), Cloud Run can send egress traffic to a VPC network with Direct VPC egress (Google's recommendation) or a VPC connector. In GKE, Google documents that by default all Pods in a cluster can communicate with each other until network policy is enforced. Each of these defaults is a setting someone has to change on purpose, and the change belongs in the review record with the date it was made. ## Azure documents a private setup for Foundry and open defaults elsewhere Microsoft documents a Standard Setup for [Foundry Agent Service](/agent-platforms/foundry-agent-service/) with private networking: no public egress, and a delegated subnet from your virtual network. Private networking is documented for prompt agents, and hosted agents support a bring-your-own virtual network, with each session in a VM-isolated sandbox connected to your VNet. For [Container Apps and AKS](/agent-platforms/self-hosted-azure/) the defaults are open. Microsoft documents that Container Apps is publicly accessible by default and can talk only to internet-accessible endpoints until you supply your own virtual network, and that in AKS all pods can send and receive traffic without limitations by default. ## Filter outbound traffic and keep the cloud APIs private Four checks cover the network. Run agent compute on a private network ([check](/checklist/private-network-no-direct-internet/)). Allow outbound traffic only to destinations the agent needs, using security groups and AWS Network Firewall, VPC firewall rules, or network security groups and Azure Firewall ([check](/checklist/allowlist-egress/)). Reach cloud APIs through private endpoints ([check](/checklist/private-endpoints-for-cloud-apis/)). Keep the agent's code from reading node or neighbor credentials ([check](/checklist/credential-isolation/)). A perimeter that stops copying of data out is a separate control ([check](/checklist/data-perimeter/)). Record the gaps too: any cell in [the checklist](/checklist/) that reads none found is a question for the platform team. Each answer belongs in the review record next to the platform it came from. A perimeter around the data is the last layer, and each cloud has a different form. AWS documents an endpoint policy as a resource-based policy attached to a VPC endpoint, which does not override identity-based or resource-based policies. Google documents VPC Service Controls as protecting against the risk of data exfiltration. Microsoft documents virtual network service endpoint policies as allowing data exfiltration only to specific Azure Storage accounts. Each is built by the network or platform team, not by the agent's builder, so the review should ask who owns it and when it was last tested. ## Related pages - [Logging what an AI agent did: audit, data and model logs](/guides/logging-what-an-agent-did/) - [AI agent permissions: how to cut an agent's access](/guides/least-privilege-for-ai-agents/) - [Amazon Bedrock agent permissions: roles, trust and what the console creates](/guides/bedrock-agent-permissions/) - [All of them](/guides/) ## Sources - AWS: interface VPC endpoints for Amazon Bedrock: https://docs.aws.amazon.com/bedrock/latest/userguide/vpc-interface-endpoints.html (read 10 Nov 2026) - AWS: AgentCore Runtime and tools in a VPC: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/agentcore-vpc.html (read 10 Nov 2026) - AWS: connect a Lambda function to a VPC: https://docs.aws.amazon.com/lambda/latest/dg/configuration-vpc.html (read 10 Nov 2026) - Google Cloud: Private Service Connect interface with Agent Runtime: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/private-service-connect-interface (read 10 Nov 2026) - Google Cloud: connect Cloud Run to a VPC network: https://cloud.google.com/run/docs/configuring/connecting-vpc (read 10 Nov 2026) - Microsoft: private networking for Foundry Agent Service: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/virtual-networks (read 10 Nov 2026) - Microsoft: networking in Azure Container Apps: https://learn.microsoft.com/en-us/azure/container-apps/networking (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/guides/network-reachability-for-agents/ --- # The logs that show what an agent did are mostly off until someone turns them on, and they must name the agent Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. After an incident the question is which agent did what, to which data. This guide covers the three log layers on AWS, Google Cloud and Azure (control plane, data access, agent trace), which ones are off by default, and whether each entry names the agent. ## Three layers of log answer three different questions A control-plane log shows who changed a role, a setting or a policy. A data-level log shows which objects, rows or keys the agent read and wrote. An agent trace shows the steps the agent took: the model calls, the tool calls and the order. A review needs all three, because each answers something the others cannot. The control-plane log will not show that the agent read a customer table, and the trace will not show that someone widened the agent's role at midnight. Each layer also has a different default, and the defaults are where the gaps sit. The rest of this guide goes cloud by cloud. ## AWS records API calls, but data events and model logs need switching on AWS documents that CloudTrail captures all Amazon Bedrock API calls as events, including calls from the console and from code. For [Agents Classic](/agent-platforms/bedrock-agents/), InvokeAgent and InvokeInlineAgent calls are logged as data events that you must enable with advanced event selectors for the AWS::Bedrock::AgentAlias resource type. AWS documents that CloudTrail trails do not log data events by default, and that extra charges apply. Model invocation logging, which collects the full request and response data for supported calls, is disabled by default and can deliver to CloudWatch Logs or Amazon S3. On [AgentCore](/agent-platforms/bedrock-agentcore/), metrics, spans and logs are stored in Amazon CloudWatch. AWS documents CloudTrail coverage for only some AgentCore capabilities, and coverage for Runtime API calls was none found on 2026-11-10 in the pages read. For a Lambda agent, the CloudTrail userIdentity element records the type of IAM identity that made each request. AWS documents that Runtime and other AgentCore data goes to CloudWatch, so retention and who can read those log groups are settings to check as well. ## Google Cloud separates always-on admin logs from data access logs Google documents Admin Activity audit logs as written for user-driven API calls that change configuration, and as impossible to disable. Data Access audit logs are disabled by default for all services except a few, so reads and writes of the agent's data appear only after you enable them. For [Agent Runtime](/agent-platforms/agent-runtime-gcp/), Google documents that the agent's stdout and stderr go to Cloud Logging under the log IDs reasoning_engine_stdout and reasoning_engine_stderr, that agent identity shows up in logs across Google Cloud services, and that for user-delegated flows the logs show both the user and the agent. Cloud Logging is not supported for child resources such as Sessions and Memory Bank. Confirm the Data Access setting for each service the agent touches before the review. ## Azure splits identity logs, activity logs and traces Microsoft documents three sources for [Foundry Agent Service](/agent-platforms/foundry-agent-service/): Entra sign-in logs, which show which agent identity obtained a token for which resource, including on behalf of which user; Entra audit logs, which record blueprint and agent identity creation and deletion; and Foundry tracing in the connected Application Insights resource for tool calls. The Azure Monitor activity log records management operations on Azure resources. Resource logs, which carry data operations such as blob reads, are not collected until you create a diagnostic setting. For [Container Apps and AKS](/agent-platforms/self-hosted-azure/), console and system logs go to Log Analytics. ## Cost and volume are the usual reasons logs stay off Data-level logs are large, so they are the first thing switched off. AWS documents that CloudTrail data events are often high-volume, that trails do not log them by default, and that additional charges apply. AWS also documents that advanced event selectors can filter on fields such as eventName, so the log can be limited to the agent alias or the bucket that matters instead of the whole account. Google documents that Data Access audit logs can be configured per service and per project, folder or organization. The review should accept a narrow data log on the agent's own resources and refuse an empty one. Where the logs go matters as much as whether they exist. AWS documents organization trails, which log events for all accounts in an organization. Google documents that entries written by Cloud Audit Logs are immutable. A log destination that the agent's own role can edit or delete cannot serve as evidence of what the agent did. Ask who administers the log destination, and whether that group includes anyone who can also change the agent's role. ## Four checks make the logs usable in an investigation Turn on organization-wide control-plane logging ([check](/checklist/control-plane-audit-logs/)). Log reads and writes to the data the agent touches ([check](/checklist/log-data-access/)). Make each logged action traceable to one agent identity, which needs the identity work in [the first guide](/guides/the-identity-an-agent-runs-as/) to be done ([check](/checklist/attribute-actions-to-the-agent/)). Alert on denied calls and on changes to the agent's role or policies ([check](/checklist/alert-on-denied-and-role-changes/)). Also decide where the logs live: a log the agent's own role can edit or delete is weak evidence. [The map](/map/) lists the logging entry for each platform. ## Related pages - [AI agent permissions: how to cut an agent's access](/guides/least-privilege-for-ai-agents/) - [Amazon Bedrock agent permissions: roles, trust and what the console creates](/guides/bedrock-agent-permissions/) - [Vertex AI agent service account: Agent Runtime identities in Google Cloud](/guides/vertex-ai-agent-service-account/) - [All of them](/guides/) ## Sources - AWS: monitor Amazon Bedrock API calls using CloudTrail: https://docs.aws.amazon.com/bedrock/latest/userguide/logging-using-cloudtrail.html (read 10 Nov 2026) - AWS: Bedrock model invocation logging: https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html (read 10 Nov 2026) - AWS: logging data events in CloudTrail: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/logging-data-events-with-cloudtrail.html (read 10 Nov 2026) - Google Cloud: enable Data Access audit logs: https://cloud.google.com/logging/docs/audit/configure-data-access (read 10 Nov 2026) - Google Cloud: logging for Agent Runtime: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/logging (read 10 Nov 2026) - Microsoft: agent identity in Foundry Agent Service: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 10 Nov 2026) - Microsoft: monitor Azure Blob Storage: https://learn.microsoft.com/en-us/azure/storage/blobs/monitor-blob-storage (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/guides/logging-what-an-agent-did/ --- # AI agent permissions should be written from what the tools call, then trimmed with usage data, not copied from a starter Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. AI agent permissions start too wide when a starter role or broad managed policy gets the agent working fast. This guide shows how to write the policy from the tools' real calls, set a ceiling above it, remove unused grants and test changes, with each cloud's documented mechanism. ## Start from what the tools call, not from the starter policy Least privilege for an agent begins with a list: every API call the agent's tools can make, with the resources each call touches. The policy is written from that list ([check](/checklist/least-privilege-policy/)). Starter material does the opposite. AWS documents that the BedrockAgentCoreFullAccess managed policy grants broad permissions and recommends a custom policy with only what the application needs. For deployment, AWS says to use broad permissions only for development and testing, and to create custom policies for production. A policy built from the tool list is shorter, and it makes a hijacked tool call fail at the cloud instead of succeeding quietly. ## Each platform starts the agent with a different amount The starting point differs. On [Foundry Agent Service](/agent-platforms/foundry-agent-service/), Microsoft documents that a new agent identity starts without resource permissions and that Foundry assigns no RBAC roles to it. For people who only call agents, Microsoft names Foundry Agent Consumer as the least-privilege role. On [Agent Runtime](/agent-platforms/agent-runtime-gcp/), an agent with agent identity starts with the Agent Default Access role, covering basic project-wide logging and model calling, and the Agent Context Editor role, limited to its own sessions, memories and sandboxes. On [Bedrock Agents Classic](/agent-platforms/bedrock-agents/), the documented service role policy covers invoking named foundation models, reading the schema objects in S3 and querying attached knowledge bases. AWS tells you to narrow it further. Its Bedrock guidance says to replace the wildcard with specific agent IDs once the agents exist, and the same habit applies to every resource ARN in the policy ([check](/checklist/narrowest-scope/)). Microsoft gives the same advice for Azure RBAC: avoid assigning broader roles at broader scopes, even when that seems more convenient. Whatever the starting point, the tools matter as much as the agent: give each tool function its own narrow role instead of the agent's role ([check](/checklist/tool-own-role/)). ## Set a ceiling above the role, because role policies get edited A role's own policy can be widened by whoever administers the account. A ceiling set higher in the hierarchy still holds when that happens ([check](/checklist/set-permission-ceiling/)). AWS documents service control policies as specifying the maximum available permissions for IAM roles in an organization, and permissions boundaries as setting the maximum an identity-based policy can grant to a role. Google documents Principal Access Boundary policies as defining the resources a principal may access. For Azure, Microsoft documents that customer-managed deny assignments are in private preview, and that management groups let you apply governance conditions above subscriptions. Pair the ceiling with an explicit deny of IAM changes and destructive actions ([check](/checklist/deny-iam-changes-and-destructive-actions/)). How to write those policies is the work of other references; here the review only needs to see that one exists. ## Scope the grant by resource and by condition A role granted on a whole account, project or subscription covers every resource in it, including ones created later. Each cloud documents a way to narrow that. AWS documents attribute-based access control, where tags on IAM entities and resources are the attributes ([check](/checklist/condition-based-data-access/)). Google documents IAM Conditions, which make a role binding apply only when the conditions are met. Microsoft documents Azure role assignment conditions, which add attribute-based filtering to a role assignment. Where the agent reads data, a condition keyed on a tag means a new data store does not become reachable by default, which a list of resource names cannot promise once it goes stale. ## Trim with usage data after the agent has run Permissions granted during a pilot tend to stay. Each cloud documents a tool for finding the unused part ([check](/checklist/remove-unused-permissions/)). AWS describes IAM Access Analyzer unused access analyzers as helping identify unused access. Google documents IAM role recommendations as helping you identify and remove excess permissions. Microsoft documents Entra access reviews as a way to manage role assignments on a recurring basis. Schedule the first trim for a set number of weeks after launch, with an owner named in the review. An agent's real call pattern is clear only once it has handled real traffic, and the grants it never used are the cheapest to remove. ## Test each change before it is enforced Tightening an agent's access can break a tool in ways nobody saw coming, so test first ([check](/checklist/test-before-enforcing/)). AWS documents the IAM policy simulator for testing identity-based policies. Google documents organization policies in dry-run mode, which monitor how a policy would behave without enforcing it. Microsoft documents the enforcementMode property on an Azure Policy assignment, where DoNotEnforce lets you see the outcome on existing resources without initiating the policy effect. Keep the result of each test with the review record. The [full checklist](/checklist/) lists the other 29 checks, and [the map](/map/) shows what each platform grants by default. ## Related pages - [Amazon Bedrock agent permissions: roles, trust and what the console creates](/guides/bedrock-agent-permissions/) - [Vertex AI agent service account: Agent Runtime identities in Google Cloud](/guides/vertex-ai-agent-service-account/) - [What can an AI agent access in AWS: the four layers to read](/guides/what-can-an-ai-agent-access-in-aws/) - [All of them](/guides/) ## Sources - AWS: IAM permissions for AgentCore Runtime: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html (read 10 Nov 2026) - AWS: create a service role for Amazon Bedrock Agents: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html (read 10 Nov 2026) - AWS: IAM Access Analyzer: https://docs.aws.amazon.com/IAM/latest/UserGuide/what-is-access-analyzer.html (read 10 Nov 2026) - Google Cloud: IAM recommender overview: https://cloud.google.com/iam/docs/recommender-overview (read 10 Nov 2026) - Google Cloud: Principal Access Boundary policies: https://cloud.google.com/iam/docs/principal-access-boundary-policies (read 10 Nov 2026) - Microsoft: Azure RBAC best practices: https://learn.microsoft.com/en-us/azure/role-based-access-control/best-practices (read 10 Nov 2026) - Microsoft: Azure Policy assignment structure: https://learn.microsoft.com/en-us/azure/governance/policy/concepts/assignment-structure (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/guides/least-privilege-for-ai-agents/ --- # A Bedrock agent runs as a service role, and the trust policy and the console-created role decide how far that role reaches Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Bedrock agents run as IAM roles that the service assumes. This guide covers Agents Classic and AgentCore: the trust policy and its condition keys, the documented permission statements, the role the console creates, the Lambda resource-based policy for tools, and the gaps in AWS documentation. ## Agents Classic runs as a service role that Bedrock assumes AWS documents that [Amazon Bedrock Agents Classic](/agent-platforms/bedrock-agents/) runs as an IAM service role. You can let Bedrock create the role automatically or supply a custom one. AWS's status page says you can no longer sign up for Agents Classic, that existing users can continue, and that AgentCore is the service it points to for similar capabilities. Anyone reviewing an existing agent still needs to read the role, because the agent is what that role can do. The role has two parts that matter: the trust policy, which says who may assume it, and the identity-based policy, which says what it may do once assumed. AWS states that Agents Classic will no longer be open to new customers starting July 30, 2026, a date that has passed as of this reading. ## The trust policy should name the account and the agent The documented trust policy allows the bedrock.amazonaws.com service principal to assume the role. AWS shows two optional conditions: aws:SourceAccount set to your account ID and aws:SourceArn matching agents in your Region and account. The sample ends the ARN with a wildcard, and AWS adds a note that, as a best practice, you should replace the * with specific agent IDs after you have created them. Without the conditions, AWS describes the confused deputy problem: an entity without permission to perform an action coerces a more-privileged entity to perform it. Scoping the trust to one agent is the checklist item [scope trust to the agent](/checklist/scope-trust-to-the-agent/). ## The documented policy is short, and the console role is not shown The documented identity-based policy has three standing statements: bedrock:InvokeModel on the named foundation models, s3:GetObject on the action group schema objects, and bedrock:Retrieve and bedrock:RetrieveAndGenerate on attached knowledge bases. Optional statements cover Provisioned Throughput, agent collaborators, guardrails, KMS decryption and files for code interpretation. The role that Bedrock creates automatically has no policy listed on that page, so its exact contents were none found on 2026-11-10. If your agent uses the automatic role, read the attached policy in IAM and compare it with the documented one before approval ([check](/checklist/least-privilege-policy/)). ## Tools add a second role and a resource-based policy Each action group is defined by an OpenAPI schema or function details and is usually backed by a Lambda function. AWS documents that you must attach a resource-based policy to the function so Bedrock may invoke it. The function then runs under its own Lambda execution role, so the tool's reach is the function's role, not the agent's. Give each function a narrow role ([check](/checklist/tool-own-role/)) and restrict the resource-based policy to the Bedrock service principal for one agent ARN ([check](/checklist/restrict-tool-invokers/)). Also limit who holds iam:PassRole for the agent's role ([check](/checklist/restrict-who-can-attach-role/)), since AWS documents PassRole as the permission needed to hand a role to a service. ## Gateway and Identity add credentials beyond the execution role AgentCore Gateway and AgentCore Identity carry access of their own. For targets that use IAM-based outbound authorization, the gateway service role signs the call with AWS Signature Version 4. AWS documents that the console or CLI can create that role for you with the permissions the target needs, and that each time you add a target the necessary permissions are attached. AgentCore Identity is a separate service for managing credentials so that agents can reach AWS resources and third-party services on behalf of users, and AWS says agent identities are implemented as workload identities. These credentials do not appear in the execution role, so a review that reads only that role will under-count the agent's reach. ## A read order for a Bedrock agent review Use the same order for both generations of the service. First, name the role and read its trust policy. Second, read its attached policies and compare them with the documented minimum. Third, list the tools, the function behind each, and the role each function runs as. Fourth, read the resource-based policy on each function. Fifth, check who holds iam:PassRole and the AgentCore or Bedrock administrative policies. Sixth, check the logging: Bedrock CloudTrail data events for the agent alias, and model invocation logging, which AWS documents as disabled by default. Record each result with the date. The AWS pages above change, and [the map](/map/) shows the read date for each platform entry. ## AgentCore uses an execution role, and the console can create it On [AgentCore](/agent-platforms/bedrock-agentcore/), Runtime needs an IAM execution role. The documented trust policy names the bedrock-agentcore.amazonaws.com service principal, with SourceAccount and a SourceArn that ends in a wildcard for the account. For the console to create the role, AWS lists iam:CreateRole limited to role names starting AmazonBedrockAgentCoreRuntimeDefaultServiceRole-, plus matching policy permissions. What that automatic role contains was none found on 2026-11-10. Gateway can create a service role and attach permissions each time you add a target, so the role grows with the target list. Review the role after every target is added, and replace the BedrockAgentCoreFullAccess policy on the people and pipelines that use AgentCore with a custom one, as AWS recommends. ## Related pages - [Vertex AI agent service account: Agent Runtime identities in Google Cloud](/guides/vertex-ai-agent-service-account/) - [What can an AI agent access in AWS: the four layers to read](/guides/what-can-an-ai-agent-access-in-aws/) - [Approving an AI agent for production: running the review meeting](/guides/approving-an-agent-for-production/) - [All of them](/guides/) ## Sources - AWS: create a service role for Amazon Bedrock Agents: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html (read 10 Nov 2026) - AWS: Lambda functions for Bedrock action groups: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-lambda.html (read 10 Nov 2026) - AWS: Bedrock Agents Classic maintenance mode: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-classic-maintenance-mode.html (read 10 Nov 2026) - AWS: IAM permissions for AgentCore Runtime: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html (read 10 Nov 2026) - AWS: AgentCore Gateway outbound authorization: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html (read 10 Nov 2026) - AWS: the confused deputy problem: https://docs.aws.amazon.com/IAM/latest/UserGuide/confused-deputy.html (read 10 Nov 2026) - AWS: granting a user permissions to pass a role: https://docs.aws.amazon.com/IAM/latest/UserGuide/id_roles_use_passrole.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/guides/bedrock-agent-permissions/ --- # An agent on Agent Runtime can run as its own agent identity, the default service agent or a shared service account Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Agent Runtime, formerly Vertex AI Agent Engine, lets an agent run as a per-agent identity, a Google-managed service agent or a custom service account. This guide compares the three, the roles each starts with, how to write the principal in a policy, and what a Private Service Connect interface adds. ## The product has a new name and the old names remain in places Google renamed the product that was Vertex AI Agent Engine in the Agent Builder documentation. The old URLs redirect to the Gemini Enterprise Agent Platform documentation, where it is called [Agent Runtime](/agent-platforms/agent-runtime-gcp/), and some role names still read Vertex AI Agent Engine. Google also labels some features Preview or Pre-GA on these pages, so check the label before relying on a behavior. When a ticket says "the Vertex agent's service account", find out which of the three identities below is meant, because the answer changes who else shares the grants. [Gemini Enterprise](/agent-platforms/gemini-enterprise/) agents that you register run under the Agent Runtime identity described here. Agent Runtime is documented as the place where an agent receives a runtime identity, a network position and logs, so the three guides on [identity](/guides/the-identity-an-agent-runs-as/), [network](/guides/network-reachability-for-agents/) and [logging](/guides/logging-what-an-agent-did/) apply to it directly. ## Three identity choices, one of them recommended Google creates and manages service agents, and they are not listed by default on the IAM page. That makes the default hard to see in a review, so ask for the role grants of the Reasoning Engine Service Agent explicitly rather than assuming it holds nothing. Google documents these choices for an agent on Agent Runtime. Ask the builder for the identity in writing, and check it against the setting in the console. - Agent identity: a per-agent IAM principal, which Google recommends. It is tied to the lifecycle of the agent and, Google says, is a more secure principal than a service account. - The AI Platform Reasoning Engine Service Agent: a Google-managed service agent, used by default if you choose nothing. - A custom service account: shared across every agent you deploy with it, and likely to need the Agent Platform User role plus whatever roles your code requires. ## Agent identity starts small and is named in policies by principal With agent identity, Google documents that the agent starts with the Agent Default Access role, which covers basic project-wide logging and model calling, and the Agent Context Editor role, limited to its own sessions, memories and sandboxes. Further grants are made to a principal written as principal://agents.global.org-ORGANIZATION_ID.system.id.goog/resources/aiplatform/projects/PROJECT_NUMBER/locations/LOCATION/reasoningEngines/AGENT_ENGINE_ID. Google warns that granting roles to the wrong principal produces permission errors. Once IAM is configured, Application Default Credentials in the Agent Platform SDK use the agent identity. Each grant to that principal is a deliberate choice, which is the point of having a per-agent identity: the list of its roles is the list of what this one agent can reach. ## Deny policies can target agent identities as a group Google's agent identity page includes a deny policy example that denies all agents across an organization the ability to create or delete buckets. That is the kind of ceiling the checklist asks for ([check](/checklist/deny-iam-changes-and-destructive-actions/)), and it sits above any role granted to a single agent ([check](/checklist/set-permission-ceiling/)). Google documents agent identity as tied to the lifecycle of the agent ([check](/checklist/retire-identity-with-agent/)). A service account is a separate resource, so retiring the agent leaves the account and its grants in place until someone deletes them. Limit who can attach a service account to a resource by granting the Service Account User role narrowly ([check](/checklist/restrict-who-can-attach-role/)). Google documents that the Service Account User role lets a principal attach a service account to a resource, which gives the resource's workloads that account's access. ## Gemini Enterprise registers agents that run here Agents registered in [Gemini Enterprise](/agent-platforms/gemini-enterprise/) run on Agent Runtime, so the identity above applies to them. Google documents that registering an ADK agent requires the Gemini Enterprise Admin role for the person doing it, and that Gemini Enterprise gives ADK agents the user's email address. An agent that must reach Google Cloud resources on a user's behalf needs authorization details set up: an OAuth client, an authorization resource registered with Gemini Enterprise, and user consent. A Gemini Enterprise administrator can see the agent's identity, typically its SPIFFE ID. Permissions that Gemini Enterprise itself grants a registered agent were none found on 2026-11-10 in the pages read. Read it with the Agent Runtime pages to get the full reach of a registered agent. ## Network access needs its own decision The identity says what the agent may call. The network says where the call can go. Google documents that the agent runs in a Google-managed network with no access to your VPC unless you configure a Private Service Connect interface. The interface needs the Agent Platform Service Agent of the host project to hold either the Compute Network Admin role or a custom role with listed permissions, which is itself a grant to review. With VPC Service Controls, the interface is required for internet access. The [network guide](/guides/network-reachability-for-agents/) covers the rest, and [the checklist](/checklist/) has the 30 checks. ## Related pages - [What can an AI agent access in AWS: the four layers to read](/guides/what-can-an-ai-agent-access-in-aws/) - [Approving an AI agent for production: running the review meeting](/guides/approving-an-agent-for-production/) - [AI agent permissions in Azure: Foundry Agent Service, identities and RBAC](/guides/ai-agent-permissions-in-azure/) - [All of them](/guides/) ## Sources - Google Cloud: agent identity on Agent Runtime: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 10 Nov 2026) - Google Cloud: set up Agent Runtime: https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime/setup (read 10 Nov 2026) - Google Cloud: types of service accounts and service agents: https://cloud.google.com/iam/docs/service-account-types (read 10 Nov 2026) - Google Cloud: Private Service Connect interface with Agent Runtime: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/private-service-connect-interface (read 10 Nov 2026) - Google Cloud: service account permissions: https://cloud.google.com/iam/docs/service-account-permissions (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/guides/vertex-ai-agent-service-account/ --- # An agent's access in AWS is its role's policy, plus what its tools reach, plus resource policies, minus the ceilings above it Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. To answer what an agent can reach in an AWS account, read four layers: the role it assumes, the tools and Lambda functions behind it, the resource-based policies on the targets, and the ceilings and network controls above them. This guide gives the order and the AWS pages that document each layer. ## Access is a sum and a subtraction, read in four layers No single page in the console answers what an agent can access. The answer comes from four layers. The first is the identity-based policy on the role the agent assumes. The second is what the agent's tools can reach, because a tool can run under a different role. The third is the resource-based policies on the targets, which can name the agent's role as a principal. The fourth is the ceilings and network controls above the role, which can remove access the role's own policy grants. Read them in that order and write down each result. [The map](/map/) records the first two layers for the AWS platforms, and the [AWS cloud page](/clouds/aws/) sums them up. ## Layer one is the role the agent assumes For [Bedrock Agents Classic](/agent-platforms/bedrock-agents/), the documented service role policy covers invoking named foundation models, reading the OpenAPI schema objects in S3 and querying attached knowledge bases. For [AgentCore](/agent-platforms/bedrock-agentcore/), Runtime runs under an execution role that AgentCore assumes. For an agent on [Lambda or EKS](/agent-platforms/self-hosted-aws/), the function runs as its execution role, and an EKS pod gets the role mapped to its service account by an EKS Pod Identity association. AWS documents that a container never has access to credentials used by containers in other pods, but that pods can reach the node's role if access to the instance metadata service is not restricted. Read the attached policies, not the role's name. ## Layer two is what the tools reach, and they may not use that role In Bedrock Agents Classic, an action group is usually a Lambda function. AWS documents that the agent reaches it through a resource-based policy on the function, and that the function then runs under its own Lambda execution role. An action can also return control to the calling application, which then runs the action with its own credentials. On AgentCore, Gateway turns APIs, Lambda functions and existing services into tools, and each target chooses how its call is authorized: the gateway service role, the caller's credentials, OAuth, token passthrough or an API key. List every tool and name the identity it runs as ([check](/checklist/tool-own-role/)). An agent with five tools has up to six roles to read. ## Layers three and four add policies on the target and ceilings above Resource-based policies live on the thing being accessed. AWS documents a KMS key policy as a resource policy and the primary way to control access to a key, so the agent's reach into encrypted data depends on that policy as well as its role ([check](/checklist/control-the-keys/)). S3 bucket policies, Lambda function policies and VPC endpoint policies work the same way. Above the role, AWS documents that service control policies specify the maximum available permissions for roles in an organization, and that a permissions boundary sets the maximum an identity-based policy can grant to a role ([check](/checklist/set-permission-ceiling/)). Network controls decide whether the call can leave at all ([check](/checklist/data-perimeter/)). ## Three defaults catch reviewers out Three documented defaults tend to surprise reviewers. When you create a function in the Lambda console, AWS documents that it creates an execution role with minimal permissions, so a role that looks small may have been widened later by hand. On EKS, if access to the instance metadata service is not restricted, a pod's containers can also reach the node's IAM role. And the BedrockAgentCoreFullAccess policy, for the person or pipeline using AgentCore, grants broad permissions. None of the three appears in the agent's own role policy, so none shows up if the review reads only that. AWS's data perimeter whitepaper describes organizations that want perimeter controls to help protect against unintended access, built from identity, resource and network controls together. For an agent this is the layer that decides whether data the role can read can also leave. Ask which organization-level controls apply to the account the agent runs in. The checklist item [put data the agent reads inside a perimeter](/checklist/data-perimeter/) records the documented form on each cloud. ## A worksheet for the review meeting Fill one worksheet per agent. Each row is a question the AWS documentation lets you answer, and each answer is a fact about the account, not a promise from the builder. Rows left empty are open questions for the meeting described in [the approval guide](/guides/approving-an-agent-for-production/). Add a sixth row for the network path once the first five are filled. Keep the finished sheet with the review record. | Question | Where to look in AWS | Checklist item | | --- | --- | --- | | Which role does the agent run as? | The service role or execution role and its trust policy | [own identity per agent](/checklist/own-identity-per-agent/) | | What does that role allow? | Attached identity-based policies | [least-privilege policy](/checklist/least-privilege-policy/) | | What do the tools run as? | Each Lambda execution role, each Gateway target | [tool role](/checklist/tool-own-role/) | | What caps the role? | SCPs and permissions boundaries | [ceiling](/checklist/set-permission-ceiling/) | | Who can call the tools? | Resource-based policies on the functions | [tool invokers](/checklist/restrict-tool-invokers/) | ## Related pages - [Approving an AI agent for production: running the review meeting](/guides/approving-an-agent-for-production/) - [AI agent permissions in Azure: Foundry Agent Service, identities and RBAC](/guides/ai-agent-permissions-in-azure/) - [AI agent authorization in AWS, Azure and Google Cloud](/ai-agent-authorization/) - [All of them](/guides/) ## Sources - AWS: create a service role for Amazon Bedrock Agents: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html (read 10 Nov 2026) - AWS: Lambda functions for Bedrock action groups: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-lambda.html (read 10 Nov 2026) - AWS: AgentCore Gateway outbound authorization: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html (read 10 Nov 2026) - AWS: EKS Pod Identity: https://docs.aws.amazon.com/eks/latest/userguide/pod-identities.html (read 10 Nov 2026) - AWS: AWS KMS key policies: https://docs.aws.amazon.com/kms/latest/developerguide/key-policies.html (read 10 Nov 2026) - AWS: service control policies: https://docs.aws.amazon.com/organizations/latest/userguide/orgs_manage_policies_scps.html (read 10 Nov 2026) - AWS: permissions boundaries: https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_boundaries.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/guides/what-can-an-ai-agent-access-in-aws/ --- # Approve an agent in one meeting by walking the checklist in seven sections and recording an answer, an owner and a date for each gap Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. A production review for an AI agent works when it follows a fixed order and ends with recorded decisions. This guide gives the pre-read, the agenda built on the 30-check Agent Boundary Checklist, how to score a check, what to do with none found, and what the meeting must leave behind. ## The pre-read is a one-page description of the agent Ask the builder for a page before the meeting, not a slide deck. It names the platform ([the map](/map/) has the entry for each), the identity the agent runs as, the roles attached to it, every tool with the identity each tool runs as, the data stores it touches, the network it sits on, and the person who owns it. Without the page, the meeting spends its hour discovering facts that the builder could have written down in twenty minutes. With it, each of the 30 checks on [the checklist](/checklist/) becomes a yes or no question that someone in the room can answer from the cloud console, not from memory. ## Four people need to be in the room Invite the person who approves, the person who built the agent, the owner of the cloud platform it runs on, and the owner of the data it touches. The approver decides. The builder answers questions from the pre-read and shows settings. The platform owner can show the account-level controls (ceilings, logging, network) that the builder cannot see. The data owner says whether the access described is acceptable for that data. If one of the four is absent, mark the checks that need them as open instead of guessing. A review held without the platform owner is the usual reason the ceiling and logging checks are marked met without anyone looking. ## Walk the checklist in its seven sections, in order The checklist groups its 30 checks into identity, permissions, data, network, tools, change control and logging. Take them in that order, because the early sections decide how much weight the later ones carry. If the agent runs as a shared default account, the permission answers are about every workload on that account. Start with [one identity per agent](/checklist/own-identity-per-agent/). Move to [the permission policy](/checklist/least-privilege-policy/) and [the ceiling above it](/checklist/set-permission-ceiling/). Then cover data, network, [tool authority](/checklist/decide-user-or-agent-authority/), [testing before enforcing](/checklist/test-before-enforcing/) and [attributing logged actions to the agent](/checklist/attribute-actions-to-the-agent/). Each check has a how-to for AWS, Google Cloud and Azure with a provider source. Allow roughly an hour for a typical agent. A table helps the chair keep time. The counts below come from the checklist. Move on when the time for a section is used, and mark whatever remains as open instead of running over; the open list is the output of the meeting, and a long open list is a finding in itself, and it should reach the approver in writing before anyone signs. | Section | Checks | Share of the hour | | --- | --- | --- | | Identity | 5 | 12 minutes | | Permissions | 5 | 12 minutes | | Data | 4 | 8 minutes | | Network | 4 | 8 minutes | | Tools | 4 | 8 minutes | | Change control | 4 | 6 minutes | | Logging | 4 | 6 minutes | ## Score each check as met, not met or open Use three answers. Met means someone showed the setting in the account during the meeting. Not met means the setting is absent or wrong. Open means the documentation gives no way to meet the check, or nobody in the room can show it. The checklist itself contains three open cells, each reading none found on 2026-11-10: limiting who can pass or impersonate the role on Azure, avoiding default accounts on Azure, and tying the identity's removal to the agent on AWS. Treat those as open and assign an owner to find out. A provider's silence in the pages we read is not proof that the control is missing, and it is not proof that it exists. ## Hold a few facts the providers document about defaults Three documented defaults tend to settle arguments in the room. AWS says to use broad deployment permissions only for development and testing, and that BedrockAgentCoreFullAccess grants broad permissions. Microsoft documents that a new Foundry agent identity starts without resource permissions and that Foundry does not assign RBAC roles to it, so a pass on a permission check means someone granted roles on purpose. AWS documents that CloudTrail trails do not log data events by default, and Google documents that Data Access audit logs are disabled by default for most services, so a claim that everything is logged needs the setting shown ([check](/checklist/log-data-access/)). ## Leave the room with owners, dates and a trim review A review that ends with "approved" and nothing else cannot be revisited. Record four things. The decision: approved, approved with conditions, or not approved. The conditions, each with an owner and a date. The checks scored open and who will close them. A date for the first trim of unused permissions ([check](/checklist/remove-unused-permissions/)), set after the agent has run on real traffic. Store the record next to the pre-read and the dated provider pages used as evidence. When the agent is later retired, the same record tells the platform team which identity and access to remove ([check](/checklist/retire-identity-with-agent/)). ## Related pages - [AI agent permissions in Azure: Foundry Agent Service, identities and RBAC](/guides/ai-agent-permissions-in-azure/) - [AI agent authorization in AWS, Azure and Google Cloud](/ai-agent-authorization/) - [AI agent guardrails: SCPs, Azure Policy and org policy](/ai-agent-guardrails/) - [All of them](/guides/) ## Sources - AWS: IAM permissions for AgentCore Runtime: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html (read 10 Nov 2026) - Microsoft: agent identity in Foundry Agent Service: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 10 Nov 2026) - AWS: logging data events in CloudTrail: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/logging-data-events-with-cloudtrail.html (read 10 Nov 2026) - Google Cloud: enable Data Access audit logs: https://cloud.google.com/logging/docs/audit/configure-data-access (read 10 Nov 2026) - Google Cloud: agent identity on Agent Runtime: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/guides/approving-an-agent-for-production/ --- # In Azure an agent's access is the RBAC roles granted to its Entra agent identity or managed identity, and a new identity starts with none Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Microsoft Foundry Agent Service gives each agent a Microsoft Entra agent identity that starts without roles. This guide covers agent identities, project and managed identities, the Foundry RBAC roles and their renaming, how tools authenticate, key-based access that bypasses roles, and the documented limits on deny assignments. ## Each Foundry agent has an Entra agent identity that starts empty Microsoft documents that each agent in [Foundry Agent Service](/agent-platforms/foundry-agent-service/) gets a Microsoft Entra agent identity, a special service principal created under an agent blueprint. Hosted agents get a dedicated agent identity at deploy time. Microsoft also documents that a newly created agent identity has no resource permissions and that Foundry does not automatically assign RBAC roles to it. An administrator grants access through the target service, such as a storage account or Key Vault. The practical result is that a Foundry agent's reach equals the role assignments someone made for its identity. ## Shared project identities weaken the separation A project-wide managed identity also exists. Microsoft documents that legacy agents sharing the project's managed identity have no per-agent separation, which means their access is pooled ([check](/checklist/own-identity-per-agent/)). Managed identities in general come in two kinds. A system-assigned identity is enabled on one resource and is deleted with it. A user-assigned identity is a standalone resource that you can assign to several resources. Microsoft documents that managed identities remove the need to manage credentials ([check](/checklist/short-lived-credentials/)). For agents you run yourself on [Container Apps or AKS](/agent-platforms/self-hosted-azure/), choose the identity type deliberately: the first ties access to one app, and the second makes the sharing visible. ## Foundry roles were renamed and one role fits agent callers Microsoft renamed its Foundry roles: Foundry User, Foundry Owner, Foundry Account Owner and Foundry Project Manager were previously called Azure AI User, Azure AI Owner, Azure AI Account Owner and Azure AI Project Manager. Role IDs and core permissions are unchanged, and the old names still appear in places. A person or service principal that only needs to call agents, for example through the Responses API, should hold Foundry Agent Consumer instead of Foundry User, which Microsoft calls the least-privilege access for agent consumers. You can assign Foundry roles at agent scope, so access to one agent's endpoints does not extend to all agents in the project ([check](/checklist/restrict-tool-invokers/)). ## Keys bypass the roles, and tools may use another identity Microsoft documents that RBAC roles apply when you authenticate with Microsoft Entra ID, and that with key-based authentication the key grants full access without role restrictions. So the review must find out whether keys are in use. On the tools side, MCP and A2A tools receive a scoped agent identity token and the downstream resource checks that identity's role assignments. Tool connections can instead use keys, the agent's or the project's managed identity, or OAuth identity passthrough, and OpenAPI tools support anonymous, API key or project managed identity authentication. With function calling, your own application executes the function, under whatever identity your code holds. List each tool and the authentication it uses ([check](/checklist/decide-user-or-agent-authority/)). ## Logs and networks complete the Azure picture Microsoft documents that Entra sign-in logs show which agent identity obtained a token for which resource, including on behalf of which user, and that Entra audit logs record blueprint and agent identity creation and deletion. Foundry tracing in Application Insights covers tool calls. For network access, Microsoft documents a Standard Setup with private networking and no public egress, and a private endpoint for the Foundry account and its projects. The [network guide](/guides/network-reachability-for-agents/) and the [logging guide](/guides/logging-what-an-agent-did/) cover both in more detail, and [the map](/map/) lists the Foundry entry beside the AWS and Google Cloud ones. ## Self-hosted agents use managed identities and Workload ID An agent you run yourself on [Azure Container Apps](/agent-platforms/self-hosted-azure/) uses a managed identity, and Microsoft documents that you can control which managed identities are available to the app. On AKS, Microsoft Entra Workload ID is how pods authenticate to Entra-protected services such as Key Vault and Microsoft Graph, using the cluster's OIDC issuer. Microsoft documents that on AKS Automatic clusters workload identity is preconfigured, while on AKS Standard you enable it yourself. A default role grant for either platform was none found on 2026-11-10 in the pages read, so the role assignments someone made for the identity set its reach. ## Scope assignments narrowly, and treat deny assignments as limited Microsoft's RBAC guidance is to grant the least privilege needed and to avoid broader roles at broader scopes ([check](/checklist/narrowest-scope/)). Assign at resource or resource group scope, not subscription or management group scope. Azure role assignment conditions can filter a role by attributes ([check](/checklist/condition-based-data-access/)). Deny assignments block listed actions even when a role grants them, but Microsoft documents customer-managed deny assignments as currently in private preview, so for now a ceiling comes from management group governance and from keeping the agent in its own subscription ([check](/checklist/dedicated-account-project-subscription/)). Review role assignments on a schedule with Entra access reviews ([check](/checklist/remove-unused-permissions/)). Microsoft also documents that Azure Policy helps you audit and govern resource state for Foundry resources, which is a way to report when an agent project drifts. ## Related pages - [AI agent authorization in AWS, Azure and Google Cloud](/ai-agent-authorization/) - [AI agent guardrails: SCPs, Azure Policy and org policy](/ai-agent-guardrails/) - [AI access control in cloud accounts: roles, policy, network](/ai-access-control/) - [All of them](/guides/) ## Sources - Microsoft: agent identity in Foundry Agent Service: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 10 Nov 2026) - Microsoft: role-based access control for Microsoft Foundry: https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/rbac-foundry (read 10 Nov 2026) - Microsoft: managed identities for Azure resources: https://learn.microsoft.com/en-us/entra/identity/managed-identities-azure-resources/overview (read 10 Nov 2026) - Microsoft: Azure RBAC best practices: https://learn.microsoft.com/en-us/azure/role-based-access-control/best-practices (read 10 Nov 2026) - Microsoft: Azure deny assignments: https://learn.microsoft.com/en-us/azure/role-based-access-control/deny-assignments (read 10 Nov 2026) - Microsoft: managed identities in Azure Container Apps: https://learn.microsoft.com/en-us/azure/container-apps/managed-identity (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/guides/ai-agent-permissions-in-azure/ --- # AI agent authorization is decided by the agent's identity, the resource's policy and, where used, the user's token Facts read 16 Nov 2026. Page reviewed 16 Nov 2026. AI agent authorization in a cloud account is one decision made on up to three inputs: the identity the agent runs as, the policies attached to the resource it calls, and, on platforms that support it, a token that carries the user. AWS, Azure and Google Cloud documentation read on 16 Nov 2026 describes each input. ## The cloud decides on the identity that signed the call A cloud API does not see an agent. It sees a request signed by an identity, and it decides whether that identity may perform that action on that resource. AWS documents the sequence: it first authenticates the principal, then processes the request context to find which policies apply, then evaluates all policy types to allow or deny ([AWS: policy evaluation logic](https://docs.aws.amazon.com/IAM/latest/UserGuide/reference_policies_evaluation-logic.html), read 16 Nov 2026). For an agent, the identity is the role, service account or agent identity the platform runs it as. [The identity an AI agent runs as](/guides/the-identity-an-agent-runs-as/) lists it per platform. The rest of this page covers what the cloud then checks against that identity, and where a user can enter the decision. ## Authorization inputs in each cloud | Cloud | Identity the decision is made on | Where the grant sits | Limits applied on top | | --- | --- | --- | --- | | AWS | An IAM role, for example the execution role of AgentCore Runtime or the service role of a Bedrock agent | Identity-based policies on the role, and resource-based policies on the resource; AWS documents that within one account the result is the union of the two | SCPs, RCPs, permissions boundaries and session policies; AWS documents the result as the intersection with the grant | | Azure | An Entra agent identity or a managed identity | Azure RBAC role assignments at a scope (management group, subscription, resource group or resource), or a service's own data-plane roles | Deny assignments, which Azure Resource Manager checks first, and optional role assignment conditions that filter a grant by attributes | | Google Cloud | An agent identity principal, the default service agent, or a service account | Allow policies that bind roles to the principal on a resource, inherited down the hierarchy | Deny policies, which IAM checks before allow policies, and Principal Access Boundary policies on the principal | ## The role or identity decides what the agent is allowed to try On AWS, AWS documents that the AgentCore Runtime execution role is the identity the agent assumes, and that the console can create it when the console role holds the listed IAM create permissions ([AWS: IAM permissions for AgentCore Runtime](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html)). A Bedrock agent runs as a service role whose trust policy names bedrock.amazonaws.com ([AWS: service role for Bedrock Agents](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html)). Anything that role's policies allow is something the agent can attempt. On Azure, Microsoft documents that a new Foundry agent identity starts without resource permissions and that Foundry assigns it no RBAC roles ([Microsoft: agent identity concepts](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)). The grant is whatever an administrator adds, and Microsoft names the object ID of the agent identity as the assignee. On Google Cloud, Google documents that a deployed agent identity receives the Agent Default Access and Agent Context Editor roles by default, and advises against granting it project-level roles such as Vertex AI User ([Google Cloud: agent identity on Agent Runtime](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)). ## The resource's own policy can allow or refuse the same call A resource can carry its own policy, and it is evaluated beside the identity's. AWS lists IAM role trust policies and S3 bucket policies as resource-based policies and documents that a resource-based policy must name a principal ([AWS: policies and permissions in IAM](https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies.html)). AWS also documents that AgentCore Runtime resources can be protected with resource-based policies. An explicit deny in either the identity policy or the resource policy overrides an allow. Azure and Google Cloud place most of this in the hierarchy rather than on the resource. Azure Resource Manager retrieves the role assignments and deny assignments that apply to the resource and blocks the call if a deny assignment applies ([Microsoft: Azure RBAC overview](https://learn.microsoft.com/en-us/azure/role-based-access-control/overview)). Google documents that IAM always checks relevant deny policies before allow policies ([Google Cloud: deny policies](https://cloud.google.com/iam/docs/deny-overview)). Put those ceilings in place first; [the guardrails page](/ai-agent-guardrails/) covers them. ## The on-behalf-of user is a third input, and not every platform adds it Authorization can run on the agent's authority, the user's, or both. In the Microsoft identity platform, the on-behalf-of flow uses delegated scopes only: roles stay attached to the user and never to the application acting for them ([Microsoft: on-behalf-of flow](https://learn.microsoft.com/en-us/entra/identity-platform/v2-oauth2-on-behalf-of-flow)). For a Foundry agent, Microsoft documents delegated access as needing both the agent's authorization and the user's access to the resource. AWS documents several outbound options for AgentCore Gateway: the caller's IAM credentials, a user-delegated authorization code grant, a token exchange that carries both the user and agent identities, and a client credentials grant that acts on the application's own behalf ([AWS: Gateway outbound authorization](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html)). The AgentCore Runtime page warns that the BedrockAgentCoreFullAccess managed policy includes GetWorkloadAccessTokenForUserId, which issues workload access tokens from a caller-supplied user ID string without checking the identity provider, and advises production use of GetWorkloadAccessTokenForJWT instead. Google describes a delegated identity that combines the user and the agent when the agent calls downstream services for the user ([Google Cloud: identity delegation architecture](https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/identity-delegation-architecture)). The map records which platforms document each option, and [why an agent's permissions are not its user's](/guides/why-an-agents-permissions-are-not-its-users/) explains the gap when the platform does not carry the user. ## What to check in a review - Name the identity the agent runs as, and confirm it is not shared with another agent ([check: own identity per agent](/checklist/own-identity-per-agent/)). - List the identity's grants and the resource policies on what it reaches. - Find the ceilings above it: organization policies, deny rules and boundaries ([check: set a ceiling](/checklist/set-permission-ceiling/)). - For each tool, record whether it runs on the agent's authority or the user's ([check: user or agent authority](/checklist/decide-user-or-agent-authority/)). - Read [the Agent Access Map](/map/) for the platform's documented defaults. ## Related pages - [AI agent guardrails: SCPs, Azure Policy and org policy](/ai-agent-guardrails/) - [AI access control in cloud accounts: roles, policy, network](/ai-access-control/) - [Agent authorization patterns: five ways to scope an agent](/agent-authorization-patterns/) - [All of them](/guides/) ## Sources - AWS: policy evaluation logic: https://docs.aws.amazon.com/IAM/latest/UserGuide/reference_policies_evaluation-logic.html (read 16 Nov 2026) - AWS: policies and permissions in IAM: https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies.html (read 16 Nov 2026) - AWS: IAM permissions for AgentCore Runtime: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html (read 16 Nov 2026) - AWS: create a service role for Amazon Bedrock Agents: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html (read 16 Nov 2026) - AWS: AgentCore Gateway outbound authorization: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html (read 16 Nov 2026) - Microsoft: Azure RBAC overview: https://learn.microsoft.com/en-us/azure/role-based-access-control/overview (read 16 Nov 2026) - Microsoft: agent identity concepts in Foundry: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 16 Nov 2026) - Microsoft: OAuth 2.0 on-behalf-of flow: https://learn.microsoft.com/en-us/entra/identity-platform/v2-oauth2-on-behalf-of-flow (read 16 Nov 2026) - Google Cloud: deny policies: https://cloud.google.com/iam/docs/deny-overview (read 16 Nov 2026) - Google Cloud: agent identity on Agent Runtime: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 16 Nov 2026) - Google Cloud: identity delegation architecture: https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/identity-delegation-architecture (read 16 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/ai-agent-authorization/ --- # AI agent guardrails at the access layer are organization policies, deny rules, boundaries and private endpoints Facts read 16 Nov 2026. Page reviewed 16 Nov 2026. AI agent guardrails for cloud access are the limits set above the agent's own role: AWS SCPs, RCPs and permissions boundaries, Azure Policy and deny assignments, Google Cloud organization policy, deny policies and VPC Service Controls. They cap what a misconfigured or hijacked agent can reach. Documentation read on 16 Nov 2026. ## An access guardrail limits the identity, whatever the agent decides A guardrail at the access layer works on the cloud's authorization decision, not on the model's text. It applies to the role, service account or agent identity, so it holds when a prompt is manipulated, a tool is misused or the agent's policy is wider than intended. AWS documents the principle for SCPs: they grant nothing and set limits, and effective permissions are the intersection of what the SCPs and RCPs allow and what identity-based and resource-based policies allow ([AWS: service control policies](https://docs.aws.amazon.com/organizations/latest/userguide/orgs_manage_policies_scps.html), read 16 Nov 2026). The same shape exists in each cloud: a grant written for the agent, and a ceiling written by someone else, higher up. [The ceiling check](/checklist/set-permission-ceiling/) asks for it before an agent goes to production. ## Guardrail mechanisms by cloud | Cloud | Mechanism | What it limits | Documented limit or caveat | | --- | --- | --- | --- | | AWS | Service control policies | The maximum permissions of IAM users and roles in member accounts | AWS documents that SCPs do not affect users or roles in the management account | | AWS | Resource control policies | The maximum permissions on resources, aimed at principals outside the organization | AWS documents that RCPs do not affect resources in the management account, and apply only to supported services | | AWS | Permissions boundaries | The maximum an identity-based policy can grant one role | AWS documents the effective permission as the intersection of the boundary and the identity policy | | AWS | VPC endpoint policies | Which principals can use an endpoint to reach a service | AWS documents that an endpoint policy does not replace identity or resource policies, and that services without endpoint policy support allow full access through an endpoint | | Azure | Azure Policy | Resource state: it evaluates resources and actions in Azure against policy definitions | Microsoft documents that it compares resource properties to rules, which is separate from who holds a role | | Azure | Deny assignments and role assignment conditions | Blocks a principal from actions, or filters a grant by attributes such as a tag | Microsoft documents that deny assignments are checked first by Azure Resource Manager | | Google Cloud | Organization Policy constraints | Which configurations are allowed across the hierarchy, for example which service accounts can be used | Google documents that the Organization Policy service sets constraints across the whole organization, folders and projects | | Google Cloud | IAM deny policies | Permissions a principal cannot use, whatever allow policies grant | Google documents that IAM checks deny policies before allow policies | | Google Cloud | Principal Access Boundary policies | The resources a set of principals is eligible to access | Google documents that by default principals are eligible to access any resource | | Google Cloud | VPC Service Controls | A perimeter around supported services to reduce data exfiltration | Google documents it as an extra layer beside IAM | ## Pick the guardrail by the failure it stops A permissions boundary or SCP stops an agent role from being granted IAM changes or destructive actions, which [the deny check](/checklist/deny-iam-changes-and-destructive-actions/) asks for. An RCP, a Google Cloud perimeter or a private endpoint policy stops data from leaving through a principal outside the organization ([the data perimeter check](/checklist/data-perimeter/)). Azure Policy and Google Cloud organization policy stop an agent from being deployed with a public endpoint or an unapproved identity. Microsoft documents that Azure Policy has built-in definitions for Foundry hubs and projects, which a platform administrator can assign to a subscription or use as the basis for custom definitions ([Microsoft: audit and manage Foundry with Azure Policy](https://learn.microsoft.com/en-us/azure/ai-foundry/how-to/azure-policy)). That page is marked as legacy support for hub-based projects and does not apply to Foundry projects, so check the project type before relying on it. ## Test a guardrail before enforcing it A ceiling that is wrong breaks the agent and often other workloads. AWS documents testing the effects of SCPs and RCPs before attaching them, and Google Cloud documents a dry-run mode for organization policies ([Google Cloud: dry-run mode](https://cloud.google.com/resource-manager/docs/organization-policy/dry-run-policy)). [The test-before-enforcing check](/checklist/test-before-enforcing/) lists the mechanism per cloud. The Agent Access Map shows the boundaries each platform page already names, for example on [AgentCore](/agent-platforms/bedrock-agentcore/) and [Agent Runtime](/agent-platforms/agent-runtime-gcp/). ## Model-output guardrails are a different layer Amazon Bedrock Guardrails and Azure AI Content Safety are guardrails on what a model reads and writes. AWS describes Bedrock Guardrails as safeguards to detect and filter undesirable content and protect sensitive information in user inputs and model responses. Microsoft describes Azure AI Content Safety as a service that detects harmful user-generated and AI-generated content and, in its guardrail controls, tool use by agents that is misaligned or unintended. They act on content, not on the identity's permissions, so a content filter does not remove a grant the role holds. This site covers the access layer only and does not assess them. ## Related pages - [AI access control in cloud accounts: roles, policy, network](/ai-access-control/) - [Agent authorization patterns: five ways to scope an agent](/agent-authorization-patterns/) - [Amazon Q Business default permissions: roles it creates](/default-permissions/amazon-q-business/) - [All of them](/guides/) ## Sources - AWS: service control policies: https://docs.aws.amazon.com/organizations/latest/userguide/orgs_manage_policies_scps.html (read 16 Nov 2026) - AWS: resource control policies: https://docs.aws.amazon.com/organizations/latest/userguide/orgs_manage_policies_rcps.html (read 16 Nov 2026) - AWS: permissions boundaries: https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_boundaries.html (read 16 Nov 2026) - AWS: VPC endpoint policies: https://docs.aws.amazon.com/vpc/latest/privatelink/vpc-endpoints-access.html (read 16 Nov 2026) - Microsoft: Azure Policy overview: https://learn.microsoft.com/en-us/azure/governance/policy/overview (read 16 Nov 2026) - Microsoft: Azure RBAC overview: https://learn.microsoft.com/en-us/azure/role-based-access-control/overview (read 16 Nov 2026) - Microsoft: Azure ABAC and role assignment conditions: https://learn.microsoft.com/en-us/azure/role-based-access-control/conditions-overview (read 16 Nov 2026) - Microsoft: audit and manage Foundry with Azure Policy: https://learn.microsoft.com/en-us/azure/ai-foundry/how-to/azure-policy (read 16 Nov 2026) - Google Cloud: Organization Policy overview: https://cloud.google.com/resource-manager/docs/organization-policy/overview (read 16 Nov 2026) - Google Cloud: deny policies: https://cloud.google.com/iam/docs/deny-overview (read 16 Nov 2026) - Google Cloud: Principal Access Boundary policies: https://cloud.google.com/iam/docs/principal-access-boundary-policies (read 16 Nov 2026) - Google Cloud: VPC Service Controls overview: https://cloud.google.com/vpc-service-controls/docs/overview (read 16 Nov 2026) - AWS: Amazon Bedrock Guardrails: https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html (read 16 Nov 2026) - Microsoft: Azure AI Content Safety overview: https://learn.microsoft.com/en-us/azure/ai-services/content-safety/overview (read 16 Nov 2026) - Google Cloud: dry-run mode for organization policy: https://cloud.google.com/resource-manager/docs/organization-policy/dry-run-policy (read 16 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/ai-agent-guardrails/ --- # AI access control in a cloud account is four layers around the identity the AI workload runs as Facts read 16 Nov 2026. Page reviewed 16 Nov 2026. AI access control for AI workloads in a cloud account has four layers: the identity the workload runs as, the policies that grant it access, the ceilings above those policies, and the network paths it can use. This page lists each layer for AWS, Azure and Google Cloud from documentation read on 16 Nov 2026. ## Access control for an AI workload is the cloud's access control applied to a new kind of caller An AI workload is an agent, an inference app or a notebook that calls cloud APIs. The cloud controls its access the way it controls any workload: by identity, by grants, by ceilings and by network path. Nothing about the model changes that. What changes is the number of identities, because platforms create roles and service accounts when a console wizard or a quick start runs, and the scope of the grants those identities receive. This page covers the controls in the cloud account. It does not cover prompt filtering or content moderation, which [the guardrails page](/ai-agent-guardrails/) sets apart. ## The four layers, by cloud | Layer | AWS | Azure | Google Cloud | | --- | --- | --- | --- | | Identity | IAM role (AgentCore execution role, Bedrock service role, Lambda execution role) | Entra agent identity or managed identity | Agent identity principal, service agent or service account | | Grant | Identity-based and resource-based policies | Azure RBAC role assignments at a scope; service data-plane roles | IAM allow policies binding roles, inherited down the hierarchy | | Ceiling | SCPs, RCPs, permissions boundaries, session policies | Deny assignments and role assignment conditions, with Azure Policy on resource state | Organization policy constraints, deny policies, Principal Access Boundary policies | | Network | VPC endpoints with endpoint policies | Private networking for the platform, where it offers it | VPC Service Controls perimeters | ## Identity: one per workload, and know the platform default Platforms differ on the default. Microsoft documents that a new Foundry agent identity starts with no RBAC roles ([Microsoft: agent identity concepts](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)). Google documents that a deployed Agent Runtime identity receives the Agent Default Access and Agent Context Editor roles ([Google Cloud: agent identity on Agent Runtime](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)). AWS documents that the AgentCore console can create the execution role on your behalf ([AWS: IAM permissions for AgentCore Runtime](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html)). The contents of that auto-created role: none found on 16 Nov 2026 in the pages read. A shared identity is the common failure. If two agents run as one role, every grant for one is a grant for both, and the logs cannot tell them apart ([check: own identity per agent](/checklist/own-identity-per-agent/)). The per-platform entries are on [the platforms page](/agent-platforms/). ## Grants and ceilings: what the workload receives, and what it cannot exceed AWS documents that for a request inside an organization the result is the intersection of the identity policy, the SCP and the RCP ([AWS: policy evaluation logic](https://docs.aws.amazon.com/IAM/latest/UserGuide/reference_policies_evaluation-logic.html)). Microsoft documents that Azure Resource Manager checks deny assignments first, and that a role assignment condition filters a grant by an attribute such as a tag ([Microsoft: Azure ABAC](https://learn.microsoft.com/en-us/azure/role-based-access-control/conditions-overview)). Google documents that Principal Access Boundary policies attach to sets of principals, and that by default principals are eligible to access any resource ([Google Cloud: Principal Access Boundary policies](https://cloud.google.com/iam/docs/principal-access-boundary-policies)). Write the grant from what the workload calls ([check: write the policy from the tools' calls](/checklist/least-privilege-policy/)), then grant at the narrowest scope that works ([check: narrowest scope](/checklist/narrowest-scope/)). ## Network: where the workload's calls can go AWS documents that a VPC endpoint policy is a resource-based policy controlling which principals can use an endpoint, and that it does not replace identity or resource policies ([AWS: VPC endpoint policies](https://docs.aws.amazon.com/vpc/latest/privatelink/vpc-endpoints-access.html)). Google documents VPC Service Controls as a perimeter that reduces data exfiltration risk from services such as Cloud Storage ([Google Cloud: VPC Service Controls](https://cloud.google.com/vpc-service-controls/docs/overview)). [Network reachability for AI agents](/guides/network-reachability-for-agents/) records which platforms allow the internet by default. The checklist groups its 30 checks under identity, permissions, data, network, tools, change control and logging. [The full checklist](/checklist/) lists them, and [the AWS page](/clouds/aws/) gives the AWS column. ## Related pages - [Agent authorization patterns: five ways to scope an agent](/agent-authorization-patterns/) - [Amazon Q Business default permissions: roles it creates](/default-permissions/amazon-q-business/) - [Copilot Studio default permissions: tool authentication](/default-permissions/copilot-studio/) - [All of them](/guides/) ## Sources - AWS: policy evaluation logic: https://docs.aws.amazon.com/IAM/latest/UserGuide/reference_policies_evaluation-logic.html (read 16 Nov 2026) - AWS: IAM permissions for AgentCore Runtime: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html (read 16 Nov 2026) - AWS: VPC endpoint policies: https://docs.aws.amazon.com/vpc/latest/privatelink/vpc-endpoints-access.html (read 16 Nov 2026) - Microsoft: Azure ABAC and role assignment conditions: https://learn.microsoft.com/en-us/azure/role-based-access-control/conditions-overview (read 16 Nov 2026) - Microsoft: agent identity concepts in Foundry: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 16 Nov 2026) - Google Cloud: agent identity on Agent Runtime: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 16 Nov 2026) - Google Cloud: Principal Access Boundary policies: https://cloud.google.com/iam/docs/principal-access-boundary-policies (read 16 Nov 2026) - Google Cloud: VPC Service Controls overview: https://cloud.google.com/vpc-service-controls/docs/overview (read 16 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/ai-access-control/ --- # Five agent authorization patterns scope an agent: a role per agent, a role per tool, a session policy, a token exchange and a conditioned resource policy Facts read 16 Nov 2026. Page reviewed 16 Nov 2026. Five authorization patterns scope an AI agent in a cloud account: a dedicated role per agent, a role per tool, session policies that scope down credentials, on-behalf-of token exchange, and resource-based policies with conditions. Each has a documented mechanism in AWS, Azure or Google Cloud, read on 16 Nov 2026. ## The five patterns at a glance | Pattern | AWS mechanism | Azure mechanism | Google Cloud mechanism | | --- | --- | --- | --- | | Dedicated role per agent | One IAM role per agent, trust scoped to the service and source ARN | One Entra agent identity per agent, created with the agent in Foundry | One agent identity principal per agent, instead of a shared service account | | Role per tool | A separate execution role for each Lambda function that implements a tool | RBAC assigned to the identity each tool uses | A service account chosen per tool (documented for Dialogflow CX tools) | | Session policy or scoped credentials | Session policy passed to AssumeRole, up to 10 managed policies | Role assignment at resource scope, with a condition | Scoped OAuth delegation for the user's access | | On-behalf-of token exchange | AgentCore Gateway token exchange; GetWorkloadAccessTokenForJWT | Entra on-behalf-of flow with delegated scopes | Delegated identity with 3-legged OAuth | | Resource policy with conditions | Resource-based and endpoint policies with condition keys | Azure ABAC role assignment conditions | Deny policies and IAM conditions | ## A dedicated role per agent The pattern: each agent runs as its own identity, so a grant to one does not reach another and the logs name it. On AWS, a Bedrock agent runs under a service role whose trust policy names bedrock.amazonaws.com ([AWS: service role for Bedrock Agents](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html)), and the Q Business plugin role example scopes its trust policy with aws:SourceAccount and aws:SourceArn conditions ([AWS: plugin role](https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/plugin-iam-role.html)). On Azure, Microsoft documents that creating a Foundry agent provisions a unique agent blueprint and an identity for its default instance. On Google Cloud, Google recommends the per-agent identity over the shared default service agent. See [the own-identity check](/checklist/own-identity-per-agent/) and [the trust check](/checklist/scope-trust-to-the-agent/). ## A role per tool The pattern: the agent's own identity can invoke tools but holds no data permissions, and each tool function runs under a narrow role of its own. AWS documents that a Lambda function runs under its execution role ([AWS: Lambda execution role](https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html)), so a tool built as a function can carry its own. Google documents that a Dialogflow CX tool can authenticate as a service account that you create and grant roles ([Google Cloud: Dialogflow CX tools](https://cloud.google.com/dialogflow/cx/docs/concept/playbook/tool)). The compromise of the agent then does not hand over what the tools hold. The cost is more roles to review. [The per-tool role check](/checklist/tool-own-role/) covers the setup, and [self-hosted agents on AWS](/agent-platforms/self-hosted-aws/) show the Lambda case. ## Session policies and scoped-down credentials The pattern: when the agent obtains temporary credentials, the request itself carries a smaller policy. AWS documents that session policies are passed as a parameter when creating a temporary session, that the session's permissions are the intersection of the role's identity-based policies and the session policies, and that a call can pass one inline document or up to 10 managed policy ARNs ([AWS: policies and permissions in IAM](https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies.html)). This lets an orchestrator hand each task only what that task needs. A per-call session policy parameter on Azure or Google Cloud: none found on 16 Nov 2026 in the pages read. The documented moves there are narrow scopes and conditions on the role assignment or binding. ## On-behalf-of token exchange The pattern: the agent trades the user's inbound token for a downstream token that carries both identities. AWS documents this as the token exchange grant in AgentCore Gateway: the exchanged token carries the user's identity and the agent's identity, so downstream services can authorize at every hop without more consent flows ([AWS: Gateway outbound authorization](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html)). Microsoft's on-behalf-of flow passes the user's identity and uses delegated scopes only ([Microsoft: on-behalf-of flow](https://learn.microsoft.com/en-us/entra/identity-platform/v2-oauth2-on-behalf-of-flow)). Google describes a delegated identity that combines the user and the agent ([Google Cloud: identity delegation architecture](https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/identity-delegation-architecture)). [The user-or-agent check](/checklist/decide-user-or-agent-authority/) decides which tools use it. ## Resource-based policies with conditions The pattern: the data side refuses calls that do not match, whatever the agent's role holds. AWS documents resource-based policies, including S3 bucket policies, and VPC endpoint policies that restrict which principals can use an endpoint ([AWS: VPC endpoint policies](https://docs.aws.amazon.com/vpc/latest/privatelink/vpc-endpoints-access.html)). Microsoft documents role assignment conditions that can require, for example, a blob to carry a specific tag ([Microsoft: Azure ABAC](https://learn.microsoft.com/en-us/azure/role-based-access-control/conditions-overview)). Google documents deny rules whose conditions decide whether a rule applies ([Google Cloud: deny policies](https://cloud.google.com/iam/docs/deny-overview)). See [the condition check](/checklist/condition-based-data-access/) and [the authorization page](/ai-agent-authorization/). ## Related pages - [Amazon Q Business default permissions: roles it creates](/default-permissions/amazon-q-business/) - [Copilot Studio default permissions: tool authentication](/default-permissions/copilot-studio/) - [Azure OpenAI Assistants default permissions and access](/default-permissions/azure-openai-assistants/) - [All of them](/guides/) ## Sources - AWS: policies and permissions in IAM: https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies.html (read 16 Nov 2026) - AWS: VPC endpoint policies: https://docs.aws.amazon.com/vpc/latest/privatelink/vpc-endpoints-access.html (read 16 Nov 2026) - AWS: AgentCore Gateway outbound authorization: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html (read 16 Nov 2026) - AWS: create a service role for Amazon Bedrock Agents: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html (read 16 Nov 2026) - AWS: IAM role for Amazon Q Business plugins: https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/plugin-iam-role.html (read 16 Nov 2026) - AWS: Lambda execution role: https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html (read 16 Nov 2026) - Microsoft: agent identity concepts in Foundry: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 16 Nov 2026) - Microsoft: Azure ABAC and role assignment conditions: https://learn.microsoft.com/en-us/azure/role-based-access-control/conditions-overview (read 16 Nov 2026) - Microsoft: OAuth 2.0 on-behalf-of flow: https://learn.microsoft.com/en-us/entra/identity-platform/v2-oauth2-on-behalf-of-flow (read 16 Nov 2026) - Google Cloud: agent identity on Agent Runtime: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 16 Nov 2026) - Google Cloud: identity delegation architecture: https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/identity-delegation-architecture (read 16 Nov 2026) - Google Cloud: deny policies: https://cloud.google.com/iam/docs/deny-overview (read 16 Nov 2026) - Google Cloud: Dialogflow CX tools: https://cloud.google.com/dialogflow/cx/docs/concept/playbook/tool (read 16 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/agent-authorization-patterns/ --- # Amazon Q Business runs on service roles the console creates, and each one is narrow but includes a few wildcard resources Facts read 16 Nov 2026. Page reviewed 16 Nov 2026. Amazon Q Business runs as separate IAM service roles for the application, web experience, data sources and plugins, which the console creates if you let it. AWS documents that the application role writes logs and metrics and the plugin role reads one secret. AWS also states Q Business is closed to new customers. Read 16 Nov 2026. ## Status first: Amazon Q Business is closed to new customers Each Amazon Q Business page that AWS publishes carries the note that it is no longer open to new customers, and points to Amazon Quick for similar capabilities ([AWS: what is Amazon Q Business](https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/what-is.html), read 16 Nov 2026). Existing applications still run, so the roles below matter for accounts that already have one. It is not one of the platforms in [the Agent Access Map](/map/). ## The identity it runs as is a set of service roles, not one role AWS documents that when you create an application or a web experience, or connect a data source, Q Business needs access to AWS resources, and that you supply an IAM role for each. With the console you can create new roles or pick existing ones, and the console lists roles with qbusiness or QBusiness in the name. If the console creates the roles, AWS documents that the listed policies are created on your behalf ([AWS: IAM roles for Amazon Q Business](https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/iam-roles.html)). The topics list six role types: application, web experience, data source connectors, plugins, custom document enrichment and the Amazon Kendra retriever. Q Business also answers on behalf of end users. AWS documents that its responses are based on the content the end user has permission to access, with access managed through AWS IAM Identity Center or IAM. That is a user-side control layered on top of the service roles. ## What the documented roles let it reach | Role | Documented permissions | Trust policy | | --- | --- | --- | | Application role | cloudwatch:PutMetricData limited to the AWS/QBusiness namespace; logs:DescribeLogGroups on all resources; log group, stream and event writes under /aws/qbusiness/* | Service principal qbusiness.amazonaws.com may call sts:AssumeRole | | Plugin service access role | secretsmanager:GetSecretValue on the one Secrets Manager secret holding the third-party credentials; optionally the KMS key that encrypts it | Service principal qbusiness.amazonaws.com, with aws:SourceAccount and aws:SourceArn conditions tying it to one application | | Web experience, data source, enrichment and retriever roles | Listed on their own AWS pages; the data source role depends on the connector | See each page | ## What that lets an attacker or a mistake reach The application role reaches logging and metrics, not data. The plugin role reaches the credential for the third-party system, such as Jira, Salesforce, ServiceNow or Zendesk, that the plugin was built for. AWS documents that plugins let users perform tasks such as changing the status of a ticket or creating an incident, up to 25 plugins per application ([AWS: configuring actions](https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/actions.html)). So the reach into a third-party service is set by the credential in that secret, and a broad credential there is a broad grant. The reach into AWS data comes from the data source roles and the retriever, which depend on what you connect. Two details to read in the policy text: logs:DescribeLogGroups is allowed on all resources, and the shown plugin trust example substitutes a placeholder for the secret ARN that you must replace with the real one. ## How to narrow it - Keep one set of roles per application, and add the source account and source ARN conditions to every trust policy; AWS shows them on the plugin role ([check: scope the trust](/checklist/scope-trust-to-the-agent/)). - Restrict the plugin secret's resource policy and the IAM permissions to read it, and store a third-party credential scoped to the one task the plugin needs ([check: write the policy from the calls](/checklist/least-privilege-policy/)). - Attach a permissions boundary or an SCP over the Q Business roles so nothing can widen them ([check: set a ceiling](/checklist/set-permission-ceiling/)). - Review data source roles against the buckets and sites the connector truly indexes, using [the least privilege guide](/guides/least-privilege-for-ai-agents/). - Read [the AWS cloud page](/clouds/aws/) for the other AWS platforms and their defaults. ## Related pages - [Copilot Studio default permissions: tool authentication](/default-permissions/copilot-studio/) - [Azure OpenAI Assistants default permissions and access](/default-permissions/azure-openai-assistants/) - [Conversational Agents (Dialogflow CX) default permissions](/default-permissions/conversational-agents/) - [All of them](/guides/) ## Sources - AWS: what is Amazon Q Business: https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/what-is.html (read 16 Nov 2026) - AWS: IAM roles for Amazon Q Business: https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/iam-roles.html (read 16 Nov 2026) - AWS: IAM role for an Amazon Q Business application: https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/create-application-iam-role.html (read 16 Nov 2026) - AWS: IAM role for Amazon Q Business plugins: https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/plugin-iam-role.html (read 16 Nov 2026) - AWS: configuring actions in Amazon Q Business: https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/actions.html (read 16 Nov 2026) - AWS: identity and access management for Amazon Q Business: https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/security-iam.html (read 16 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/default-permissions/amazon-q-business/ --- # A Copilot Studio agent reaches data through the credential each tool uses, either the maker's or the signed-in user's Facts read 16 Nov 2026. Page reviewed 16 Nov 2026. In Microsoft Copilot Studio an agent's reach depends on how each tool authenticates: agent author authentication, which Microsoft recommends for low-risk or implicit access, or user authentication, which uses the signed-in user's connection. Microsoft documents data policies in the Power Platform admin center as the admin control. Documentation read on 16 Nov 2026. ## Copilot Studio is a low-code builder, so the identity is per tool Microsoft describes Copilot Studio as a graphical, low-code studio for building and managing AI agents and workflows, which you connect to your organization's data and systems and publish to channels such as Microsoft Teams ([Microsoft: Copilot Studio overview](https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio), read 16 Nov 2026). Unlike a Bedrock or Foundry agent, there is no single execution role or service principal in the documentation pages read. The credential is chosen on each tool. Microsoft also documents that some agents can be given their own account, and that for organizations that onboard Microsoft Agent 365, Copilot Studio agents can be represented as identities in Microsoft Entra and governed with Conditional Access and role-based and attribute-based access controls. The default identity before onboarding Agent 365: none found on 16 Nov 2026 in the pages read. ## The two authentication choices for a tool | Choice | What Microsoft documents | What it reaches | | --- | --- | --- | | Agent author authentication | For access to the service that is implicit or for low-risk use cases, such as a weather API | Whatever the credential supplied for the agent can reach, for every user who talks to the agent (a reading of the name, not a Microsoft statement) | | User authentication | For restricting data access to specific groups or individuals, or doing work on the user's behalf | Only what the signed-in user's connection can reach; Copilot Studio checks for a valid connection and shows a connection card if none exists | ## The default posture Microsoft lists, among its security and governance controls, that makers can configure tools to use the user's credentials by default ([Microsoft: security and governance](https://learn.microsoft.com/en-us/microsoft-copilot-studio/security-and-governance)). It lists makers' security warnings when default configurations are modified, and an automatic security scan with real-time risk assessment while configuring knowledge, tools and actions. The exact default selection for a new tool: none found on 16 Nov 2026 in the pages read, so read it in each tool's settings. The practical reading: an agent author tool is a shared credential. It behaves like a service account that every chat user borrows. A user authentication tool carries the user's own reach, which is the pattern in [why an agent's permissions are not its user's](/guides/why-an-agents-permissions-are-not-its-users/). ## How an administrator narrows it - Data policies in the Power Platform admin center govern maker and user authentication, knowledge sources, actions and connectors, HTTP requests, publication to channels and triggers, per Microsoft's list. - Environment routing gives makers a separate environment to build in. - The tenant admin can disable the publishing of agents that use generative AI features. - Audit logs for maker activity are available in Microsoft Purview, and activity alerts in Microsoft Sentinel. - With Agent 365, Microsoft documents Entra network egress and ingress controls for Copilot Studio agents. ## What this page leaves out User authentication for tools is supported on some channels only; Microsoft lists Teams, SharePoint, a custom website and Omnichannel live chat as supported, and Azure Bot Service channels, the demo website, Facebook and the mobile app as not supported ([Microsoft: user authentication for tools](https://learn.microsoft.com/en-us/microsoft-copilot-studio/configure-enduser-authentication)). On an unsupported channel a tool cannot use the user's own connection, which pushes it toward a maker credential. Copilot Studio reaches Azure resources only through connectors, so the checks in [the Azure cloud page](/clouds/azure/) apply to the identity behind each connector. Compare the Foundry model on [the Foundry Agent Service page](/agent-platforms/foundry-agent-service/). ## Related pages - [Azure OpenAI Assistants default permissions and access](/default-permissions/azure-openai-assistants/) - [Conversational Agents (Dialogflow CX) default permissions](/default-permissions/conversational-agents/) - [The identity an AI agent runs as in AWS, Google Cloud and Azure](/guides/the-identity-an-agent-runs-as/) - [All of them](/guides/) ## Sources - Microsoft: Copilot Studio overview: https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio (read 16 Nov 2026) - Microsoft: Copilot Studio security and governance: https://learn.microsoft.com/en-us/microsoft-copilot-studio/security-and-governance (read 16 Nov 2026) - Microsoft: configure user authentication for tools: https://learn.microsoft.com/en-us/microsoft-copilot-studio/configure-enduser-authentication (read 16 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/default-permissions/copilot-studio/ --- # Azure OpenAI Assistants has no identity of its own: callers use a key or an Entra role on the resource, and the API is retired Facts read 16 Nov 2026. Page reviewed 16 Nov 2026. Azure OpenAI Assistants (Preview) is documented by Microsoft as retired in favor of Foundry Agent Service. The quickstart authenticates with an API key; Entra ID with the Cognitive Services OpenAI User role is the documented alternative. The assistant itself holds no cloud role, so its reach is its tools. Read 16 Nov 2026. ## Status first: the Assistants API is retired Microsoft's page on getting started with Azure OpenAI Assistants (Preview) carries the note that the Assistants API is retired, that the generally available Microsoft Foundry Agents service replaces it, and that a migration guide exists ([Microsoft: getting started with Azure OpenAI Assistants (classic)](https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/assistant), read 16 Nov 2026). The page applies only to the Foundry (classic) portal. Accounts that still run assistants should treat the access below as something to move or remove. The replacement is on [the Foundry Agent Service page](/agent-platforms/foundry-agent-service/). ## The identity: the caller's, because the assistant has none In the documented quickstart the application creates the client with an API key read from an environment variable. An assistant object is created through the API with instructions, a model deployment and tools. Nothing in the pages read describes a managed identity or service principal assigned to the assistant. Microsoft documents that the Azure OpenAI resource can instead be called with Microsoft Entra ID: your code requests a short-lived token for an identity that holds a role on the resource, which can be your developer account, a system-assigned or user-assigned managed identity, or a service principal ([Microsoft: Entra ID authentication for Azure OpenAI](https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/managed-identity)). An assistant-specific role: none found on 16 Nov 2026. ## What the defaults let a caller do | Credential | What it allows on the resource | Documented limit | | --- | --- | --- | | Resource API key | Calls to the resource's endpoints, including the Assistants API; the quickstart authenticates this way | Microsoft lists viewing and copying keys as a permission separate from the OpenAI User and Contributor roles | | Cognitive Services OpenAI User role | Chat, completions and inference API calls with Entra ID; view deployments | Cannot create resources or deployments, copy keys, or create customized guardrails | | Cognitive Services OpenAI Contributor role | Everything in the User role, plus fine-tuning, stored completions and creating or editing deployments | Cannot create resources, view keys or access quota | | Cognitive Services Contributor role | Create resources in the resource group; view and regenerate keys | Typically granted at the resource group with other roles | ## What the assistant's tools reach Microsoft documents that an individual assistant can access up to 128 tools, including code interpreter and file search, and custom functions. Code interpreter gives the model a sandboxed Python environment. Custom functions are tools you define, so what they reach depends on the code and the identity behind them. The cloud reach of an assistant is therefore the files you attach plus whatever those functions' identity can reach. See [the identity guide](/guides/the-identity-an-agent-runs-as/) for that second identity. ## How to narrow it - Call the resource with Microsoft Entra ID and a role of Cognitive Services OpenAI User, assigned at the resource scope rather than the subscription ([check: narrowest scope](/checklist/narrowest-scope/)). - Microsoft notes that subscription-level Owner and Contributor roles are inherited and take priority over the custom Azure OpenAI roles applied at the resource group, so check for them first. - Give each application its own managed identity so calls can be traced ([check: attribute actions to the agent](/checklist/attribute-actions-to-the-agent/)). - Give any function host its own role limited to what that function does ([check: a role per tool](/checklist/tool-own-role/)). - Plan the migration to Foundry Agent Service, where a new agent identity starts with no roles, as [the Azure agent permissions guide](/guides/ai-agent-permissions-in-azure/) describes. ## Related pages - [Conversational Agents (Dialogflow CX) default permissions](/default-permissions/conversational-agents/) - [The identity an AI agent runs as in AWS, Google Cloud and Azure](/guides/the-identity-an-agent-runs-as/) - [Why an AI agent's permissions are not its user's permissions](/guides/why-an-agents-permissions-are-not-its-users/) - [All of them](/guides/) ## Sources - Microsoft: getting started with Azure OpenAI Assistants (classic): https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/assistant (read 16 Nov 2026) - Microsoft: Azure RBAC for Azure OpenAI (classic): https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/role-based-access-control (read 16 Nov 2026) - Microsoft: Microsoft Entra ID authentication for Azure OpenAI (classic): https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/managed-identity (read 16 Nov 2026) - Microsoft: agent identity concepts in Foundry: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 16 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/default-permissions/azure-openai-assistants/ --- # A Conversational Agents agent calls tools as the Dialogflow service agent unless you pick a service account, and the project owner sees everything Facts read 16 Nov 2026. Page reviewed 16 Nov 2026. Conversational Agents (Dialogflow CX) calls tools and webhooks as the Dialogflow CX Service Agent, or as a service account you name. Google documents that the project owner has full access to every agent in the project and that sharing an agent adds the Dialogflow Reader role by default. Read 16 Nov 2026. ## Two directions of access: who reaches the agent, and what the agent reaches A Dialogflow CX agent has two access questions. Who may build, query or share the agent is an IAM question about people and applications. What the agent reaches when it calls a tool or webhook depends on the credential the tool uses. Google's access control page covers the first; the tool and webhook pages cover the second ([Google Cloud: Dialogflow CX access control](https://cloud.google.com/dialogflow/cx/docs/concept/access-control), read 16 Nov 2026). The Conversational Agents console is the newer interface that replaces the Dialogflow CX console. ## Who can reach the agent by default Google documents that the IAM Project Owner role is granted by default to the user who created the project that owns the agent, and that this owner has full access to all agents in the project. Agent sharing in the console adds agent roles, and the Dialogflow Reader role is added by default to a shared principal, which users need to open the Conversational Agents console. Setting agent roles in the console needs the Project IAM Admin role. The agent roles are IAM roles with IAM conditions that limit them to one agent or part of it. Google lists Conversational Reader, Conversational Admin and Conversational Client, among others, mapping to Dialogflow API Reader, Admin and Client. A service account added to the project can call the agent from an application, which is the way to give an application detect-intent access without admin rights. ## Which identity the agent's tools and webhooks use | Option | What Google documents | What it reaches | | --- | --- | --- | | Dialogflow CX Service Agent ID token | Dialogflow generates an ID token as the service agent, an address at gcp-sa-dialogflow.iam.gserviceaccount.com, and adds it to the Authorization header | Cloud Run and Cloud Run functions, after you grant roles/run.invoker or roles/cloudfunctions.invoker to the service agent; same-project services need no extra permission | | Service account | Dialogflow generates an access token for a service account you create; you need roles/iam.serviceAccountUser, and the service agent needs roles/iam.serviceAccountTokenCreator | Any Google API the service account has roles for | | API key | Key name, location and value are provided in the tool | The service the key opens | | OAuth client credentials | Server-to-server; Google recommends storing the client secret in Secret Manager | The third-party API the client may call | ## What the defaults imply The service agent's own default permissions: none found on 16 Nov 2026 in the pages read. What Google documents is that the tool authenticates as the service agent only for Cloud Run and Cloud Run functions where you grant the invoker role, and that a service account tool is limited to the roles of that account. For webhooks, Google documents a priority order when several authentication options are enabled: authentication headers, basic auth, third-party OAuth, then service agent authentication ([Google Cloud: Dialogflow CX webhooks](https://cloud.google.com/dialogflow/cx/docs/concept/webhook)). A webhook that carries a header credential will use it ahead of the service agent. ## How to narrow it - Use a dedicated service account per tool with only the roles that tool needs, and grant TokenCreator only on that account ([check: a role per tool](/checklist/tool-own-role/)). - Grant run.invoker on the single Cloud Run service, not the project ([check: restrict tool invokers](/checklist/restrict-tool-invokers/)). - Give people Conversational Reader or Client on one agent through the sharing tab, and keep Project Owner rare. - Where service accounts cross projects, Google notes that the organization policy must support it; review the constraint ([check: scope the trust](/checklist/scope-trust-to-the-agent/)). - Compare the Gemini platforms on [the Agent Runtime page](/agent-platforms/agent-runtime-gcp/) and [the Google Cloud page](/clouds/gcp/). ## Related pages - [The identity an AI agent runs as in AWS, Google Cloud and Azure](/guides/the-identity-an-agent-runs-as/) - [Why an AI agent's permissions are not its user's permissions](/guides/why-an-agents-permissions-are-not-its-users/) - [Network reachability for AI agents on AWS, Google Cloud and Azure](/guides/network-reachability-for-agents/) - [All of them](/guides/) ## Sources - Google Cloud: Dialogflow CX access control: https://cloud.google.com/dialogflow/cx/docs/concept/access-control (read 16 Nov 2026) - Google Cloud: Dialogflow CX tools: https://cloud.google.com/dialogflow/cx/docs/concept/playbook/tool (read 16 Nov 2026) - Google Cloud: Dialogflow CX webhooks: https://cloud.google.com/dialogflow/cx/docs/concept/webhook (read 16 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/default-permissions/conversational-agents/ --- # 29 terms, each with a sourced example Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Plain definitions of the terms in the map and checklist, from execution role and agent identity to permissions boundary, Principal Access Boundary and data event. Each entry has a sourced example from a cloud provider page read on 10 Nov 2026. - [Action group](/glossary/action-group/): In Amazon Bedrock Agents, the set of actions an agent may perform, defined by an API schema or function details and usually backed by a Lambda function. - [Agent runtime](/glossary/agent-runtime/): A managed cloud service that hosts agent code, runs it for each session and gives it an identity and a network position. - [Agent tool](/glossary/agent-tool/): A function, API or service that an agent can call to act or to fetch data. - [AI agent](/glossary/ai-agent/): A program that uses a language model to decide which actions to take, then carries them out by calling tools, APIs and data sources on its own. - [AI workload](/glossary/ai-workload/): Any program running in a cloud account whose job involves a model: an agent, a retrieval service, a batch scoring job or an inference endpoint. - [AWS PrivateLink](/glossary/aws-privatelink/): The AWS technology behind interface VPC endpoints. - [Cloud Audit Logs](/glossary/cloud-audit-logs/): Google Cloud's audit trail of who did what, where and when. - [CloudTrail data events](/glossary/cloudtrail-data-events/): AWS CloudTrail records of operations on or inside a resource, such as reading an S3 object or invoking a function, as opposed to management events that record changes to configuration. - [Confused deputy](/glossary/confused-deputy/): A security problem where something without permission to do an action tricks a more-privileged party into doing it. - [Data access audit logs](/glossary/data-access-audit-logs/): The Google Cloud audit logs that record reads and writes of user-provided data. - [Egress control](/glossary/egress-control/): Limiting where outbound traffic from a workload may go, by destination, protocol or domain. - [EKS Pod Identity](/glossary/eks-pod-identity/): An Amazon EKS feature that maps an IAM role to a Kubernetes service account in a namespace, so the pods using that service account receive the role's temporary credentials. - [Execution role](/glossary/execution-role/): An IAM role that a function or agent runtime assumes when it runs. - [IAM roles for service accounts](/glossary/iam-roles-for-service-accounts/): The earlier Amazon EKS method, often shortened to IRSA, for giving a pod an IAM role. - [Identity propagation](/glossary/identity-propagation/): Carrying who is asking, and who is acting, through each hop between a user, an agent and the services it calls. - [Knowledge base](/glossary/knowledge-base/): In Amazon Bedrock, a store of your documents that an agent can query for extra context when it builds an answer. - [Managed identity](/glossary/managed-identity/): An Azure identity assigned to a compute resource so that the resource can get Microsoft Entra tokens without any stored secret. - [Model invocation logging](/glossary/model-invocation-logging/): An Amazon Bedrock setting that records the full request data, response data and metadata of model calls in an account and Region. - [On-behalf-of flow](/glossary/on-behalf-of-flow/): An authorization pattern where a service exchanges the token of the signed-in user for a new token aimed at a downstream resource, so the downstream call is judged against the user. - [Private endpoint](/glossary/private-endpoint/): An Azure network interface with a private IP address from your virtual network that connects privately to a service on Azure Private Link. - [Private Service Connect](/glossary/private-service-connect/): A Google Cloud networking capability that lets consumers reach managed services privately from their own VPC networks using internal IP addresses. - [Resource-based policy](/glossary/resource-based-policy/): A permission policy attached to the resource being accessed, such as a Lambda function, an S3 bucket or a KMS key, naming who may use it. - [Retrieval augmented generation](/glossary/retrieval-augmented-generation/): A pattern where a system looks up relevant documents first and gives them to a language model along with the question, so the answer draws on those documents. - [Service account](/glossary/service-account/): A non-human identity in Google Cloud that a workload runs as. - [Service agent](/glossary/service-agent/): A Google-managed service account that a Google Cloud service uses to act in your project, such as the account Agent Runtime uses by default. - [Service role](/glossary/service-role/): An IAM role that an AWS service assumes to act in your account on your behalf. - [VPC endpoint](/glossary/vpc-endpoint/): A private connection from an AWS VPC to an AWS service or another service, without a route through an internet gateway. - [Workload identity](/glossary/workload-identity/): An identity held by software instead of a person. - [Workload identity federation](/glossary/workload-identity-federation/): A way for software running outside a cloud, or in another system, to trade a token from its own identity provider for short-lived cloud credentials. ## Related pages - [News](/news/) - [Agent platforms in AWS, Google Cloud and Azure](/agent-platforms/) - [Tools scored on AI access coverage](/tools/) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/ --- # AI agent Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. A program that uses a language model to decide which actions to take, then carries them out by calling tools, APIs and data sources on its own. In a cloud account it is a workload like any other: it runs as an identity, holds permissions, and every call it makes is judged against them. ## An example, sourced AWS documents how an Amazon Bedrock agent works. Its orchestration step interprets the user input, invokes action groups and queries knowledge bases, then returns output to the user or feeds it into the next step of the orchestration. In each cycle the agent predicts which action in an action group it should invoke or which knowledge base it should query. The model\'s choices become real only through those calls, and each call is a request that the cloud account authorizes or refuses. ## Why it matters here A reviewer cannot approve behavior in the abstract. The thing to approve is the identity, the tools and the network path behind the agent. Treating the agent as a workload turns a vague risk into a list of permissions that can be read and cut. ## Related pages - [AI workload: meaning and an example](/glossary/ai-workload/) - [Execution role: meaning and an example](/glossary/execution-role/) - [Service role: meaning and an example](/glossary/service-role/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/bedrock/latest/userguide/agents-how.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/ai-agent/ --- # AI workload Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Any program running in a cloud account whose job involves a model: an agent, a retrieval service, a batch scoring job or an inference endpoint. It gets an identity, permissions, network access and logs like other compute, and it is reviewed on those, not on the model it calls. ## An example, sourced Microsoft documents hosted agents in Foundry Agent Service as workloads you package yourself. You build the agent as a container image and push it to Azure Container Registry. When you deploy, Agent Service pulls the image, assigns a dedicated Microsoft Entra identity, and exposes a dedicated endpoint. Microsoft says you own the code that runs inside the sandbox, while the platform owns the endpoint, identity, scaling and session state around it. ## Why it matters here Splitting responsibility this way tells a reviewer where to look. The platform sets the identity and the endpoint. The builder's code decides what the agent does with them. Both halves need an answer before production. Both halves of that split need an owner named in the review. ## Related pages - [Execution role: meaning and an example](/glossary/execution-role/) - [Service role: meaning and an example](/glossary/service-role/) - [Service account: meaning and an example](/glossary/service-account/) - [All of them](/glossary/) ## Sources - https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/hosted-agents (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/ai-workload/ --- # Execution role Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. An IAM role that a function or agent runtime assumes when it runs. The role's policies decide which AWS APIs the code can call, and the service assumes the role automatically on each invocation, so the code holds no long-lived key. ## An example, sourced AWS documents the Lambda execution role as the role that Lambda automatically assumes when you invoke your function. The page, titled as defining function permissions with an execution role, also covers creating the role in the IAM console and granting least privilege access to it. AWS documents that when you create a function in the Lambda console, Lambda creates an execution role with minimal permissions, covering basic logging to CloudWatch Logs. Anything the function code does beyond that depends on policies someone attached. ## Why it matters here For a function-based or AgentCore agent, the execution role is the agent's whole reach into AWS. Reading its attached policies is the first step of a review, and a role created by a console wizard still needs reading. A role created by a wizard may also have been widened since. ## Related pages - [Service role: meaning and an example](/glossary/service-role/) - [Service account: meaning and an example](/glossary/service-account/) - [Service agent: meaning and an example](/glossary/service-agent/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/execution-role/ --- # Service role Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. An IAM role that an AWS service assumes to act in your account on your behalf. The role has a trust policy naming the service and permission policies listing what the service may do, and it is either created for you by a console flow or written by you. ## An example, sourced AWS documents a service role for Amazon Bedrock Agents. You can use a custom service role instead of the one Amazon Bedrock automatically creates. The documented identity-based policy gives Bedrock access to the base models, to the S3 objects holding the action group schemas, and to the knowledge bases attached to the agent, with optional statements for items such as guardrails and KMS decryption. ## Why it matters here The service role is the identity of a Bedrock Classic agent. Whoever can edit it can change what the agent reaches, so the people allowed to pass or modify it belong in the review as well as the policy itself. ## Related pages - [Service account: meaning and an example](/glossary/service-account/) - [Service agent: meaning and an example](/glossary/service-agent/) - [Managed identity: meaning and an example](/glossary/managed-identity/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/service-role/ --- # Service account Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. A non-human identity in Google Cloud that a workload runs as. You grant it IAM roles, attach it to a resource such as a Cloud Run service, and the workload calls Google Cloud APIs as that account. Several workloads can share one, which makes the sharing a review question. ## An example, sourced Google documents user-managed service accounts that you create in a project, and default service accounts that Google Cloud creates for you. For Cloud Run, Google documents that the Compute Engine default service account is the default service identity if you pick none. Google also documents that, depending on organization policy, that account might automatically be granted the Editor role on the project, and recommends replacing Editor with less permissive roles. ## Why it matters here An agent that inherits a default service account inherits access nobody chose for it. Naming the account in the review, and listing its roles, is how that inheritance is found. The roles on the account, not the account's name, set the real reach. ## Related pages - [Service agent: meaning and an example](/glossary/service-agent/) - [Managed identity: meaning and an example](/glossary/managed-identity/) - [Workload identity: meaning and an example](/glossary/workload-identity/) - [All of them](/glossary/) ## Sources - https://cloud.google.com/run/docs/securing/service-identity (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/service-account/ --- # Service agent Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. A Google-managed service account that a Google Cloud service uses to act in your project, such as the account Agent Runtime uses by default. Google creates and manages it, and your own access policy decides whether it can reach your resources. ## An example, sourced Google documents service agents as accounts that Google Cloud creates and manages, separate from the user-managed service accounts you create in a project. Google states that service agents are not created in your projects, and that by default they are not listed on the IAM page of the console, where you must choose to include Google-provided role grants. Google adds that both the creation time and the email address format for service agents are subject to change, so a policy should not hard-code them. ## Why it matters here A default you cannot see in the IAM page is a default a reviewer may miss. Ask for the service agent's role grants by name when an agent uses it. It also helps to know which project the agent runs in, because the grants are made per project. ## Related pages - [Managed identity: meaning and an example](/glossary/managed-identity/) - [Workload identity: meaning and an example](/glossary/workload-identity/) - [Workload identity federation: meaning and an example](/glossary/workload-identity-federation/) - [All of them](/glossary/) ## Sources - https://cloud.google.com/iam/docs/service-account-types (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/service-agent/ --- # Managed identity Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. An Azure identity assigned to a compute resource so that the resource can get Microsoft Entra tokens without any stored secret. It is either system-assigned, tied to one resource, or user-assigned, a standalone resource that several resources can use. Role assignments decide what it can reach. ## An example, sourced Microsoft documents that managed identities eliminate the need for developers to manage credentials, and that applications use them to obtain Microsoft Entra tokens. A system-assigned identity is enabled on one resource, and its service principal always carries the same name as that resource. A user-assigned identity is created as a standalone resource and can be assigned to more than one. Either kind can then be authorized to reach downstream resources, such as a storage account, a SQL database or Cosmos DB. ## Why it matters here Managed identities remove the leaked-key risk but not the question of scope. Whether an agent shares one with other apps, and what roles it holds, still has to be read from the role assignments. The reviewer also needs the list of resources the identity can be assigned to. ## Related pages - [Workload identity: meaning and an example](/glossary/workload-identity/) - [Workload identity federation: meaning and an example](/glossary/workload-identity-federation/) - [EKS Pod Identity: meaning and an example](/glossary/eks-pod-identity/) - [All of them](/glossary/) ## Sources - https://learn.microsoft.com/en-us/entra/identity/managed-identities-azure-resources/overview (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/managed-identity/ --- # Workload identity Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. An identity held by software instead of a person. Cloud providers issue them to applications, containers and functions so that code can authenticate to other services. The type differs by provider (role, service account, service principal), but the review question is the same: which one, and what can it reach. ## An example, sourced Microsoft Entra documents that machine identities consist of device and workload identities, and that in Microsoft Entra, workload identities are applications, service principals and managed identities. Microsoft describes a managed identity as an identity assigned to an Azure compute resource, such as a virtual machine, a scale set, a Service Fabric cluster or an Azure Kubernetes cluster. Once assigned, it can be authorized, directly or indirectly, to reach downstream resources such as a storage account, a SQL database or Cosmos DB. ## Why it matters here The term lets a reviewer ask one question across clouds: which workload identity does this agent present? The answer then maps to a role, a service account or a service principal on each platform. It also shows which products are in scope for the identity inventory. ## Related pages - [Workload identity federation: meaning and an example](/glossary/workload-identity-federation/) - [EKS Pod Identity: meaning and an example](/glossary/eks-pod-identity/) - [IAM roles for service accounts: meaning and an example](/glossary/iam-roles-for-service-accounts/) - [All of them](/glossary/) ## Sources - https://learn.microsoft.com/en-us/entra/identity/managed-identities-azure-resources/overview (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/workload-identity/ --- # Workload identity federation Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. A way for software running outside a cloud, or in another system, to trade a token from its own identity provider for short-lived cloud credentials. The cloud trusts the provider, so the workload holds no stored key. Google Cloud, Microsoft Entra and AWS each offer a version. ## An example, sourced Microsoft documents workload identity federation as configuring a user-assigned managed identity or app registration in Microsoft Entra ID to trust tokens from an external identity provider, such as GitHub or Google. Once the trust relationship exists, the external software workload exchanges trusted tokens from that provider for access tokens from the Microsoft identity platform. Microsoft lists GitHub Actions and Kubernetes workloads as scenarios. ## Why it matters here Federation is the usual way to remove keys from agents that run outside the cloud they call. A reviewer should check which external subject is trusted, because a broad trust lets any workload from that provider in. Ask also how long the exchanged credentials last. ## Related pages - [EKS Pod Identity: meaning and an example](/glossary/eks-pod-identity/) - [IAM roles for service accounts: meaning and an example](/glossary/iam-roles-for-service-accounts/) - [Action group: meaning and an example](/glossary/action-group/) - [All of them](/glossary/) ## Sources - https://learn.microsoft.com/en-us/entra/workload-id/workload-identity-federation (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/workload-identity-federation/ --- # EKS Pod Identity Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. An Amazon EKS feature that maps an IAM role to a Kubernetes service account in a namespace, so the pods using that service account receive the role's temporary credentials. It lets two agents in one cluster hold different AWS permissions. ## An example, sourced AWS documents that each EKS Pod Identity association maps a role to a service account in a namespace in a specified cluster. If you run the same application in several clusters, AWS says you can make identical associations in each cluster without modifying the trust policy of the role. AWS states that a container never has access to credentials that are used by other containers in other pods, and that Pod Identity works unchanged with isolated and confidential pods that run in Nitro Sandboxes. ## Why it matters here This is the control that gives each self-hosted agent on EKS its own role. Without an association per service account, agents in a cluster tend to share the node's access. The trust policy on each role should name the service account and namespace. ## Related pages - [IAM roles for service accounts: meaning and an example](/glossary/iam-roles-for-service-accounts/) - [Action group: meaning and an example](/glossary/action-group/) - [Agent tool: meaning and an example](/glossary/agent-tool/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/eks/latest/userguide/pod-identities.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/eks-pod-identity/ --- # IAM roles for service accounts Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. The earlier Amazon EKS method, often shortened to IRSA, for giving a pod an IAM role. A Kubernetes service account is linked to a role, and the pod's containers use the role's temporary credentials through the AWS SDK. EKS Pod Identity is the newer option. ## An example, sourced AWS documents that applications in a pod's containers can use an AWS SDK or the AWS CLI to make API requests to AWS services using IAM permissions, and that they must sign requests with AWS credentials. AWS describes IAM roles for service accounts as providing the ability to manage credentials for your applications, similar to the way Amazon EC2 instance profiles provide credentials to EC2 instances. The credentials come through the role instead of a stored key. ## Why it matters here Clusters often carry both methods. A reviewer has to find which one the agent uses, because the trust policy to read differs: OIDC conditions for one, session tags for the other. Both can coexist in a cluster, and the older one is easier to leave in place by accident. ## Related pages - [Action group: meaning and an example](/glossary/action-group/) - [Agent tool: meaning and an example](/glossary/agent-tool/) - [Agent runtime: meaning and an example](/glossary/agent-runtime/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/eks/latest/userguide/iam-roles-for-service-accounts.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/iam-roles-for-service-accounts/ --- # Action group Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. In Amazon Bedrock Agents, the set of actions an agent may perform, defined by an API schema or function details and usually backed by a Lambda function. Each action group is a tool the agent can call, so its backing function's permissions are part of the agent's reach. ## An example, sourced AWS documents that an action group consists of an OpenAPI schema and a Lambda function that define what API operations an agent can call and how the agent should handle requests and responses. Each action group can use a different schema, and the schema holds the parameters that the agent must elicit from the user. During orchestration the agent predicts which action in an action group it should invoke. The Lambda function behind each action is therefore part of the agent\'s reach, along with the agent\'s own role. ## Why it matters here An action group adds a second identity to an agent: the function's role. Reviewers should read that role as well as the agent's own, since a tool can reach more than the agent was meant to. The invocation policy on the function matters too, since it says who may trigger the tool. ## Related pages - [Agent tool: meaning and an example](/glossary/agent-tool/) - [Agent runtime: meaning and an example](/glossary/agent-runtime/) - [On-behalf-of flow: meaning and an example](/glossary/on-behalf-of-flow/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/bedrock/latest/userguide/agents-how.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/action-group/ --- # Agent tool Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. A function, API or service that an agent can call to act or to fetch data. The tool runs under some identity, which may be the agent's, the user's or its own, and the choice sets what a call can reach and whom the logs name. ## An example, sourced AWS documents that Amazon Bedrock AgentCore Gateway converts APIs, Lambda functions and existing services into Model Context Protocol (MCP) compatible tools. AWS adds that Gateway fronts other agents and HTTP services through passthrough targets, including Agent-to-Agent traffic, and routes inference requests across model providers through one endpoint. A tool in this sense is any target the agent can reach through the gateway, so each target is a path from the agent to something else. ## Why it matters here Every tool is a path from the agent to something else. A review lists the tools and records which identity each one runs as, since a narrow agent role does not narrow a tool that runs under a broad one. ## Related pages - [Agent runtime: meaning and an example](/glossary/agent-runtime/) - [On-behalf-of flow: meaning and an example](/glossary/on-behalf-of-flow/) - [Identity propagation: meaning and an example](/glossary/identity-propagation/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/agent-tool/ --- # Agent runtime Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. A managed cloud service that hosts agent code, runs it for each session and gives it an identity and a network position. Google Cloud calls its service Agent Runtime, and AWS has Amazon Bedrock AgentCore Runtime. The runtime decides the default identity and network path. ## An example, sourced Google documents that on Agent Runtime, the product formerly called Vertex AI Agent Engine, using agent identity provides a secure, per-agent identity, tied to the lifecycle of the agent, that enables a least-privilege approach. Google documents two other choices: the Google-managed Reasoning Engine Service Agent, used by default, and a custom service account. Which one an agent uses sets the starting roles it holds. ## Why it matters here Product names change, as the Vertex AI Agent Engine rename shows. A reviewer should name the runtime and its identity setting in the record, not just the product label from the vendor deck. It also tells the reviewer which logs and network settings to look for. ## Related pages - [On-behalf-of flow: meaning and an example](/glossary/on-behalf-of-flow/) - [Identity propagation: meaning and an example](/glossary/identity-propagation/) - [Model invocation logging: meaning and an example](/glossary/model-invocation-logging/) - [All of them](/glossary/) ## Sources - https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/agent-runtime/ --- # On-behalf-of flow Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. An authorization pattern where a service exchanges the token of the signed-in user for a new token aimed at a downstream resource, so the downstream call is judged against the user. The new token can carry both the user's identity and the calling agent's identity. ## An example, sourced AWS documents the token exchange grant (On-behalf-of) for Amazon Bedrock AgentCore Gateway. The gateway exchanges the inbound user's access token for a new, scoped access token that targets a downstream resource. AWS says the exchanged token carries both the user's identity and the agent's identity, so downstream services can enforce fine-grained authorization at every hop without triggering additional consent flows. ## Why it matters here This is the mechanism that keeps an agent from acting with more authority than the person asking. A reviewer should ask which tools use it and which fall back to the agent's own role. Flows like this are newer than plain service roles, so confirm the feature status in the provider page. ## Related pages - [Identity propagation: meaning and an example](/glossary/identity-propagation/) - [Model invocation logging: meaning and an example](/glossary/model-invocation-logging/) - [CloudTrail data events: meaning and an example](/glossary/cloudtrail-data-events/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/on-behalf-of-flow/ --- # Identity propagation Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Carrying who is asking, and who is acting, through each hop between a user, an agent and the services it calls. Propagation lets a downstream service apply the right authority and lets the logs name both parties. Where it is missing, calls arrive as the agent alone. ## An example, sourced Google documents a three-identity framework for agents: the user identity establishes caller context at ingress, the agent identity authenticates the workload within the runtime boundary, and a delegated identity combines both authorities when the agent accesses downstream services on behalf of the user. Google's architecture page describes how each identity layer propagates across the trust boundaries on the way. Google places the user, the agent and the delegated identity in separate trust boundaries. ## Why it matters here Propagation decides whether a downstream log line names the person, the agent or only a shared role. If a platform does not propagate, the review should say so and rely on controls elsewhere. Treat a missing link in the chain as a finding and not as a detail. ## Related pages - [Model invocation logging: meaning and an example](/glossary/model-invocation-logging/) - [CloudTrail data events: meaning and an example](/glossary/cloudtrail-data-events/) - [Cloud Audit Logs: meaning and an example](/glossary/cloud-audit-logs/) - [All of them](/glossary/) ## Sources - https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/identity-delegation-architecture (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/identity-propagation/ --- # Model invocation logging Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. An Amazon Bedrock setting that records the full request data, response data and metadata of model calls in an account and Region. It delivers to CloudWatch Logs or Amazon S3, and it is off until someone turns it on. The records can include the prompts and answers themselves. ## An example, sourced AWS documents that model invocation logging collects invocation logs, model input data and model output data for all invocations in your account used in Amazon Bedrock in a Region. AWS says logging is disabled by default, supports destinations in Amazon CloudWatch Logs and Amazon S3 from the same account and Region, and covers operations such as Converse and InvokeModel made through the bedrock-runtime endpoint. ## Why it matters here Prompts and responses are the evidence for what an agent was asked and told. A default of off means an incident team finds no record unless someone enabled the setting before the event. Turning it on also creates stored copies of prompts, which need their own access control and retention decision. ## Related pages - [CloudTrail data events: meaning and an example](/glossary/cloudtrail-data-events/) - [Cloud Audit Logs: meaning and an example](/glossary/cloud-audit-logs/) - [Data access audit logs: meaning and an example](/glossary/data-access-audit-logs/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/model-invocation-logging/ --- # CloudTrail data events Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. AWS CloudTrail records of operations on or inside a resource, such as reading an S3 object or invoking a function, as opposed to management events that record changes to configuration. They are high volume, so trails do not record them unless you ask. ## An example, sourced AWS documents that Amazon Bedrock logs all Agents for Amazon Bedrock Runtime API operations, such as InvokeAgent and InvokeInlineAgent, to CloudTrail as data events. To log InvokeAgent calls, you configure advanced event selectors to record data events for the AWS::Bedrock::AgentAlias resource type. AWS also notes that CloudTrail does not log data events by default because they are often high-volume. The Bedrock page gives the event selector for the agent alias type. ## Why it matters here Data events are where an agent's reads and invocations show up. A trail with only management events cannot answer which objects the agent touched. Cost is the usual reason they are off, which should be an explicit decision in the review. ## Related pages - [Cloud Audit Logs: meaning and an example](/glossary/cloud-audit-logs/) - [Data access audit logs: meaning and an example](/glossary/data-access-audit-logs/) - [VPC endpoint: meaning and an example](/glossary/vpc-endpoint/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/bedrock/latest/userguide/logging-using-cloudtrail.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/cloudtrail-data-events/ --- # Cloud Audit Logs Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Google Cloud's audit trail of who did what, where and when. It has several log types, including Admin Activity logs for configuration changes and Data Access logs for reads and writes of data, and its entries are written to Cloud Logging. ## An example, sourced Google documents Admin Activity audit logs as log entries written by user-driven API calls or other actions that modify the configuration or metadata of resources. The Cloud Audit Logs overview names several log types, including Admin Activity and Policy Denied logs, and tells you which roles are needed to read them. Google states that log entries written by Cloud Audit Logs are immutable, which matters when the log has to serve as evidence of what an agent or its administrators did. ## Why it matters here Audit logs are the record of what happened to an agent's role and settings. Because entries are immutable, they hold up better as evidence than logs the agent's own identity could edit. A reviewer should check which log types are enabled for the project the agent runs in. ## Related pages - [Data access audit logs: meaning and an example](/glossary/data-access-audit-logs/) - [VPC endpoint: meaning and an example](/glossary/vpc-endpoint/) - [AWS PrivateLink: meaning and an example](/glossary/aws-privatelink/) - [All of them](/glossary/) ## Sources - https://cloud.google.com/logging/docs/audit (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/cloud-audit-logs/ --- # Data access audit logs Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. The Google Cloud audit logs that record reads and writes of user-provided data. They can be large, so they are disabled by default for most services and must be enabled per project, folder or organization before they hold anything. They are separate from the Admin Activity logs. ## An example, sourced Google documents that Data Access audit logs record calls that create, modify or read user-provided resource data. Because these logs are disabled by default for most services, Google tells you to explicitly enable, configure or disable them for your projects, billing accounts, folders and organization. Without that step, an agent's reads of a Cloud Storage bucket would leave no entry. ## Why it matters here This is the setting that turns "the agent has read access" into "this is what the agent read". A review should ask for the audit configuration of the services the agent touches. It is also the log type to enable before an agent is given read access to sensitive data. ## Related pages - [VPC endpoint: meaning and an example](/glossary/vpc-endpoint/) - [AWS PrivateLink: meaning and an example](/glossary/aws-privatelink/) - [Private Service Connect: meaning and an example](/glossary/private-service-connect/) - [All of them](/glossary/) ## Sources - https://cloud.google.com/logging/docs/audit/configure-data-access (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/data-access-audit-logs/ --- # VPC endpoint Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. A private connection from an AWS VPC to an AWS service or another service, without a route through an internet gateway. An endpoint can carry its own policy that limits which principals and actions may use it. Traffic through it stays on the provider's network. ## An example, sourced AWS documents that an endpoint policy is a resource-based policy that you attach to a VPC endpoint to control which AWS principals can use the endpoint to access an AWS service. AWS states that an endpoint policy does not override or replace identity-based policies or resource-based policies. A call through the endpoint must be allowed by the endpoint policy and by the other policies that apply. ## Why it matters here Endpoint policies can limit what a stolen credential does from inside the network. For an agent they act as an extra layer, set by the network team, on top of the agent's own role. A reviewer should ask who owns the policy on each endpoint the agent uses. ## Related pages - [AWS PrivateLink: meaning and an example](/glossary/aws-privatelink/) - [Private Service Connect: meaning and an example](/glossary/private-service-connect/) - [Private endpoint: meaning and an example](/glossary/private-endpoint/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/vpc/latest/privatelink/vpc-endpoints-access.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/vpc-endpoint/ --- # AWS PrivateLink Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. The AWS technology behind interface VPC endpoints. It gives a private connection between a VPC and a service, so traffic reaches the service as if it were inside the VPC, without an internet gateway or NAT device on the path. ## An example, sourced AWS documents that you can use AWS PrivateLink to create a private connection between your VPC and Amazon Bedrock, and that you can access Amazon Bedrock as if it were in your VPC. The page covers considerations, creating an interface endpoint, creating an endpoint policy and connecting to Amazon Bedrock through the endpoint. It also notes that Amazon Bedrock and Amazon Bedrock Agents endpoints are available in multiple Availability Zones, which affects how an agent\'s network is laid out. ## Why it matters here PrivateLink keeps API calls from the agent's network on the AWS network. It does not set what the agent may do, so it complements the role policy rather than replacing it. Check the endpoint policy and the security groups on the endpoint interface too. ## Related pages - [Private Service Connect: meaning and an example](/glossary/private-service-connect/) - [Private endpoint: meaning and an example](/glossary/private-endpoint/) - [Egress control: meaning and an example](/glossary/egress-control/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/bedrock/latest/userguide/vpc-interface-endpoints.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/aws-privatelink/ --- # Private Service Connect Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. A Google Cloud networking capability that lets consumers reach managed services privately from their own VPC networks using internal IP addresses. Producers can publish services through it, and a Private Service Connect interface can give a managed service private egress into your network. ## An example, sourced Google documents Private Service Connect as a capability of Google Cloud networking that allows consumers to access managed services privately from inside their VPC network. With it, consumers use their own internal IP addresses to access services without leaving their VPC networks, and Google states that traffic remains entirely within Google Cloud. Service producers publish their applications to consumers, and Private Service Connect interfaces let managed services initiate connections to consumers, which is the form the Agent Runtime feature takes. ## Why it matters here Agent Runtime agents sit in a Google-managed network by default. The interface is how a reviewer gets the agent onto a network the organization controls and can filter. The grants the interface needs are a separate item for the review to read. ## Related pages - [Private endpoint: meaning and an example](/glossary/private-endpoint/) - [Egress control: meaning and an example](/glossary/egress-control/) - [Knowledge base: meaning and an example](/glossary/knowledge-base/) - [All of them](/glossary/) ## Sources - https://cloud.google.com/vpc/docs/private-service-connect (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/private-service-connect/ --- # Private endpoint Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. An Azure network interface with a private IP address from your virtual network that connects privately to a service on Azure Private Link. It brings the service into your network, so traffic to it need not cross the public internet. ## An example, sourced Microsoft documents that a private endpoint is a network interface that uses a private IP address from your virtual network. The interface connects you privately and securely to a service powered by Azure Private Link. Microsoft says that by enabling a private endpoint, you bring the service into your virtual network. A storage account or a key vault behind a private endpoint can then be reached from the network the organization controls, while public access can be restricted on the service itself. ## Why it matters here Private endpoints let the network team close public access to a data store or a Foundry project. The agent then reaches it only from the network the team controls. Public network access on the service must be turned off for the endpoint to close the public path. ## Related pages - [Egress control: meaning and an example](/glossary/egress-control/) - [Knowledge base: meaning and an example](/glossary/knowledge-base/) - [Retrieval augmented generation: meaning and an example](/glossary/retrieval-augmented-generation/) - [All of them](/glossary/) ## Sources - https://learn.microsoft.com/en-us/azure/private-link/private-endpoint-overview (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/private-endpoint/ --- # Egress control Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Limiting where outbound traffic from a workload may go, by destination, protocol or domain. For an agent, egress rules decide whether a hijacked prompt can send data to an attacker's server or fetch tools from one. Firewalls, security groups and network rules are the usual tools. ## An example, sourced AWS documents that AWS Network Firewall is a stateful, managed network firewall and intrusion detection and prevention service for your VPC. With it, AWS says you can filter traffic at the perimeter of your VPC, and the filtering includes traffic going to and coming from an internet gateway, a NAT gateway, a transit gateway, or over VPN. For an agent, rules at that perimeter can limit outbound traffic to a short list of destinations that the review has approved. ## Why it matters here Identity limits what a call may do, and egress limits where it can go. Open egress means a failed permission check is not the only thing between an agent and an outside server. Ask for the rules in force, not for the intention behind them. ## Related pages - [Knowledge base: meaning and an example](/glossary/knowledge-base/) - [Retrieval augmented generation: meaning and an example](/glossary/retrieval-augmented-generation/) - [Confused deputy: meaning and an example](/glossary/confused-deputy/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/network-firewall/latest/developerguide/what-is-aws-network-firewall.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/egress-control/ --- # Knowledge base Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. In Amazon Bedrock, a store of your documents that an agent can query for extra context when it builds an answer. Access to it is a permission on the agent's role, and the data in it can reach every user of the agent. ## An example, sourced AWS documents that you can associate knowledge bases with an agent, and that the agent queries the knowledge base for extra context to augment response generation and input into steps of the orchestration process. AWS also documents that the base prompt templates the agent builds contain instructions, action descriptions, knowledge base descriptions and conversation history. Orchestration includes predicting whether the agent should invoke an action or query a knowledge base, so retrieval is a standing part of the loop. ## Why it matters here Everything in a knowledge base is readable by whoever can steer the agent. Reviewers should ask what documents it holds and whether the agent's users are all allowed to see them. The same question applies to the store behind it, which may have its own access list. ## Related pages - [Retrieval augmented generation: meaning and an example](/glossary/retrieval-augmented-generation/) - [Confused deputy: meaning and an example](/glossary/confused-deputy/) - [Resource-based policy: meaning and an example](/glossary/resource-based-policy/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/bedrock/latest/userguide/agents-how.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/knowledge-base/ --- # Retrieval augmented generation Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. A pattern where a system looks up relevant documents first and gives them to a language model along with the question, so the answer draws on those documents. Often shortened to RAG. The lookup step is where an agent touches your data. ## An example, sourced AWS documents the components of Amazon Q Business as working together to create your Retrieval Augmented Generation (RAG) solution. The what-is page describes the service as generating responses by analyzing information across all enterprise content that it has access to, confining responses to existing enterprise data, and citing sources. It integrates with data sources such as Amazon S3, Microsoft SharePoint and Salesforce. The page also states that Amazon Q Business is no longer open to new customers. ## Why it matters here Retrieval has its own permissions, separate from the model call. A reviewer should check what the retrieval identity can read, and whether document-level permissions of the source system carry through to the answer a given user receives. The same applies to any index built from them, and to the connector's own credentials. ## Related pages - [Confused deputy: meaning and an example](/glossary/confused-deputy/) - [Resource-based policy: meaning and an example](/glossary/resource-based-policy/) - [AI agent: meaning and an example](/glossary/ai-agent/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/what-is.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/retrieval-augmented-generation/ --- # Confused deputy Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. A security problem where something without permission to do an action tricks a more-privileged party into doing it. A cloud service acting on your behalf is a typical deputy, so the trust on its role has to say which of its customers or resources may use it. ## An example, sourced AWS documents the confused deputy problem as a security issue where an entity that lacks permission to perform an action can coerce a more-privileged entity to perform the action. AWS says it provides tools to protect your account when you give third parties (cross-account) or other AWS services (cross-service) access to resources in your account, and recommends keys such as aws:SourceArn and aws:SourceAccount in resource policies. ## Why it matters here Agent platforms are deputies by design: a service assumes your role. Conditions on the trust policy tie the role to one agent so another customer's agent, or another of yours, cannot use it. It is a cheap condition to add and an easy one to omit. ## Related pages - [Resource-based policy: meaning and an example](/glossary/resource-based-policy/) - [AI agent: meaning and an example](/glossary/ai-agent/) - [AI workload: meaning and an example](/glossary/ai-workload/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/IAM/latest/UserGuide/confused-deputy.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/confused-deputy/ --- # Resource-based policy Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. A permission policy attached to the resource being accessed, such as a Lambda function, an S3 bucket or a KMS key, naming who may use it. It works alongside the caller's own identity-based policy and always names a principal. It is the counterpart of the policy attached to the caller. ## An example, sourced AWS documents that for an Amazon Bedrock agent to use a Lambda function, you must attach a resource-based policy to the function to provide permissions for the agent. The step is separate from the agent\'s own role: the policy lives on the function and says who may call it. The page points to the AWS Lambda Developer Guide for more information about resource-based policies in Lambda, which is where the policy syntax and the principals it can name are documented. ## Why it matters here The same role can reach a resource through its own policy or through the resource's policy. A review that reads only the role misses grants written on the targets. It also shows which other principals, including other accounts, were granted the same resource. ## Related pages - [AI agent: meaning and an example](/glossary/ai-agent/) - [AI workload: meaning and an example](/glossary/ai-workload/) - [Execution role: meaning and an example](/glossary/execution-role/) - [All of them](/glossary/) ## Sources - https://docs.aws.amazon.com/bedrock/latest/userguide/agents-lambda.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/glossary/resource-based-policy/ --- # What changed in the last 90 days Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Dated news that changes the map or the ranking: renames, retirements, preview launches and acquisitions, such as Bedrock Agents becoming Agents Classic and Apono joining 1Password in June 2026. Each item names its source and the date the page was read. - [AWS Nitro Sandboxes give pods, tasks and agents their own hardware-isolated VM](/news/nitro-sandboxes-eks-ecs-ec2-agents/): 9 Nov 2026 - [Amazon Quick adds smart approvals for actions that agents take](/news/quick-smart-approvals-agentic-actions/): 28 Oct 2026 - [WorkSpaces for AI agents adds OAuth sessions inside existing network boundaries](/news/workspaces-for-ai-agents-oauth-sessions/): 20 Oct 2026 - [Act Security and HUB International build a no-credentials coding agent environment](/news/act-hub-international-agentcore-coding-agents/): 19 Oct 2026 - [Act Security argues AI governance should limit what agents can reach](/news/act-real-ai-risk-what-agents-can-reach/): 18 Oct 2026 - [Amazon Bedrock Managed Agents powered by OpenAI enters preview](/news/bedrock-managed-agents-openai-preview/): 29 Sep 2026 - [Security Command Center flags reasoning engines that can modify IAM policies](/news/scc-risk-engine-reasoning-engines-modify-iam/): 28 Sep 2026 - [IAM outbound identity federation supports VPC endpoints for OIDC discovery](/news/iam-outbound-federation-vpc-endpoints-oidc/): 25 Sep 2026 - [New AgentCore Runtime keeps hardware-enforced session isolation](/news/agentcore-runtime-v2-session-isolation/): 18 Sep 2026 - [Policy Troubleshooter can now explain access denials for agents](/news/policy-troubleshooter-agent-access-denials/): 15 Sep 2026 - [Google agent platform shell and computer use sandboxes reach GA with VPC-SC](/news/agent-platform-sandboxes-vpc-sc-psc-ga/): 9 Sep 2026 - [Google IAM unified access policies for agents reach general availability](/news/agent-platform-iam-access-policies-ga/): 31 Aug 2026 - [AgentCore Memory adds fine-grained access control through Cedar policies](/news/agentcore-memory-fine-grained-access-control/): 28 Aug 2026 - [Amazon Quick custom permissions can deny new AI capabilities by default](/news/quick-custom-permissions-deny-by-default/): 19 Aug 2026 ## Related pages - [Agent platforms in AWS, Google Cloud and Azure](/agent-platforms/) - [Tools scored on AI access coverage](/tools/) - [Picks by situation](/situations/) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/news/ --- # AWS Nitro Sandboxes give pods, tasks and agents their own hardware-isolated VM Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. AWS announced AWS Nitro Sandboxes on 9 November 2026: lightweight VMs managed by the Nitro Hypervisor for an EKS pod, ECS task or VM, including AI agents running inside them. Each workload gets its own isolation boundary, and AWS says secrets are released only after attestation. AWS announced on 9 November 2026 that AWS Nitro Sandboxes are available as lightweight VMs managed by the Nitro Hypervisor. They can be used for an Amazon EKS pod, an Amazon ECS task or a VM, including AI agents running inside them, each in its own hardware-isolated virtual machine. AWS says this allows untrusted, multi-tenant or regulated workloads to share the resources of a single instance. AWS also says workloads can be validated through cryptographic attestation so that secrets are released only to Nitro Sandboxes that have proven their integrity, and that LLM-generated code can run in an isolated environment of its own. Enabling it in EKS or ECS is described as a one-line change, with Pod Identity and VPC controls staying native. It is available on most Nitro-based EC2 instances at no extra charge. ## Related pages - [Amazon Quick adds smart approvals for actions that agents take](/news/quick-smart-approvals-agentic-actions/) - [WorkSpaces for AI agents adds OAuth sessions inside existing network boundaries](/news/workspaces-for-ai-agents-oauth-sessions/) - [Act Security and HUB International build a no-credentials coding agent environment](/news/act-hub-international-agentcore-coding-agents/) - [All of them](/news/) ## Sources - https://aws.amazon.com/about-aws/whats-new/2026/11/nitro-sandboxes-eks-ecs-ec2/ (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/news/nitro-sandboxes-eks-ecs-ec2-agents/ --- # Amazon Quick adds smart approvals for actions that agents take Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. AWS announced on 28 October 2026 that Amazon Quick supports smart approvals. A separate classifier model reviews each eligible agent action, approves routine ones and sends consequential ones, such as changing permissions or deleting content, to the user. AWS states that smart approvals never grants new access. Amazon Quick now supports smart approvals, a review layer for actions agents take for a user, such as sending email, updating records or creating tickets, per an AWS announcement dated 28 October 2026. A separate classifier model evaluates each eligible action against the request, approves routine actions and refers consequential ones to the user. AWS names sending content outside the organization, changing permissions, deleting content and making payments as higher-impact examples. AWS states that existing policies come first: it only reviews actions where the user setting is ask each time, denied actions stay denied, and smart approvals never grants new access. It is on by default, administrators control who can use it through custom permissions, and CloudTrail records each review. The limit on what an agent can touch remains the permission policy underneath. ## Related pages - [WorkSpaces for AI agents adds OAuth sessions inside existing network boundaries](/news/workspaces-for-ai-agents-oauth-sessions/) - [Act Security and HUB International build a no-credentials coding agent environment](/news/act-hub-international-agentcore-coding-agents/) - [Act Security argues AI governance should limit what agents can reach](/news/act-real-ai-risk-what-agents-can-reach/) - [All of them](/news/) ## Sources - https://aws.amazon.com/about-aws/whats-new/2026/10/quick-smart-approvals-agentic-actions/ (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/news/quick-smart-approvals-agentic-actions/ --- # WorkSpaces for AI agents adds OAuth sessions inside existing network boundaries Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. AWS said on 20 October 2026 that WorkSpaces for AI agents now handles authentication and desktop connections in its managed MCP service. Sessions use existing AWS credentials through OAuth 2.0, agents run inside the VPC controls and data-loss policies that govern staff, and the service adds four Regions. AWS announced on 20 October 2026 that WorkSpaces for AI agents now handles authentication and desktop connections as part of its managed Model Context Protocol service. You can create sessions with your existing AWS credentials through OAuth 2.0, and the service manages the connection lifecycle. Agents reach the right desktop without seeing the underlying streaming connection, and you can query which stack, fleet and session an agent is using. AWS also says agent sessions run inside the same network isolation and compliance boundaries as the human workforce, inheriting VPC controls, data-loss prevention policies and session-level security posture. The service is now in four more Regions: Sao Paulo, Milan, Spain and Malaysia. For reach, an agent on a cloud desktop can only get as far as that desktop can. ## Related pages - [Act Security and HUB International build a no-credentials coding agent environment](/news/act-hub-international-agentcore-coding-agents/) - [Act Security argues AI governance should limit what agents can reach](/news/act-real-ai-risk-what-agents-can-reach/) - [Amazon Bedrock Managed Agents powered by OpenAI enters preview](/news/bedrock-managed-agents-openai-preview/) - [All of them](/news/) ## Sources - https://aws.amazon.com/about-aws/whats-new/2026/10/workspaces-ai-agents-oauth-connection-management (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/news/workspaces-for-ai-agents-oauth-sessions/ --- # Act Security and HUB International build a no-credentials coding agent environment Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Act Security published a blog post on 19 October 2026 about HUB International moving Claude Code into cloud-hosted sandboxes on Amazon Bedrock AgentCore in two days. Act says the agent holds no credentials and every action is attributed to a named developer under central control. Act Security published a blog post dated 19 October 2026 about HUB International. Act says that working with AWS, HUB moved the local Claude Code experience into cloud-hosted sandboxes in two days, with sessions that survive a closed laptop, no credentials in the agent, and every action attributed to a named developer. Act says the work took place in September 2026 inside HUB's own AWS account. The post describes native AWS pieces: AgentCore Runtime for per-session sandboxes, AgentCore Gateway as the registered path to tools, AgentCore Identity for per-user credentials, a Cedar policy engine at the gateway, and service control and resource control policies as account guardrails. This is Act's own account of a customer project, labelled as such, and it shows one way to limit what a coding agent can reach. ## Related pages - [Act Security argues AI governance should limit what agents can reach](/news/act-real-ai-risk-what-agents-can-reach/) - [Amazon Bedrock Managed Agents powered by OpenAI enters preview](/news/bedrock-managed-agents-openai-preview/) - [Security Command Center flags reasoning engines that can modify IAM policies](/news/scc-risk-engine-reasoning-engines-modify-iam/) - [All of them](/news/) ## Sources - https://act.security/resources/how-hub-international-built-a-secure-home-for-ai-coding-agents-in-two-days-using-amazon-bedrock-agentcore (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/news/act-hub-international-agentcore-coding-agents/ --- # Act Security argues AI governance should limit what agents can reach Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Act Security published a news piece on 18 October 2026 titled The Real AI Risk Isn't How Agents Think. It's What They Can Reach. Act argues that agent reasoning cannot be made reliable by prompts, so leadership should set strict access boundaries from the first day. Act Security posted a news item dated 18 October 2026 titled The Real AI Risk Isn't How Agents Think. It's What They Can Reach. In its summary, Act says agent reasoning is inherently unpredictable and vulnerable to manipulation, and that no amount of prompt engineering can prevent it from failing. It argues that leadership should set strict boundaries from day one instead of reacting after a crisis. Act frames AI governance as controlling how much damage an agent can do the moment its reasoning goes wrong, and not as ensuring an agent always reasons correctly. The saved page shows only this summary and the date, so the full argument is not available here. It is a vendor opinion piece that matches the focus of this site on agent reach in a cloud account. ## Related pages - [Amazon Bedrock Managed Agents powered by OpenAI enters preview](/news/bedrock-managed-agents-openai-preview/) - [Security Command Center flags reasoning engines that can modify IAM policies](/news/scc-risk-engine-reasoning-engines-modify-iam/) - [IAM outbound identity federation supports VPC endpoints for OIDC discovery](/news/iam-outbound-federation-vpc-endpoints-oidc/) - [All of them](/news/) ## Sources - https://act.security/resources/the-real-ai-risk-isnt-how-agents-think-its-what-they-can-reach (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/news/act-real-ai-risk-what-agents-can-reach/ --- # Amazon Bedrock Managed Agents powered by OpenAI enters preview Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. AWS announced on 29 September 2026 a preview of Amazon Bedrock Managed Agents, built jointly with OpenAI. AWS says each agent runs with its own IAM role, supports human approval before consequential actions, and records supported API activity with AWS CloudTrail, in three US Regions. AWS announced the preview of Bedrock Managed Agents, developed jointly by AWS and OpenAI, on 29 September 2026. It is built on a customized version of OpenAI's Agents API and is described as AWS-native. AWS says you can build agents optimized for OpenAI models that run inside AWS with the identities, permissions and governance controls you already use, with durable sessions, reusable skills and tool connections, including MCP servers. On access, AWS states that each agent operates with its own IAM role, supports human approval before consequential actions, and records supported API activity with CloudTrail. The preview runs in US East (N. Virginia), US West (Oregon) and US East (Ohio), with no extra charge beyond underlying resources. Reach is set by that IAM role, so its permissions policy is the main control to review. ## Related pages - [Security Command Center flags reasoning engines that can modify IAM policies](/news/scc-risk-engine-reasoning-engines-modify-iam/) - [IAM outbound identity federation supports VPC endpoints for OIDC discovery](/news/iam-outbound-federation-vpc-endpoints-oidc/) - [New AgentCore Runtime keeps hardware-enforced session isolation](/news/agentcore-runtime-v2-session-isolation/) - [All of them](/news/) ## Sources - https://aws.amazon.com/about-aws/whats-new/2026/09/bedrock-managed-agents-preview/ (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/news/bedrock-managed-agents-openai-preview/ --- # Security Command Center flags reasoning engines that can modify IAM policies Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Google Cloud said on 28 September 2026 that the Security Command Center Risk Engine now detects and reports reasoning engines that can modify IAM policies and perform lateral movement. The findings are generated as a toxic combination, so a risky agent permission pairing shows up as one issue. Security Command Center release notes dated 28 September 2026 say that Risk Engine detects and reports reasoning engines that can modify IAM policies and perform lateral movement. These findings are generated as a toxic combination, which Google uses for risk that comes from several conditions together instead of a single misconfiguration, so the report covers the pairing and not each permission alone. The note does not say how the engines are identified, how many are affected, or what remediation Google advises. For teams that deploy agents on Google Cloud, it points at one specific question: whether an agent runtime can change IAM policy, and whether it could use that to move elsewhere. The finding is a prompt to review the permissions attached to the engine. ## Related pages - [IAM outbound identity federation supports VPC endpoints for OIDC discovery](/news/iam-outbound-federation-vpc-endpoints-oidc/) - [New AgentCore Runtime keeps hardware-enforced session isolation](/news/agentcore-runtime-v2-session-isolation/) - [Policy Troubleshooter can now explain access denials for agents](/news/policy-troubleshooter-agent-access-denials/) - [All of them](/news/) ## Sources - https://cloud.google.com/security-command-center/docs/release-notes (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/news/scc-risk-engine-reasoning-engines-modify-iam/ --- # IAM outbound identity federation supports VPC endpoints for OIDC discovery Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. AWS announced on 25 September 2026 that IAM outbound identity federation supports interface VPC endpoints for OIDC discovery and JWKS key endpoints through AWS PrivateLink. Workloads that get short-lived tokens from AWS STS can have them verified without internet access, instead of using long-lived credentials. AWS Identity and Access Management outbound identity federation now supports Amazon VPC endpoints for the OpenID Connect discovery APIs, per an AWS announcement dated 25 September 2026. You can reach the OIDC discovery metadata and the JSON Web Key Set verification key endpoints from inside your VPC with AWS PrivateLink, so the traffic that retrieves verification keys stays inside the AWS network and does not cross the public internet. AWS explains that outbound federation removes the need for long-lived credentials when AWS workloads access external services: workloads request short-lived JSON Web Tokens from AWS STS and external services verify them with public keys. Until now the discovery endpoints were reachable only over the public internet, so a verifier in a VPC without internet access could not fetch them. The feature costs only standard PrivateLink pricing. ## Related pages - [New AgentCore Runtime keeps hardware-enforced session isolation](/news/agentcore-runtime-v2-session-isolation/) - [Policy Troubleshooter can now explain access denials for agents](/news/policy-troubleshooter-agent-access-denials/) - [Google agent platform shell and computer use sandboxes reach GA with VPC-SC](/news/agent-platform-sandboxes-vpc-sc-psc-ga/) - [All of them](/news/) ## Sources - https://aws.amazon.com/about-aws/whats-new/2026/09/aws-sts-vpc-oidc/ (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/news/iam-outbound-federation-vpc-endpoints-oidc/ --- # New AgentCore Runtime keeps hardware-enforced session isolation Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. AWS announced on 18 September 2026 the next generation of AgentCore Runtime, the serverless microVM compute in Amazon Bedrock AgentCore. It keeps hardware-enforced session isolation and scale to zero, reclaims unused memory, and AWS reports a P75 cold start of 1.9 to 2.0 seconds in testing. AWS announced the next generation of AgentCore Runtime, the serverless microVM compute within Amazon Bedrock AgentCore, on 18 September 2026. The new Runtime reclaims unused memory during a session so you pay for actual usage, and prepares the agent environment once and snapshots it so cold starts stay consistent. In AWS testing, P75 cold start was 1.9 to 2.0 seconds for images from 200 MB to 2 GB, against 5.4 to 30 seconds with V1. AWS states the serverless model is unchanged, with no pre-provisioning, scale to zero and hardware-enforced session isolation. It is available in five Regions: us-east-1, us-east-2, us-west-2, eu-west-1 and ap-northeast-1, and you opt in by setting platformVersion to V2. For access, the isolation limits what one agent session can see of another, while the runtime role still sets what an agent reaches in the account. ## Related pages - [Policy Troubleshooter can now explain access denials for agents](/news/policy-troubleshooter-agent-access-denials/) - [Google agent platform shell and computer use sandboxes reach GA with VPC-SC](/news/agent-platform-sandboxes-vpc-sc-psc-ga/) - [Google IAM unified access policies for agents reach general availability](/news/agent-platform-iam-access-policies-ga/) - [All of them](/news/) ## Sources - https://aws.amazon.com/about-aws/whats-new/2026/09/new-agentcore-runtime-generally-available (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/news/agentcore-runtime-v2-session-isolation/ --- # Policy Troubleshooter can now explain access denials for agents Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Google Cloud announced on 15 September 2026 that Policy Troubleshooter supports troubleshooting access for agent identities. You can check IAM allow policies, deny policies and principal access boundary policies for agents acting under their own authority by entering the principal identifier or an error ID from a denial. Google Cloud added a Policy Intelligence feature on 15 September 2026: Policy Troubleshooter now supports troubleshooting access for agent identities. The release note says you can troubleshoot IAM allow policies, deny policies and principal access boundary policies for agents acting under their own authority. You do this by entering the agent principal identifier or by using an error ID from an access denial event. For anyone asking why an agent could not reach a resource, or why it could, this puts the three policy layers in one check. Google lists it as a feature without a preview label in that entry, and links to its Troubleshooting access documentation. It explains existing policy decisions and does not change what an agent is allowed to do. ## Related pages - [Google agent platform shell and computer use sandboxes reach GA with VPC-SC](/news/agent-platform-sandboxes-vpc-sc-psc-ga/) - [Google IAM unified access policies for agents reach general availability](/news/agent-platform-iam-access-policies-ga/) - [AgentCore Memory adds fine-grained access control through Cedar policies](/news/agentcore-memory-fine-grained-access-control/) - [All of them](/news/) ## Sources - https://cloud.google.com/policy-intelligence/docs/release-notes (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/news/policy-troubleshooter-agent-access-denials/ --- # Google agent platform shell and computer use sandboxes reach GA with VPC-SC Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Google Cloud made Computer Use and Shell sandboxes in Gemini Enterprise Agent Platform generally available on 9 September 2026. The release adds VPC Service Controls, private ingress and egress through Private Service Connect, customer-managed encryption keys, and pausing and resuming of idle sandboxes. Google Cloud recorded on 9 September 2026 that Computer Use and Shell sandboxes in Gemini Enterprise Agent Platform are generally available. Shell sandboxes run untrusted shell commands, install packages and manipulate files in an isolated Linux container through direct API calls. The same entry lists new sandbox features, including support for VPC Service Controls, private ingress endpoints and private egress routing with Private Service Connect. The note also lists customer-managed encryption keys for sandbox data at rest, and pausing and resuming idle sandboxes while keeping file system state and connection identity. For agent reach, the VPC Service Controls support means a sandbox can be placed inside the same perimeter as the data it should be able to touch, with private routing as the network path. Teams still need to configure those controls themselves. ## Related pages - [Google IAM unified access policies for agents reach general availability](/news/agent-platform-iam-access-policies-ga/) - [AgentCore Memory adds fine-grained access control through Cedar policies](/news/agentcore-memory-fine-grained-access-control/) - [Amazon Quick custom permissions can deny new AI capabilities by default](/news/quick-custom-permissions-deny-by-default/) - [All of them](/news/) ## Sources - https://cloud.google.com/gemini-enterprise-agent-platform/release-notes (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/news/agent-platform-sandboxes-vpc-sc-psc-ga/ --- # Google IAM unified access policies for agents reach general availability Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Google Cloud made IAM Unified Access Policies for Gemini Enterprise Agent Platform generally available on 31 August 2026. Agent Gateway uses Identity-Aware Proxy and these policies to govern communication between agent principals and tools, with allow and deny rules, CEL conditions and a dry-run mode. Google Cloud logged on 31 August 2026 that IAM Unified Access Policies for Gemini Enterprise Agent Platform are generally available. The release note says Agent Gateway uses Identity-Aware Proxy and access policies to govern agentic communication between agent principals and destination resources, such as MCP servers, other agents, and registered or unregistered endpoints. Google lists key capabilities for the release. Listed capabilities include multiple allow and deny rules in one policy, Common Expression Language conditions on tool names, read-only constraints, HTTP methods and URL paths, and dry-run and enforcement modes to test evaluation before blocking agent traffic. The note also mentions end-to-end agent authentication with mutual TLS. For readers, the policy is where the question of which tools an agent may call gets written down and tested. ## Related pages - [AgentCore Memory adds fine-grained access control through Cedar policies](/news/agentcore-memory-fine-grained-access-control/) - [Amazon Quick custom permissions can deny new AI capabilities by default](/news/quick-custom-permissions-deny-by-default/) - [AWS Nitro Sandboxes give pods, tasks and agents their own hardware-isolated VM](/news/nitro-sandboxes-eks-ecs-ec2-agents/) - [All of them](/news/) ## Sources - https://cloud.google.com/gemini-enterprise-agent-platform/release-notes (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/news/agent-platform-iam-access-policies-ga/ --- # AgentCore Memory adds fine-grained access control through Cedar policies Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. AWS announced on 28 August 2026 that Amazon Bedrock AgentCore Memory supports fine-grained access control. Placing a Memory resource behind an AgentCore Gateway with OAuth authentication lets Cedar policies limit each caller to their own actor data, namespaces derived from token claims, and specific Memory operations. Amazon Bedrock AgentCore Memory now supports fine-grained access control, according to an AWS announcement dated 28 August 2026. You front the Memory resource with an AgentCore Gateway configured for OAuth (JWT) authentication and attach Cedar policies that restrict access based on the authenticated caller. AWS says you can make each user reach only their own actor data and limit memory records to namespaces taken from token claims. The announcement adds that the feature rests on an AgentCore Memory connector, which exposes 12 Memory operations as Cedar actions, so policies can allow or deny specific operations per caller. For anyone asking what an agent can read from stored memory, this moves the check from application code to a policy layer. It depends on routing memory access through the gateway. ## Related pages - [Amazon Quick custom permissions can deny new AI capabilities by default](/news/quick-custom-permissions-deny-by-default/) - [AWS Nitro Sandboxes give pods, tasks and agents their own hardware-isolated VM](/news/nitro-sandboxes-eks-ecs-ec2-agents/) - [Amazon Quick adds smart approvals for actions that agents take](/news/quick-smart-approvals-agentic-actions/) - [All of them](/news/) ## Sources - https://aws.amazon.com/about-aws/whats-new/2026/08/agentcorememory-fine-grained-access-control (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/news/agentcore-memory-fine-grained-access-control/ --- # Amazon Quick custom permissions can deny new AI capabilities by default Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. AWS announced on 19 August 2026 that Amazon Quick custom permissions now include deny by default. Administrators can restrict an AI capability category in a permissions profile so new capabilities are denied at launch for chosen users, roles or the whole account until each one is explicitly allowed. Amazon Quick custom permissions now include deny by default, a governance setting that restricts new AI capabilities before they reach users, according to an AWS announcement posted 19 August 2026. Previously, new AI capabilities were available to all users on release, so administrators had to react after the fact. The setting is configured in Manage account in Amazon Quick or through the AWS CLI. An administrator restricts an AI capability category in a custom permissions profile and assigns it to users, roles or the entire account. Quick then denies any new capability in that category at launch, and restricting a category also restricts the capabilities already in it. Administrators explicitly allow each capability when ready. The restriction applies only to the profile you configure, in Regions where Quick is available. ## Related pages - [AWS Nitro Sandboxes give pods, tasks and agents their own hardware-isolated VM](/news/nitro-sandboxes-eks-ecs-ec2-agents/) - [Amazon Quick adds smart approvals for actions that agents take](/news/quick-smart-approvals-agentic-actions/) - [WorkSpaces for AI agents adds OAuth sessions inside existing network boundaries](/news/workspaces-for-ai-agents-oauth-sessions/) - [All of them](/news/) ## Sources - https://aws.amazon.com/about-aws/whats-new/2026/08/amazon-quick-deny-by-default/ (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/news/quick-custom-permissions-deny-by-default/ --- # 30 questions, answered from the data Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Thirty answers for the security lead and the engineer: what identity a Bedrock, Agent Runtime or Foundry agent runs as, what it gets by default, how to log what it did, how tools are scored here, and why Act ranks first at 5.3 and where it is behind. **What identity does an Amazon Bedrock agent run as?** In Amazon Bedrock Agents Classic, AWS documents that the agent runs as an IAM service role that Bedrock assumes. You can let Bedrock create the role or supply a custom one. In Amazon Bedrock AgentCore, Runtime runs the agent under an IAM execution role that AgentCore assumes. In both cases the role's policy is the main limit on what the agent can call. **What permissions does a Bedrock AgentCore agent start with?** AWS documents that the console can create the execution role for you, and that Gateway can create a service role and attach permissions each time you add a target. The contents of the auto-created execution role: none found on 2026-11-10 in the pages read. AWS also warns that the BedrockAgentCoreFullAccess managed policy grants broad permissions, so write your own role. **Can a Vertex AI agent use my permissions?** Agent Runtime, formerly Vertex AI Agent Engine, separates three identities: you, the agent acting on its own authority, and a delegated identity when the agent acts for you with scoped 3-legged OAuth. By default the agent does not borrow your permissions. Calls to Google Cloud APIs with agent identity are evaluated against the agent identity's roles. **What does a Microsoft Foundry agent get by default?** Microsoft documents that a new Entra agent identity starts without resource permissions and that Foundry does not assign RBAC roles to it automatically. You or an administrator grant access through the target service. Legacy agents can share the project-wide managed identity, which gives no per-agent separation, so check which kind your agent uses. **What does a self-hosted agent on Lambda or EKS run as?** A Lambda function runs as its execution role. On EKS, Pod Identity associates an IAM role with a Kubernetes service account in a namespace, and IRSA is the older OIDC-based option. Every tool the code calls runs as that same role unless the code assumes another role. On EKS, restrict the instance metadata service so pods cannot reach the node role. **What is the default identity for an agent on Cloud Run or GKE?** Cloud Run runs the service as the Compute Engine default service account if you do not choose one. Google documents that, depending on organization policy, that account might automatically be granted the Editor role on the project. Enforce the iam.automaticIamGrantsForDefaultServiceAccounts constraint, or choose a dedicated service account for each agent. On GKE, use Workload Identity Federation. **What does an agent on Azure Container Apps or AKS act as?** A container app uses a system-assigned managed identity tied to the app, or user-assigned identities. On AKS, Microsoft Entra Workload ID gives pods an Entra identity. A managed identity needs role assignments for the services it calls, or calls to Key Vault and SQL Database are rejected. A default role grant: none found on 2026-11-10. Container Apps is public by default. **What is the Agent Access Map?** The Agent Access Map is a table of eight agent setups across AWS, Google Cloud and Azure. Each row gives the identity the agent runs as, its default permissions, the tools it calls, the network path, the logs and the boundary mechanisms that apply. It is built from provider documentation only, and each cell has a quote and a source URL. **How was the map built, and how current is it?** Every cell comes from AWS, Google Cloud or Microsoft documentation read on 10 Nov 2026, with a quote and a URL. Where a page gave no answer, the cell says none found on 2026-11-10. That is a statement about the pages read. Preview features are marked preview. Cloud services change quickly, so check the provider's page before you approve. **Should I start a new project on Amazon Bedrock Agents Classic?** AWS documents that you can no longer sign up for Amazon Bedrock Agents (now Amazon Bedrock Agents Classic), that existing users can continue, and that the service is in maintenance mode. AWS points to Amazon Bedrock AgentCore for similar capabilities. A new agent should be reviewed against the AgentCore row of the map, and an existing Classic agent against its own row. **How do I log what an AWS agent did?** Create an organization trail in CloudTrail and enable data events, which trails do not log by default. For Bedrock Agents Classic, log InvokeAgent calls with data events for the AWS::Bedrock::AgentAlias resource type, and turn on model invocation logging, which is off by default. Then alarm on denied calls and on changes to the agent's role. Checks ab-27 to ab-30 cover this. **How do I log what a Google Cloud agent did?** Admin Activity audit logs cannot be disabled, but Data Access audit logs must be explicitly enabled. On Agent Runtime, stdout and stderr go to Cloud Logging, agent identity appears in logs across Google Cloud services, and user-delegated flows show both the user and the agent. Cloud Logging is not supported for child resources such as Sessions and Memory Bank. **How do I log what an Azure agent did?** For a Foundry agent, Microsoft documents Entra sign-in logs (which agent identity got a token for which resource, and on behalf of which user), Entra audit logs for identity creation and deletion, and Foundry tracing in Application Insights for tool calls. Resource logs are not collected until you create a diagnostic setting. **What is the Agent Boundary Checklist?** The Agent Boundary Checklist has 30 checks in seven sections: identity, permissions, data, network, tools, change control and logging. Each check gives the reason and the AWS, Google Cloud and Azure way to do it, with a documentation quote. It covers the cloud boundary around the agent, so an approval meeting can mark each check done or open. **How do I use the checklist in an approval meeting?** Take the sections in order and ask the engineer to show the evidence for each check in the cloud the agent runs in: the role, the ceiling policy, the endpoint, the log query. A check without evidence is open. Record an owner for each open check, and approve only when the identity, permissions and logging sections are closed. **How do I cap an agent whatever its own policy says?** Set a ceiling above the agent's role. In AWS that is an SCP or a permissions boundary. In Google Cloud it is a Principal Access Boundary policy and an IAM deny policy that targets agent identities. In Azure it is a deny assignment, though Microsoft documents customer-managed deny assignments as private preview. Checks ab-07 and ab-08 cover these. **How do I stop an agent from reaching the internet?** Run the agent on a private network and allow outbound traffic only to the destinations it needs. Foundry's Standard Setup with private networking has no public egress. A Lambda function in a public subnet does not get internet access. On Agent Runtime, a Private Service Connect interface gives private egress, and under VPC Service Controls it is required for internet access. **How do I tell whether a tool call runs as the agent or as the user?** It depends on the platform and the tool. AgentCore Gateway lets each target use the gateway role, the caller's credentials or a token exchange that carries both identities. Google describes separate user, agent and delegated identities. Foundry has application-only and delegated access, the latter in preview. For Bedrock Agents Classic: none found on 2026-11-10. **Does Foundry's agent-scope role limit what the agent can reach?** No. Microsoft documents that role assignments at the agent scope are currently evaluated only for agent endpoint access, so they limit who can call a specific agent. What the agent can reach comes from Azure RBAC role assignments on its agent identity, which start empty. Assign them at resource or resource group scope rather than subscription scope. **How are the tools scored here?** Each of 14 tools gets an integer from 0 to 10 on seven criteria, from facts on its own pages and in named reports. The weights were fixed on 10 Nov 2026 before scoring: identity, network and AI access together 35, removes access 20, safe to enforce 15, class of risk 10, stays fixed 10, native controls 5, audit evidence 5. **Why does identity, network and AI access together weigh 35?** An agent's reach is the sum of identity, network and AI-service access, and a model that misses one misses the path. The rubric was fixed before any tool was scored, so the weight reflects the site's subject and was not set to suit a result. Read the situation pages for rankings under other weights, such as removes access at 25. **Why does Act rank first?** Act totals 5.3 against 4.8 for Sonrai Security and 4.2 for Wiz. Act scores 6 on identity, network and AI access together, against 4 for Sonrai, and 7 on safe to enforce, against 5. Act says it simulates changes against a digital twin of the customer's cloud and models identity, network and AI access in one Access Map. **Where is Act behind?** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents). It does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. **Who is second, and where does it lead?** Sonrai Security is second at 4.8. Sonrai leads Act on removes access, 5 against 3, because its organization-level policies block once deployed, while Act is read-only and delivers each change as a pull request or ticket. Sonrai also leads on native controls and evidence for auditors. The regulated and small-team pages put the gap at 0.1. **Which tools model the network path as well as identity?** In the pages read, Act models identity, network and AI together, and Wiz models identities and network with AI discovery. Prisma Cloud has a separate Cloud Network Analyzer, and Microsoft's security graph includes network connections and internet exposure. P0's connector documentation says it is not a fix for network reachability. For most of the others, none was described on 2026-11-10. **Why is AWS IAM Access Analyzer scored so low on coverage?** AWS IAM Access Analyzer scores 2 on identity, network and AI access together. Its unused access analyzers report unused roles, access keys and passwords, and its external and internal analyzers cover only resource-based policies. No network or AI model is described in the pages read. It still fits check ab-09, which compares granted permissions with those used. **Why is Microsoft scored on Defender for Cloud?** Microsoft's page says Entra Permissions Management was no longer available for purchase from April 1, 2025, and that on November 1, 2025 the product was to be retired. The profile therefore covers CIEM in Microsoft Defender for Cloud, which covers Entra ID, AWS IAM and Google Cloud IAM identities and shows network connections and AI agents in its graph. **Why are Amazon Q Business and Copilot Studio agents not on the map?** AWS documents Amazon Q Business actions as plugins to third-party services, not as actions on AWS resources, and says Q Business is no longer open to new customers. For Copilot Studio agents, none found on 2026-11-10 documents them acting on Azure resources. The map covers agents that act inside a cloud account. **Does this site cover MCP gateways, agent identity products or prompt filtering?** No. Those are different categories, covered by other publications, and nothing here ranks them. The tools are scored on what they show or change about the reach of an agent in a cloud account. Prompt and output inspection needs a different tool from the cloud boundary that the map and the checklist describe. **Where do I send a correction?** Write to the editor address on the contact page with the page, the sentence and a source. A confirmed error is fixed, the old and new text are recorded on the corrections page with the date, and any ranking the fix moves is noted. Cells on the map rest on a quote and a URL, so a better provider page settles most questions. ## Related pages - [The Agent Access Map: the data, downloadable](/datasets/) - [The Agent Access Map: what each agent platform runs as and can reach](/map/) - [The Agent Boundary Checklist: 30 checks before an agent goes to production](/checklist/) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/faq/ --- # The Agent Access Map as data: eight agent setups, the identity, default permissions and boundaries of each Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. The Agent Access Map is published as open data: eight agent setups across AWS, Google Cloud and Azure, with the identity, default permissions, tools, network, logging and boundary mechanisms of each, plus the 30-check checklist. Every cell has a quote, a source URL and the date it was read. The dataset holds the eight map entries, the 30 checks with their AWS, Google Cloud and Azure answers, and the scores, weights and public reasons for the 14 tools. Each fact carries a documentation quote, its source URL and the read date, which is 10 Nov 2026 for the first edition of the map and the scores. Anyone can recompute a tool's total with the stated rule: sum of score times weight divided by the sum of the weights, rounded to one decimal. If a cell on the map looks wrong, the quote and the URL show what it rests on, and the corrections page explains how to report an error. ## Downloads - [The Agent Access Map (CSV)](/datasets/agent-access-map-2026-11-10.csv): 8 rows, edition 10 Nov 2026 - [The Agent Access Map (JSON)](/datasets/agent-access-map-2026-11-10.json): 8 agent platforms: the identity each runs as, default permissions, tools, network, logging and the boundaries that apply, each with its source. - [The Agent Boundary Checklist (CSV)](/datasets/agent-boundary-checklist-2026-11-10.csv): 30 rows, edition 10 Nov 2026 - [The Agent Boundary Checklist (JSON)](/datasets/agent-boundary-checklist-2026-11-10.json): 30 checks with the mechanism in AWS, Google Cloud and Azure and the documentation source. - [Tools scored on AI access coverage (CSV)](/datasets/ai-access-tools-2026-11-10.csv): 14 rows, edition 10 Nov 2026 - [Tools scored on AI access coverage (JSON)](/datasets/ai-access-tools-2026-11-10.json): 14 tools scored 0 to 10 on seven criteria with this site's weights. ## Cite this **Cite this.** Cite this line: 'AI Agent Access Control, The Agent Access Map, edition 10 Nov 2026, https://aiagentaccesscontrol.com/datasets/' ## Changelog - 10 Nov 2026: first edition. Every row read from the source linked in the row on 10 Nov 2026. ## Related pages - [The Agent Access Map: what each agent platform runs as and can reach](/map/) - [The Agent Boundary Checklist: 30 checks before an agent goes to production](/checklist/) - [What an AI agent can access in AWS](/clouds/aws/) ## Sources - Amazon Bedrock Agents (now Amazon Bedrock Agents Classic): https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html (read 10 Nov 2026) - Amazon Bedrock AgentCore (Runtime, Gateway, Identity): https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html (read 10 Nov 2026) - Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder): https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 10 Nov 2026) - act.security/platform: https://act.security/platform (read 10 Nov 2026) - softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement: https://softwareanalyst.io/reports/the-ciso-s-guide-to-cloud-security-intent-enforcement (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/datasets/ --- # A reference on what AI agents can reach in a cloud account, with the working shown Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. AI Agent Access Control is a reference on what an AI agent can reach once it runs in AWS, Azure or Google Cloud. It has the Agent Access Map, the 30-check Agent Boundary Checklist and 14 tools scored on AI access coverage, each fact tied to a page and the date it was read. The site has one subject: the cloud boundary around an AI agent. The map shows the identity an agent platform runs as and what it can reach. The checklist gives the security lead and the engineer 30 checks to close before production. The tool pages score 14 products on seven fixed criteria, with the scores, reasons and quotes. Everything on the map and the checklist comes from AWS, Google Cloud and Microsoft documentation read on 10 Nov 2026. Everything about a product comes from that product's own pages or a named report, and when a vendor describes its own product the text says so. When nothing was found, the text says none found and gives the date. Prompt and output filtering, MCP gateways, agent identity products and cloud security posture rankings are outside the scope. They are mentioned in a sentence where a tool overlaps with them, and they are not ranked or scored on any page of this site. A reader looking for those will find them covered by other publications, and the map and the checklist stay on the cloud boundary itself. ## Related pages - [Editorial policy](/editorial-policy/) - [Corrections](/corrections/) - [Contact](/contact/) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/about/ --- # How AI Agent Access Control is written Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Scores come from a published rubric with fixed weights. Every claim about a product is attributed to the product's own pages or to a named report and carries the date it was read. Scores are integers, totals are rounded to one decimal, and ties are shown as ties. Weights and criteria were fixed on 10 Nov 2026 before any tool was scored. If a weight ever changes, the change will be listed on the corrections page with its date and the rankings it moved. The map and the checklist state only what the provider's documentation says on the read date. Where a provider page names a feature as preview, the text says preview. A renamed product stays on the map under its current name, with the old name recorded. ## Related pages - [Corrections](/corrections/) - [Contact](/contact/) - [Privacy](/privacy/) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/editorial-policy/ --- # Corrections and the correction rule Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Send a correction to the editor with the page, the sentence and a source. A confirmed error is fixed, the old and new text are recorded here with the date, and any ranking the fix moves is noted. Unconfirmed claims are not changed. No corrections have been recorded for the first edition of 10 Nov 2026. No correction has been published since the first edition of 10 Nov 2026. Before that edition the scores went through three blind scoring rounds and three reviews; 76 score changes were logged, each with the fact that caused it, its source and its read date, and the scores printed on this site are the final ones. From here on, every change to a published score is listed on this page with the old value, the new value, the fact and the date. ## Related pages - [Contact](/contact/) - [Privacy](/privacy/) - [Terms](/terms/) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/corrections/ --- # Contact AI Agent Access Control Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Write to editor@aiagentaccesscontrol.com for corrections, evidence a vendor wants considered, or questions about a score or a map entry. Include the page address and, for a correction, the exact sentence and a link to the source, so the editor can check it against the page and the read date. Replies come from the editor by email. The site has no forms, no accounts and no newsletter. ## Related pages - [Privacy](/privacy/) - [Terms](/terms/) - [Seven criteria, fixed weights, and coverage of identity, network and AI access carrying the most](/method/) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/contact/ --- # Privacy Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. AI Agent Access Control sets no cookies and runs no analytics script. The web server keeps ordinary access logs (address, page, time and browser string) for operating the site and for finding faults, and nothing else is collected about visitors, so there is no profile to request or delete. Email sent to the editor address is kept only as long as needed to answer it. Outbound links to vendor sites may carry source and campaign parameters in the address, and the vendor's own privacy policy then applies. ## Related pages - [Terms](/terms/) - [Seven criteria, fixed weights, and coverage of identity, network and AI access carrying the most](/method/) - [30 questions about ai agent access in the cloud](/faq/) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/privacy/ --- # Terms of use Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. The content is a reference, as of the read dates shown. Cloud services change quickly, so check the provider's current documentation before you approve an agent. Scores and map entries describe the pages read on those dates and are not a guarantee of how a service behaves in your account. You may quote short passages and reuse the dataset with credit, using the cite line on the datasets page. Product names are the property of their owners. ## Related pages - [Seven criteria, fixed weights, and coverage of identity, network and AI access carrying the most](/method/) - [30 questions about ai agent access in the cloud](/faq/) - [The Agent Access Map: the data, downloadable](/datasets/) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/terms/ --- # The Agent Access Map: the identity each agent platform runs as and what it can reach Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. The Agent Access Map lists eight agent setups across AWS, Google Cloud and Azure: five managed agent platforms and three self-hosted patterns. For each it gives the identity the agent runs as, its default permissions, the tools it calls, the network path, the logs and the boundary mechanisms that apply, all from provider documentation. Read the map one row at a time. Start with the identity column: it says whose authority the agent acts with. Then read default permissions, which show what the agent holds before you change anything, and the boundary mechanisms, which show what you can use to cap it. Where the provider documents no answer, the cell says none found on 2026-11-10, which is a statement about the pages read. The map was built from provider documentation only, read on 10 Nov 2026. Every cell carries a quote and a source URL from AWS, Google Cloud or Microsoft pages. Three renames are recorded: Amazon Bedrock Agents is now Amazon Bedrock Agents Classic, Agent Runtime was formerly Vertex AI Agent Engine, and Microsoft Foundry Agent Service was formerly Azure AI Foundry Agent Service. Amazon Q Business and Copilot Studio agents are left out, with the reasons recorded. ## The identity it runs as | Agent platform | Cloud | The identity it runs as | | --- | --- | --- | | [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic)](/agent-platforms/bedrock-agents/) | AWS | AWS documents that the agent runs as an IAM service role that Amazon Bedrock assumes. You can let Bedrock create the role automatically or supply a custom service role. The trust policy names the bedrock.amazonaws.com service principal. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html)) | | [Amazon Bedrock AgentCore (Runtime, Gateway, Identity)](/agent-platforms/bedrock-agentcore/) | AWS | Runtime runs the agent under an IAM execution role that AgentCore assumes. Gateway uses a gateway service role for IAM-based outbound calls. AgentCore Identity is a separate service that manages credentials so agents can reach AWS resources and third-party services on behalf of users. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html)) | | [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder)](/agent-platforms/agent-runtime-gcp/) | Google Cloud | Google documents three choices: a per-agent IAM agent identity (recommended, a principal tied to the individual agent); the Google-managed AI Platform Reasoning Engine Service Agent used by default; or a custom service account. Service accounts are shared across agents you deploy. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) | | [Gemini Enterprise (agents hosted on Agent Runtime)](/agent-platforms/gemini-enterprise/) | Google Cloud | Gemini Enterprise registers an ADK agent that runs on Agent Runtime, so the agent runs under the Agent Runtime identity described in that entry. A Gemini Enterprise administrator can see the agent's identity, typically its SPIFFE ID, or the Agent Registry resource ID if no SPIFFE ID is published. ([source](https://cloud.google.com/gemini/enterprise/docs/agents-overview)) | | [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service)](/agent-platforms/foundry-agent-service/) | Azure | Each agent gets a Microsoft Entra agent identity, a special service principal created under an agent blueprint. Hosted agents get a dedicated Entra agent identity at deploy time. A project-wide managed identity also exists, and legacy agents share it, with no per-agent separation. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) | | [Self-hosted agent on AWS Lambda or Amazon EKS](/agent-platforms/self-hosted-aws/) | AWS | A Lambda function runs as its execution role. On EKS, EKS Pod Identity associates an IAM role with a Kubernetes service account in a namespace; IAM roles for service accounts (IRSA) is the older OIDC-based option. ([source](https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html)) | | [Self-hosted agent on Google Cloud Run or GKE](/agent-platforms/self-hosted-gcp/) | Google Cloud | Cloud Run runs the service as a service account (the Compute Engine default service account if you do not choose one). On GKE, Workload Identity Federation for GKE assigns distinct IAM identities to each application in the cluster. ([source](https://cloud.google.com/run/docs/securing/service-identity)) | | [Self-hosted agent on Azure Container Apps or AKS](/agent-platforms/self-hosted-azure/) | Azure | A container app can have a system-assigned managed identity (tied to the app) or user-assigned identities (standalone resources). On AKS, Microsoft Entra Workload ID uses a projected Kubernetes service account token and OIDC federation to give pods an Entra identity. ([source](https://learn.microsoft.com/en-us/azure/container-apps/managed-identity)) | ## What it gets by default | Agent platform | Cloud | What it gets by default | | --- | --- | --- | | [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic)](/agent-platforms/bedrock-agents/) | AWS | The documented service role policy covers invoking named foundation models, reading the OpenAPI schema objects in S3, and querying attached knowledge bases, with optional statements for guardrails, collaborators and code interpretation. The policy of the role the console creates automatically is not listed on that page: none found on 2026-11-10 for its exact contents. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html)) | | [Amazon Bedrock AgentCore (Runtime, Gateway, Identity)](/agent-platforms/bedrock-agentcore/) | AWS | AWS documents that the console can create the execution role for you if the console role has the listed IAM create permissions, and that Gateway can create a service role and attach permissions each time you add a target. AWS warns that the BedrockAgentCoreFullAccess managed policy, for the person or application using AgentCore, grants broad permissions. The contents of the auto-created execution role: none found on 2026-11-10 in the pages read. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html)) | | [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder)](/agent-platforms/agent-runtime-gcp/) | Google Cloud | With agent identity, an agent starts with the Agent Default Access role (basic project-wide logging and model calling) and the Agent Context Editor role (own sessions, memories and sandboxes only). With the default service agent, Google documents the Reasoning Engine Service Agent role. A custom service account likely needs Agent Platform User plus roles your code needs. Google notes that some features of these roles may be in Preview. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime/setup)) | | [Gemini Enterprise (agents hosted on Agent Runtime)](/agent-platforms/gemini-enterprise/) | Google Cloud | Permissions Gemini Enterprise itself grants to a registered agent: none found on 2026-11-10. Registering an ADK agent requires the Gemini Enterprise Admin role for the person doing it; the agent's own Google Cloud access comes from Agent Runtime. ([source](https://cloud.google.com/gemini/enterprise/docs/register-and-manage-an-adk-agent)) | | [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service)](/agent-platforms/foundry-agent-service/) | Azure | Microsoft documents that a newly created agent identity starts without resource permissions and that Foundry does not automatically assign RBAC roles to it; you or an administrator grant access through the target service. For people, Microsoft names Foundry Agent Consumer as the least-privilege role for those who only interact with agents. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) | | [Self-hosted agent on AWS Lambda or Amazon EKS](/agent-platforms/self-hosted-aws/) | AWS | When you create a function in the Lambda console, AWS documents that it creates an execution role with minimal permissions (basic logging to CloudWatch Logs). On EKS, if IMDS access is not restricted, a pod's containers can also reach the node IAM role. ([source](https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html)) | | [Self-hosted agent on Google Cloud Run or GKE](/agent-platforms/self-hosted-gcp/) | Google Cloud | Google documents that, depending on organization policy, the Compute Engine default service account might automatically be granted the Editor role on the project, and that the constraint iam.automaticIamGrantsForDefaultServiceAccounts is enforced by default for organizations created after May 3, 2024. ([source](https://cloud.google.com/run/docs/securing/service-identity)) | | [Self-hosted agent on Azure Container Apps or AKS](/agent-platforms/self-hosted-azure/) | Azure | Microsoft documents that for some resources you must configure role assignments for the app's managed identity, otherwise calls to services such as Key Vault and SQL Database are rejected even with a valid token. A default role grant: none found on 2026-11-10. ([source](https://learn.microsoft.com/en-us/azure/container-apps/managed-identity)) | ## Tools and actions | Agent platform | Cloud | Tools and actions | | --- | --- | --- | | [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic)](/agent-platforms/bedrock-agents/) | AWS | Tools are action groups defined by an OpenAPI schema or function details, usually backed by a Lambda function. Bedrock invokes the function through a resource-based policy on the function; the function then runs under its own Lambda execution role. An action can instead return control to the calling application, which then runs the action itself. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-lambda.html)) | | [Amazon Bedrock AgentCore (Runtime, Gateway, Identity)](/agent-platforms/bedrock-agentcore/) | AWS | Gateway turns APIs, Lambda functions and existing services into tools reachable through Gateway endpoints. For each target, outbound authorization can use the gateway service role (SigV4), the caller's IAM credentials, OAuth (client credentials, authorization code, or token exchange), token passthrough, or an API key, so the identity of a tool call depends on which option the target uses. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html)) | | [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder)](/agent-platforms/agent-runtime-gcp/) | Google Cloud | Agent code calls Google Cloud APIs through Application Default Credentials, which use the agent identity once IAM is configured. For external services, Google documents an auth manager that brokers 2-legged or 3-legged OAuth credentials. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) | | [Gemini Enterprise (agents hosted on Agent Runtime)](/agent-platforms/gemini-enterprise/) | Google Cloud | An agent that must reach Google Cloud resources for a user needs authorization details configured: an OAuth client, an authorization resource registered with Gemini Enterprise, and user consent. Other tool and action behavior sits in the Agent Runtime entry. ([source](https://cloud.google.com/gemini/enterprise/docs/register-and-manage-an-adk-agent)) | | [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service)](/agent-platforms/foundry-agent-service/) | Azure | For MCP and A2A tools, Agent Service passes a scoped agent identity token and the downstream resource checks the agent identity's RBAC role assignments. Tool connections can also use key-based access, the agent's or project's managed identity, or OAuth identity passthrough. OpenAPI tools support anonymous, API key or project managed identity authentication. With function calling, your own app executes the function. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) | | [Self-hosted agent on AWS Lambda or Amazon EKS](/agent-platforms/self-hosted-aws/) | AWS | The agent code calls AWS APIs through an SDK, signing requests with the role's temporary credentials; every tool the code calls runs as that same role unless the code assumes another role. ([source](https://docs.aws.amazon.com/eks/latest/userguide/pod-identities.html)) | | [Self-hosted agent on Google Cloud Run or GKE](/agent-platforms/self-hosted-gcp/) | Google Cloud | The agent code calls Google Cloud APIs as its service account or Kubernetes workload principal; Google documents using IAM policies to grant Kubernetes workloads access to specific Google Cloud APIs. ([source](https://cloud.google.com/kubernetes-engine/docs/concepts/workload-identity)) | | [Self-hosted agent on Azure Container Apps or AKS](/agent-platforms/self-hosted-azure/) | Azure | Workloads on AKS need Entra application credentials or managed identities to reach Entra-protected resources such as Key Vault and Microsoft Graph; calls run as that identity. ([source](https://learn.microsoft.com/en-us/azure/aks/workload-identity-overview)) | ## The agent's permissions or the user's | Agent platform | Cloud | The agent's permissions or the user's | | --- | --- | --- | | [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic)](/agent-platforms/bedrock-agents/) | AWS | none found on 2026-11-10 | | [Amazon Bedrock AgentCore (Runtime, Gateway, Identity)](/agent-platforms/bedrock-agentcore/) | AWS | Both are documented. Runtime can fetch OAuth tokens on behalf of end users when a user ID header or a JWT is supplied. Gateway outbound calls can use the caller's IAM credentials, a user-delegated authorization code grant, a token exchange that carries the user and agent identities, or the machine-to-machine client credentials grant that acts on the application's own behalf. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html)) | | [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder)](/agent-platforms/agent-runtime-gcp/) | Google Cloud | Both are documented. Google describes three identities: the user, the agent (acting on its own authority), and a delegated identity when the agent acts for a user with scoped 3-legged OAuth. Calls to Google Cloud APIs with agent identity are evaluated against the agent identity. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/identity-delegation-architecture)) | | [Gemini Enterprise (agents hosted on Agent Runtime)](/agent-platforms/gemini-enterprise/) | Google Cloud | Both. ADK agents receive the user's email address from Gemini Enterprise, and an agent that must access Google Cloud resources on behalf of a user needs the authorization details configured. Agent-level access follows the Agent Runtime entry. ([source](https://cloud.google.com/gemini/enterprise/docs/register-and-manage-an-adk-agent)) | | [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service)](/agent-platforms/foundry-agent-service/) | Azure | Both are documented. Application-only access uses the agent's own authority; delegated access acts on behalf of the signed-in user, which needs both the agent's authorization and the user's access. Microsoft labels agent identity acting on behalf of the user as preview. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) | | [Self-hosted agent on AWS Lambda or Amazon EKS](/agent-platforms/self-hosted-aws/) | AWS | The workload uses its own role credentials. Propagation of an end user's permissions to the workload: none found on 2026-11-10 in the Lambda and EKS pages read. ([source](https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html)) | | [Self-hosted agent on Google Cloud Run or GKE](/agent-platforms/self-hosted-gcp/) | Google Cloud | The workload uses its own service account or workload identity. Propagation of an end user's permissions to the workload: none found on 2026-11-10 in the Cloud Run and GKE pages read. ([source](https://cloud.google.com/kubernetes-engine/docs/concepts/workload-identity)) | | [Self-hosted agent on Azure Container Apps or AKS](/agent-platforms/self-hosted-azure/) | Azure | The workload uses its own managed or workload identity. Propagation of an end user's permissions to the workload: none found on 2026-11-10 in the Container Apps and AKS pages read. ([source](https://learn.microsoft.com/en-us/azure/container-apps/managed-identity)) | ## Network reachability | Agent platform | Cloud | Network reachability | | --- | --- | --- | | [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic)](/agent-platforms/bedrock-agents/) | AWS | AWS documents interface endpoints (AWS PrivateLink) for the bedrock-agent and bedrock-agent-runtime APIs, with an endpoint policy you can attach. AWS's VPC page lists model customization, batch inference and knowledge bases as the features it describes for VPC protection; egress controls for the agent's own orchestration: none found on 2026-11-10. Network egress of an action group Lambda function follows that function's own VPC settings. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/vpc-interface-endpoints.html)) | | [Amazon Bedrock AgentCore (Runtime, Gateway, Identity)](/agent-platforms/bedrock-agentcore/) | AWS | AWS documents VPC connectivity for Runtime and the built-in Code Interpreter and Browser tools: AgentCore creates network interfaces in your VPC through a service-linked role. A public subnet does not give internet access; AWS says internet access needs private subnets with a route to a NAT Gateway. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/agentcore-vpc.html)) | | [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder)](/agent-platforms/agent-runtime-gcp/) | Google Cloud | Google documents that the agent runs in a Google-managed network without access to your VPC unless you configure a Private Service Connect interface, which gives private egress to your VPC and on-premises resources. With VPC Service Controls, a PSC interface is required for internet access, and the project must be in the perimeter before the agent is deployed or the agent keeps public internet access. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/private-service-connect-interface)) | | [Gemini Enterprise (agents hosted on Agent Runtime)](/agent-platforms/gemini-enterprise/) | Google Cloud | Google documents that the connection between Agent Runtime and Gemini Enterprise is compliant with VPC Service Controls. Cross-project ADK agents connect over Google Cloud private networking. ([source](https://cloud.google.com/gemini/enterprise/docs/register-and-manage-an-adk-agent)) | | [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service)](/agent-platforms/foundry-agent-service/) | Azure | Microsoft documents a Standard Setup with private networking: no public egress and a delegated subnet from your virtual network. Private networking is available for prompt agents; hosted agents support bring-your-own virtual network, with each session in a VM-isolated sandbox connected to your VNet. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/virtual-networks)) | | [Self-hosted agent on AWS Lambda or Amazon EKS](/agent-platforms/self-hosted-aws/) | AWS | AWS documents attaching a Lambda function to a VPC: private subnets for private resources, and a public subnet does not give a function internet access. Network controls for pods in EKS: none found on 2026-11-10 in the pages read. ([source](https://docs.aws.amazon.com/lambda/latest/dg/configuration-vpc.html)) | | [Self-hosted agent on Google Cloud Run or GKE](/agent-platforms/self-hosted-gcp/) | Google Cloud | Cloud Run can send egress traffic to a VPC network with Direct VPC egress (recommended) or a VPC connector, giving access to resources with internal IP addresses. In GKE, by default all Pods in a cluster can communicate with each other until network policy is enforced. ([source](https://cloud.google.com/run/docs/configuring/connecting-vpc)) | | [Self-hosted agent on Azure Container Apps or AKS](/agent-platforms/self-hosted-azure/) | Azure | By default Container Apps is publicly accessible and can communicate only with internet-accessible endpoints; supplying your own virtual network adds control over outbound traffic and access to private endpoints. In AKS, all pods can send and receive traffic without limitations by default. ([source](https://learn.microsoft.com/en-us/azure/container-apps/networking)) | ## What is logged | Agent platform | Cloud | What is logged | | --- | --- | --- | | [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic)](/agent-platforms/bedrock-agents/) | AWS | AWS documents that InvokeAgent and InvokeInlineAgent calls are logged to CloudTrail as data events, which you must enable with advanced event selectors for the AWS::Bedrock::AgentAlias resource type. Model invocation logging is off until you turn it on. The agent's step-by-step reasoning is available through traces. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/logging-using-cloudtrail.html)) | | [Amazon Bedrock AgentCore (Runtime, Gateway, Identity)](/agent-platforms/bedrock-agentcore/) | AWS | AgentCore stores metrics, spans and logs in Amazon CloudWatch. AWS documents CloudTrail coverage for only some AgentCore capabilities (Gateway and Agent Registry are the ones named); CloudTrail coverage for Runtime API calls: none found on 2026-11-10 in the pages read. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/observability.html)) | | [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder)](/agent-platforms/agent-runtime-gcp/) | Google Cloud | Agent stdout and stderr go to Cloud Logging under reasoning_engine_stdout and reasoning_engine_stderr. Google documents that agent identity shows up in logs across Google Cloud services, and for user-delegated flows logs show both the user and the agent. Cloud Logging is not supported for child resources such as Sessions and Memory Bank. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/logging)) | | [Gemini Enterprise (agents hosted on Agent Runtime)](/agent-platforms/gemini-enterprise/) | Google Cloud | Google documents Cloud Audit Logs for administrative and data-access operations on Gemini Enterprise (Vertex AI Agent Builder) resources, viewable in Logs Explorer, with an option to enable Data Access audit logs. Logging of what the agent does inside Google Cloud sits in the Agent Runtime entry. ([source](https://cloud.google.com/gemini/enterprise/docs/audit-logging)) | | [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service)](/agent-platforms/foundry-agent-service/) | Azure | Microsoft documents three sources: Entra sign-in logs (which agent identity obtained a token for which resource, including on behalf of which user), Entra audit logs (blueprint and agent identity creation and deletion), and Foundry tracing in the connected Application Insights resource for tool calls. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) | | [Self-hosted agent on AWS Lambda or Amazon EKS](/agent-platforms/self-hosted-aws/) | AWS | CloudTrail records Lambda API calls with identity information on who made each request. By default CloudTrail does not log most data events. For EKS, control plane logging can send audit logs to CloudWatch Logs. ([source](https://docs.aws.amazon.com/lambda/latest/dg/logging-using-cloudtrail.html)) | | [Self-hosted agent on Google Cloud Run or GKE](/agent-platforms/self-hosted-gcp/) | Google Cloud | Google documents Admin Activity audit logs that cannot be disabled, and Data Access audit logs that must be explicitly enabled to receive them. ([source](https://cloud.google.com/kubernetes-engine/docs/how-to/audit-logging)) | | [Self-hosted agent on Azure Container Apps or AKS](/agent-platforms/self-hosted-azure/) | Azure | Container Apps send console logs (from your app) and system logs (from the service) to Log Analytics. For AKS, the Azure Monitor activity log collects some data automatically and control plane logs are resource logs sent through a diagnostic setting. ([source](https://learn.microsoft.com/en-us/azure/container-apps/log-monitoring)) | ## Boundaries that apply | Agent platform | Mechanism | How it applies | | --- | --- | --- | | [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic)](/agent-platforms/bedrock-agents/) | IAM identity-based policy on the agent service role | The service role's policy is the main limit on what the agent itself can call. AWS shows statements scoped to specific model, knowledge base and prompt ARNs. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html)) | | [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic)](/agent-platforms/bedrock-agents/) | Trust policy conditions (aws:SourceAccount, aws:SourceArn) | AWS recommends condition keys in the role's trust policy so only agents in your account can use the role, and says to replace the wildcard with specific agent IDs once they exist. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html)) | | [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic)](/agent-platforms/bedrock-agents/) | Resource-based policy on the action group Lambda function | The function only accepts calls from Bedrock when its resource-based policy allows the bedrock.amazonaws.com principal, which AWS scopes to the account and agent ARN. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html)) | | [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic)](/agent-platforms/bedrock-agents/) | SCPs and permissions boundaries | AWS documents SCPs and permissions boundaries as limits on IAM entities in general, so they cap the permissions of the service role and of the Lambda execution role as well. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/security-iam.html)) | | [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic)](/agent-platforms/bedrock-agents/) | VPC endpoint policy | An endpoint policy on the bedrock-agent-runtime interface endpoint can limit which principals and actions use that private path. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/vpc-interface-endpoints.html)) | | [Amazon Bedrock AgentCore (Runtime, Gateway, Identity)](/agent-platforms/bedrock-agentcore/) | IAM policy on the execution role and gateway service role | AWS says the execution role is what Runtime assumes to run the agent, so its policy limits what agent code can call in your account. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html)) | | [Amazon Bedrock AgentCore (Runtime, Gateway, Identity)](/agent-platforms/bedrock-agentcore/) | Resource-based policies on AgentCore resources | AWS documents resource-based policies to control access to runtime resources. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html)) | | [Amazon Bedrock AgentCore (Runtime, Gateway, Identity)](/agent-platforms/bedrock-agentcore/) | SCPs and permissions boundaries | AWS lists permissions boundaries and SCPs among the policy types that apply to AgentCore IAM entities. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/security-iam.html)) | | [Amazon Bedrock AgentCore (Runtime, Gateway, Identity)](/agent-platforms/bedrock-agentcore/) | IAM deny of InvokeAgentRuntimeForUser | Passing a user ID header to Runtime needs a separate IAM action, so you can deny that action where user ID delegation is not needed. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-oauth.html)) | | [Amazon Bedrock AgentCore (Runtime, Gateway, Identity)](/agent-platforms/bedrock-agentcore/) | AgentCore Policy at Gateway | AWS describes AgentCore Policy as intercepting every tool call at Gateway before execution, with rules for which tools agents can use and under what conditions. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html)) | | [Amazon Bedrock AgentCore (Runtime, Gateway, Identity)](/agent-platforms/bedrock-agentcore/) | VPC security groups | When Runtime uses VPC connectivity, security groups on the network interfaces control reachable destinations. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/agentcore-vpc.html)) | | [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder)](/agent-platforms/agent-runtime-gcp/) | IAM allow policies on the agent identity | The agent's access to Google Cloud resources follows the roles you grant the agent identity. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/manage-agent-access)) | | [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder)](/agent-platforms/agent-runtime-gcp/) | IAM deny policy | Google shows a deny policy that targets all agent identities in the organization to block actions such as deleting buckets. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) | | [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder)](/agent-platforms/agent-runtime-gcp/) | Principal Access Boundary policy | Google documents a PAB policy that limits the resources an agent may access despite other permissions it might have. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) | | [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder)](/agent-platforms/agent-runtime-gcp/) | Context-Aware Access (mTLS binding) | By default, agent identity credentials are bound by a Google-managed Context-Aware Access policy to the trusted runtime, which Google says makes stolen credentials un-replayable. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) | | [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder)](/agent-platforms/agent-runtime-gcp/) | VPC Service Controls | A service perimeter can limit data exfiltration, and the project must be in the perimeter before deploying the agent. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/general/vpc-service-controls)) | | [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder)](/agent-platforms/agent-runtime-gcp/) | Organization policy for cross-project service accounts | Using a custom service account from another project requires that the iam.disableCrossProjectServiceAccountUsage organization policy is not enforced in the service account's project. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime/setup)) | | [Gemini Enterprise (agents hosted on Agent Runtime)](/agent-platforms/gemini-enterprise/) | Gemini Enterprise IAM roles | Roles such as Gemini Enterprise Admin, Discovery Engine Editor and Viewer control who manages agents; users with the discoveryengine.agents.manage permission can view a private agent's details but not use it. ([source](https://cloud.google.com/gemini/enterprise/docs/agents-overview)) | | [Gemini Enterprise (agents hosted on Agent Runtime)](/agent-platforms/gemini-enterprise/) | Agent sharing state | A private agent is meant for personal use and only its creator can chat with it. ([source](https://cloud.google.com/gemini/enterprise/docs/agents-overview)) | | [Gemini Enterprise (agents hosted on Agent Runtime)](/agent-platforms/gemini-enterprise/) | VPC Service Controls | The Agent Runtime connection is compatible with VPC-SC perimeters. ([source](https://cloud.google.com/gemini/enterprise/docs/register-and-manage-an-adk-agent)) | | [Gemini Enterprise (agents hosted on Agent Runtime)](/agent-platforms/gemini-enterprise/) | Agent Runtime IAM, deny, PAB | Limits on what the agent can call in Google Cloud are the Agent Runtime boundaries. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) | | [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service)](/agent-platforms/foundry-agent-service/) | Azure RBAC on the agent identity | The downstream resource checks the agent identity's role assignments before granting or denying access. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) | | [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service)](/agent-platforms/foundry-agent-service/) | Foundry RBAC roles and agent-scope assignments | Role assignments at the agent scope are currently evaluated only for agent endpoint access, so they limit who can call a specific agent, not what the agent can reach. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/rbac-foundry)) | | [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service)](/agent-platforms/foundry-agent-service/) | Azure Policy | Microsoft documents Azure Policy for auditing and governing Foundry resource state. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/how-to/azure-policy)) | | [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service)](/agent-platforms/foundry-agent-service/) | Virtual network isolation | The Standard Setup with private networking removes public egress and uses your delegated subnet. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/virtual-networks)) | | [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service)](/agent-platforms/foundry-agent-service/) | Hosted agent sandbox | Each hosted-agent session runs in its own VM-isolated sandbox that holds no blueprint credential. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) | | [Self-hosted agent on AWS Lambda or Amazon EKS](/agent-platforms/self-hosted-aws/) | IAM policy on the execution role or pod role | The role's permissions policy is the main limit on what the agent can do in the account. ([source](https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html)) | | [Self-hosted agent on AWS Lambda or Amazon EKS](/agent-platforms/self-hosted-aws/) | Pod Identity trust policy conditions | Session tags let the trust policy restrict which service accounts, namespaces and clusters can use the role. ([source](https://docs.aws.amazon.com/eks/latest/userguide/pod-id-role.html)) | | [Self-hosted agent on AWS Lambda or Amazon EKS](/agent-platforms/self-hosted-aws/) | SCPs and permissions boundaries | AWS documents both as maximum-permission controls that apply to IAM entities. ([source](https://docs.aws.amazon.com/lambda/latest/dg/security-iam.html)) | | [Self-hosted agent on AWS Lambda or Amazon EKS](/agent-platforms/self-hosted-aws/) | IMDS restriction | With IMDS access restricted, a pod's containers can retrieve credentials only for the role tied to their service account. ([source](https://docs.aws.amazon.com/eks/latest/userguide/pod-identities.html)) | | [Self-hosted agent on Google Cloud Run or GKE](/agent-platforms/self-hosted-gcp/) | IAM roles on the service account | The roles granted to the service account decide what the agent can reach. ([source](https://cloud.google.com/iam/docs/best-practices-service-accounts)) | | [Self-hosted agent on Google Cloud Run or GKE](/agent-platforms/self-hosted-gcp/) | Organization policy iam.automaticIamGrantsForDefaultServiceAccounts | Google recommends enforcing this constraint to stop the automatic Editor grant on default service accounts. ([source](https://cloud.google.com/run/docs/securing/service-identity)) | | [Self-hosted agent on Google Cloud Run or GKE](/agent-platforms/self-hosted-gcp/) | IAM deny policies | Deny policies set guardrails on access to Google Cloud resources regardless of allow policies. ([source](https://cloud.google.com/iam/docs/deny-overview)) | | [Self-hosted agent on Google Cloud Run or GKE](/agent-platforms/self-hosted-gcp/) | Principal Access Boundary policies | PAB policies define the resources a principal is eligible to access. ([source](https://cloud.google.com/iam/docs/principal-access-boundary-policies)) | | [Self-hosted agent on Azure Container Apps or AKS](/agent-platforms/self-hosted-azure/) | Azure RBAC role assignments | You use role-based access control to grant specific permissions to the managed identity. ([source](https://learn.microsoft.com/en-us/azure/container-apps/managed-identity)) | | [Self-hosted agent on Azure Container Apps or AKS](/agent-platforms/self-hosted-azure/) | Azure Policy | Azure Policy enforces organizational standards and assesses compliance at scale. ([source](https://learn.microsoft.com/en-us/azure/governance/policy/overview)) | | [Self-hosted agent on Azure Container Apps or AKS](/agent-platforms/self-hosted-azure/) | Deny assignments | A deny assignment attaches deny actions to a principal at a scope; Microsoft notes customer-managed deny assignments are in private preview. ([source](https://learn.microsoft.com/en-us/azure/role-based-access-control/deny-assignments)) | | [Self-hosted agent on Azure Container Apps or AKS](/agent-platforms/self-hosted-azure/) | Kubernetes network policy on AKS | Network policies define rules for traffic between pods. ([source](https://learn.microsoft.com/en-us/azure/aks/use-network-policies)) | ## Cite this **Cite this.** Cite this line: 'AI Agent Access Control, The Agent Access Map, edition 10 Nov 2026, https://aiagentaccesscontrol.com/datasets/' **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [The Agent Boundary Checklist: 30 checks before an agent goes to production](/checklist/) - [What an AI agent can access in AWS](/clouds/aws/) - [What an AI agent can access in Google Cloud](/clouds/gcp/) ## Sources - docs.aws.amazon.com/bedrock/latest/userguide/agents.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/agents-classic-maintenance-mode.htm: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-classic-maintenance-mode.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/agents-lambda.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-lambda.html (read 10 Nov 2026) - docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html: https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/agents-how.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-how.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/vpc-interface-endpoints.html: https://docs.aws.amazon.com/bedrock/latest/userguide/vpc-interface-endpoints.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/usingVPC.html: https://docs.aws.amazon.com/bedrock/latest/userguide/usingVPC.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/logging-using-cloudtrail.html: https://docs.aws.amazon.com/bedrock/latest/userguide/logging-using-cloudtrail.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html: https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/security-iam.html: https://docs.aws.amazon.com/bedrock/latest/userguide/security-iam.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/map/ --- # 8 agent platforms, and the identity each one runs as Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. The map covers 8 ways to run an AI agent in a cloud account: 5 managed agent platforms and 3 self-hosted patterns, one per cloud. Each page gives the identity the agent runs as, what it gets by default, how its tools are called, its network paths and its logs, from provider documentation. - [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic)](/agent-platforms/bedrock-agents/): AWS: AWS documents that the agent runs as an IAM service role that Amazon Bedrock assumes. - [Amazon Bedrock AgentCore (Runtime, Gateway, Identity)](/agent-platforms/bedrock-agentcore/): AWS: Runtime runs the agent under an IAM execution role that AgentCore assumes. - [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder)](/agent-platforms/agent-runtime-gcp/): Google Cloud: Google documents three choices: a per-agent IAM agent identity (recommended, a principal tied to the individual agent); the Google-managed AI Platform Reasoning Engine Service Agent used by default; or a custom service account. - [Gemini Enterprise (agents hosted on Agent Runtime)](/agent-platforms/gemini-enterprise/): Google Cloud: Gemini Enterprise registers an ADK agent that runs on Agent Runtime, so the agent runs under the Agent Runtime identity described in that entry. - [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service)](/agent-platforms/foundry-agent-service/): Azure: Each agent gets a Microsoft Entra agent identity, a special service principal created under an agent blueprint. - [Self-hosted agent on AWS Lambda or Amazon EKS](/agent-platforms/self-hosted-aws/): AWS: A Lambda function runs as its execution role. - [Self-hosted agent on Google Cloud Run or GKE](/agent-platforms/self-hosted-gcp/): Google Cloud: Cloud Run runs the service as a service account (the Compute Engine default service account if you do not choose one). - [Self-hosted agent on Azure Container Apps or AKS](/agent-platforms/self-hosted-azure/): Azure: A container app can have a system-assigned managed identity (tied to the app) or user-assigned identities (standalone resources). **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [Tools scored on AI access coverage](/tools/) - [Picks by situation](/situations/) - [Guides](/guides/) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/agent-platforms/ --- # Amazon Bedrock Agents Classic runs your agent as an IAM service role, and AWS has closed it to new sign-ups Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. AWS documents that the agent runs as an IAM service role that Amazon Bedrock assumes. The documented service role policy covers invoking named foundation models, reading the OpenAPI schema objects in S3, and querying attached knowledge bases, with optional statements for guardrails, collaborators and code interpretation. AWS documents that the agent runs as an IAM service role that Amazon Bedrock assumes. You can let Bedrock create the role automatically or supply a custom service role, and the trust policy names the bedrock.amazonaws.com service principal. The documented policy covers invoking named foundation models, reading the OpenAPI schema objects in S3 and querying attached knowledge bases. The contents of the console-created role: none found on 2026-11-10. Tools are action groups, usually backed by a Lambda function that runs under its own execution role, so the function's role is a second thing to review. AWS says you can no longer sign up for Amazon Bedrock Agents Classic, existing users can continue, and it points to Amazon Bedrock AgentCore. Logging needs two switches: CloudTrail data events for the AgentAlias resource type, and model invocation logging, which is off by default. - **Status:** Renamed Amazon Bedrock Agents Classic. AWS documents that you can no longer sign up for it, existing users can continue, and the service is in maintenance mode. AWS points to Amazon Bedrock AgentCore for similar capabilities. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents.html)) - **The identity it runs as:** AWS documents that the agent runs as an IAM service role that Amazon Bedrock assumes. You can let Bedrock create the role automatically or supply a custom service role. The trust policy names the bedrock.amazonaws.com service principal. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html)) - **What it gets by default:** The documented service role policy covers invoking named foundation models, reading the OpenAPI schema objects in S3, and querying attached knowledge bases, with optional statements for guardrails, collaborators and code interpretation. The policy of the role the console creates automatically is not listed on that page: none found on 2026-11-10 for its exact contents. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html)) - **Tools and actions:** Tools are action groups defined by an OpenAPI schema or function details, usually backed by a Lambda function. Bedrock invokes the function through a resource-based policy on the function; the function then runs under its own Lambda execution role. An action can instead return control to the calling application, which then runs the action itself. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-lambda.html)) - **The agent's permissions or the user's:** none found on 2026-11-10 - **Network reachability:** AWS documents interface endpoints (AWS PrivateLink) for the bedrock-agent and bedrock-agent-runtime APIs, with an endpoint policy you can attach. AWS's VPC page lists model customization, batch inference and knowledge bases as the features it describes for VPC protection; egress controls for the agent's own orchestration: none found on 2026-11-10. Network egress of an action group Lambda function follows that function's own VPC settings. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/vpc-interface-endpoints.html)) - **What is logged:** AWS documents that InvokeAgent and InvokeInlineAgent calls are logged to CloudTrail as data events, which you must enable with advanced event selectors for the AWS::Bedrock::AgentAlias resource type. Model invocation logging is off until you turn it on. The agent's step-by-step reasoning is available through traces. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/logging-using-cloudtrail.html)) ## Boundaries that apply - **IAM identity-based policy on the agent service role.** The service role's policy is the main limit on what the agent itself can call. AWS shows statements scoped to specific model, knowledge base and prompt ARNs. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html)) - **Trust policy conditions (aws:SourceAccount, aws:SourceArn).** AWS recommends condition keys in the role's trust policy so only agents in your account can use the role, and says to replace the wildcard with specific agent IDs once they exist. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html)) - **Resource-based policy on the action group Lambda function.** The function only accepts calls from Bedrock when its resource-based policy allows the bedrock.amazonaws.com principal, which AWS scopes to the account and agent ARN. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html)) - **SCPs and permissions boundaries.** AWS documents SCPs and permissions boundaries as limits on IAM entities in general, so they cap the permissions of the service role and of the Lambda execution role as well. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/security-iam.html)) - **VPC endpoint policy.** An endpoint policy on the bedrock-agent-runtime interface endpoint can limit which principals and actions use that private path. ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/vpc-interface-endpoints.html)) ## Checks to run for AWS - [Give each agent its own role or service account, not a shared one](/checklist/own-identity-per-agent/): Use one Lambda execution role per function, or one EKS Pod Identity association per service account, so each agent maps to its own IAM role. - [Use short-lived role credentials and no long-lived access keys for the agent](/checklist/short-lived-credentials/): Have the workload use temporary credentials through an IAM role rather than IAM user access keys. - [Limit who can pass, attach or impersonate the agent's role](/checklist/restrict-who-can-attach-role/): Grant iam:PassRole only for the specific roles a principal needs to pass to a service. - [Scope the role's trust to the specific agent, service account or workload](/checklist/scope-trust-to-the-agent/): Use trust policy conditions (for Pod Identity, session tags for service account, namespace and cluster; for services, aws:SourceArn and aws:SourceAccount). - [Do not run the agent as a default or pre-granted account](/checklist/no-default-accounts/): Replace the console-created default role with a role you wrote; AWS documents the Lambda console default as an execution role with minimal permissions (basic logging). - [Write the agent's permission policy from what its tools actually call](/checklist/least-privilege-policy/): Start from the documented minimum policy and narrow it; AWS says to use broad deployment permissions only for development and testing. - [Set an organization-level ceiling the agent's role cannot exceed](/checklist/set-permission-ceiling/): Apply SCPs, which specify the maximum available permissions for IAM roles in the organization, and permissions boundaries on the agent's role. - [Explicitly deny the agent IAM changes and destructive actions](/checklist/deny-iam-changes-and-destructive-actions/): Use an SCP or permissions boundary whose maximum excludes IAM write actions and destructive calls on protected resources. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [Amazon Bedrock AgentCore (Runtime, Gateway, Identity): identity, permissions and boundaries](/agent-platforms/bedrock-agentcore/) - [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder): identity, permissions and boundaries](/agent-platforms/agent-runtime-gcp/) - [Gemini Enterprise (agents hosted on Agent Runtime): identity, permissions and boundaries](/agent-platforms/gemini-enterprise/) ## Sources - docs.aws.amazon.com/bedrock/latest/userguide/agents.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/agents-classic-maintenance-mode.htm: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-classic-maintenance-mode.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/agents-lambda.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-lambda.html (read 10 Nov 2026) - docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html: https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/agents-how.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-how.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/vpc-interface-endpoints.html: https://docs.aws.amazon.com/bedrock/latest/userguide/vpc-interface-endpoints.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/usingVPC.html: https://docs.aws.amazon.com/bedrock/latest/userguide/usingVPC.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/logging-using-cloudtrail.html: https://docs.aws.amazon.com/bedrock/latest/userguide/logging-using-cloudtrail.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html: https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/security-iam.html: https://docs.aws.amazon.com/bedrock/latest/userguide/security-iam.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/agent-platforms/bedrock-agents/ --- # Amazon Bedrock AgentCore runs your agent as an IAM execution role you scope Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Runtime runs the agent under an IAM execution role that AgentCore assumes. AWS documents that the console can create the execution role for you if the console role has the listed IAM create permissions, and that Gateway can create a service role and attach permissions each time you add a target. Runtime runs the agent under an IAM execution role that AgentCore assumes, and Gateway uses a gateway service role for IAM-based outbound calls. AWS warns that the BedrockAgentCoreFullAccess managed policy, for the person or application using AgentCore, grants broad permissions. The contents of the auto-created execution role: none found on 2026-11-10. SCPs and permissions boundaries apply to these roles, and AgentCore Policy at Gateway intercepts tool calls before execution. For tool calls, the identity depends on the target. Outbound authorization can use the gateway service role, the caller's IAM credentials, OAuth, token exchange, token passthrough or an API key, so each target needs its own decision. Runtime can run in your VPC with security groups. AWS documents CloudTrail coverage for Gateway; for Runtime API calls: none found on 2026-11-10 in the pages read. - **Status:** AWS documents AgentCore as the platform it recommends in place of Amazon Bedrock Agents Classic. The overview page read on 2026-11-10 carries no preview or general availability label for Runtime, Gateway or Identity. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html)) - **The identity it runs as:** Runtime runs the agent under an IAM execution role that AgentCore assumes. Gateway uses a gateway service role for IAM-based outbound calls. AgentCore Identity is a separate service that manages credentials so agents can reach AWS resources and third-party services on behalf of users. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html)) - **What it gets by default:** AWS documents that the console can create the execution role for you if the console role has the listed IAM create permissions, and that Gateway can create a service role and attach permissions each time you add a target. AWS warns that the BedrockAgentCoreFullAccess managed policy, for the person or application using AgentCore, grants broad permissions. The contents of the auto-created execution role: none found on 2026-11-10 in the pages read. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html)) - **Tools and actions:** Gateway turns APIs, Lambda functions and existing services into tools reachable through Gateway endpoints. For each target, outbound authorization can use the gateway service role (SigV4), the caller's IAM credentials, OAuth (client credentials, authorization code, or token exchange), token passthrough, or an API key, so the identity of a tool call depends on which option the target uses. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html)) - **The agent's permissions or the user's:** Both are documented. Runtime can fetch OAuth tokens on behalf of end users when a user ID header or a JWT is supplied. Gateway outbound calls can use the caller's IAM credentials, a user-delegated authorization code grant, a token exchange that carries the user and agent identities, or the machine-to-machine client credentials grant that acts on the application's own behalf. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html)) - **Network reachability:** AWS documents VPC connectivity for Runtime and the built-in Code Interpreter and Browser tools: AgentCore creates network interfaces in your VPC through a service-linked role. A public subnet does not give internet access; AWS says internet access needs private subnets with a route to a NAT Gateway. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/agentcore-vpc.html)) - **What is logged:** AgentCore stores metrics, spans and logs in Amazon CloudWatch. AWS documents CloudTrail coverage for only some AgentCore capabilities (Gateway and Agent Registry are the ones named); CloudTrail coverage for Runtime API calls: none found on 2026-11-10 in the pages read. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/observability.html)) ## Boundaries that apply - **IAM policy on the execution role and gateway service role.** AWS says the execution role is what Runtime assumes to run the agent, so its policy limits what agent code can call in your account. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html)) - **Resource-based policies on AgentCore resources.** AWS documents resource-based policies to control access to runtime resources. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html)) - **SCPs and permissions boundaries.** AWS lists permissions boundaries and SCPs among the policy types that apply to AgentCore IAM entities. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/security-iam.html)) - **IAM deny of InvokeAgentRuntimeForUser.** Passing a user ID header to Runtime needs a separate IAM action, so you can deny that action where user ID delegation is not needed. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-oauth.html)) - **AgentCore Policy at Gateway.** AWS describes AgentCore Policy as intercepting every tool call at Gateway before execution, with rules for which tools agents can use and under what conditions. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html)) - **VPC security groups.** When Runtime uses VPC connectivity, security groups on the network interfaces control reachable destinations. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/agentcore-vpc.html)) ## Checks to run for AWS - [Give each agent its own role or service account, not a shared one](/checklist/own-identity-per-agent/): Use one Lambda execution role per function, or one EKS Pod Identity association per service account, so each agent maps to its own IAM role. - [Use short-lived role credentials and no long-lived access keys for the agent](/checklist/short-lived-credentials/): Have the workload use temporary credentials through an IAM role rather than IAM user access keys. - [Limit who can pass, attach or impersonate the agent's role](/checklist/restrict-who-can-attach-role/): Grant iam:PassRole only for the specific roles a principal needs to pass to a service. - [Scope the role's trust to the specific agent, service account or workload](/checklist/scope-trust-to-the-agent/): Use trust policy conditions (for Pod Identity, session tags for service account, namespace and cluster; for services, aws:SourceArn and aws:SourceAccount). - [Do not run the agent as a default or pre-granted account](/checklist/no-default-accounts/): Replace the console-created default role with a role you wrote; AWS documents the Lambda console default as an execution role with minimal permissions (basic logging). - [Write the agent's permission policy from what its tools actually call](/checklist/least-privilege-policy/): Start from the documented minimum policy and narrow it; AWS says to use broad deployment permissions only for development and testing. - [Set an organization-level ceiling the agent's role cannot exceed](/checklist/set-permission-ceiling/): Apply SCPs, which specify the maximum available permissions for IAM roles in the organization, and permissions boundaries on the agent's role. - [Explicitly deny the agent IAM changes and destructive actions](/checklist/deny-iam-changes-and-destructive-actions/): Use an SCP or permissions boundary whose maximum excludes IAM write actions and destructive calls on protected resources. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder): identity, permissions and boundaries](/agent-platforms/agent-runtime-gcp/) - [Gemini Enterprise (agents hosted on Agent Runtime): identity, permissions and boundaries](/agent-platforms/gemini-enterprise/) - [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service): identity, permissions and boundaries](/agent-platforms/foundry-agent-service/) ## Sources - docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/agents-classic-maintenance-mode.htm: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-classic-maintenance-mode.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock-agentcore/latest/devguide/identity.html: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/identity.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-oauth.html: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-oauth.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock-agentcore/latest/devguide/agentcore-vpc.html: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/agentcore-vpc.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock-agentcore/latest/devguide/observability.html: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/observability.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock-agentcore/latest/devguide/security-iam.html: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/security-iam.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/agent-platforms/bedrock-agentcore/ --- # Agent Runtime on Google Cloud gives each agent its own identity, and you grant it the roles Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Google documents three choices: a per-agent IAM agent identity (recommended, a principal tied to the individual agent); the Google-managed AI Platform Reasoning Engine Service Agent used by default; or a custom service account. Agent stdout and stderr go to Cloud Logging under reasoning_engine_stdout and reasoning_engine_stderr. Google documents three identity choices: a per-agent agent identity (recommended), the Google-managed Reasoning Engine Service Agent used by default, or a custom service account shared across agents. An agent identity starts with the Agent Default Access role (basic project-wide logging and model calling) and the Agent Context Editor role (its own sessions, memories and sandboxes). Google notes that some features of these roles may be in Preview. The product was formerly Vertex AI Agent Engine, and some role names still read that way. An IAM deny policy can target all agent identities in an organization, and a Principal Access Boundary limits the resources an agent may reach. By default the agent has no access to your VPC; a Private Service Connect interface gives private egress. Logs show the agent identity, and for user-delegated flows both the user and the agent. - **Status:** Renamed. The old agent-builder/agent-engine documentation URLs redirect to the Gemini Enterprise Agent Platform docs, where the product is called Agent Runtime; some role names still read Vertex AI Agent Engine. Google labels some features Preview or Pre-GA on these pages. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime/setup)) - **The identity it runs as:** Google documents three choices: a per-agent IAM agent identity (recommended, a principal tied to the individual agent); the Google-managed AI Platform Reasoning Engine Service Agent used by default; or a custom service account. Service accounts are shared across agents you deploy. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) - **What it gets by default:** With agent identity, an agent starts with the Agent Default Access role (basic project-wide logging and model calling) and the Agent Context Editor role (own sessions, memories and sandboxes only). With the default service agent, Google documents the Reasoning Engine Service Agent role. A custom service account likely needs Agent Platform User plus roles your code needs. Google notes that some features of these roles may be in Preview. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime/setup)) - **Tools and actions:** Agent code calls Google Cloud APIs through Application Default Credentials, which use the agent identity once IAM is configured. For external services, Google documents an auth manager that brokers 2-legged or 3-legged OAuth credentials. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) - **The agent's permissions or the user's:** Both are documented. Google describes three identities: the user, the agent (acting on its own authority), and a delegated identity when the agent acts for a user with scoped 3-legged OAuth. Calls to Google Cloud APIs with agent identity are evaluated against the agent identity. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/identity-delegation-architecture)) - **Network reachability:** Google documents that the agent runs in a Google-managed network without access to your VPC unless you configure a Private Service Connect interface, which gives private egress to your VPC and on-premises resources. With VPC Service Controls, a PSC interface is required for internet access, and the project must be in the perimeter before the agent is deployed or the agent keeps public internet access. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/private-service-connect-interface)) - **What is logged:** Agent stdout and stderr go to Cloud Logging under reasoning_engine_stdout and reasoning_engine_stderr. Google documents that agent identity shows up in logs across Google Cloud services, and for user-delegated flows logs show both the user and the agent. Cloud Logging is not supported for child resources such as Sessions and Memory Bank. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/logging)) ## Boundaries that apply - **IAM allow policies on the agent identity.** The agent's access to Google Cloud resources follows the roles you grant the agent identity. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/manage-agent-access)) - **IAM deny policy.** Google shows a deny policy that targets all agent identities in the organization to block actions such as deleting buckets. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) - **Principal Access Boundary policy.** Google documents a PAB policy that limits the resources an agent may access despite other permissions it might have. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) - **Context-Aware Access (mTLS binding).** By default, agent identity credentials are bound by a Google-managed Context-Aware Access policy to the trusted runtime, which Google says makes stolen credentials un-replayable. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) - **VPC Service Controls.** A service perimeter can limit data exfiltration, and the project must be in the perimeter before deploying the agent. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/general/vpc-service-controls)) - **Organization policy for cross-project service accounts.** Using a custom service account from another project requires that the iam.disableCrossProjectServiceAccountUsage organization policy is not enforced in the service account's project. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime/setup)) ## Checks to run for Google Cloud - [Give each agent its own role or service account, not a shared one](/checklist/own-identity-per-agent/): Create dedicated service accounts for each application, or use a per-agent agent identity on Agent Runtime. - [Use short-lived role credentials and no long-lived access keys for the agent](/checklist/short-lived-credentials/): Use workload identity or an attached service account instead of service account keys, which Google describes as powerful credentials that can present a security risk. - [Limit who can pass, attach or impersonate the agent's role](/checklist/restrict-who-can-attach-role/): Grant the Service Account User role only to principals that need to attach the service account to a resource. - [Scope the role's trust to the specific agent, service account or workload](/checklist/scope-trust-to-the-agent/): In IAM policies, grant access to the principal that corresponds to one Kubernetes workload, not to the whole cluster or project. - [Do not run the agent as a default or pre-granted account](/checklist/no-default-accounts/): Do not rely on the Compute Engine default service account, which can receive the Editor role; enforce the iam.automaticIamGrantsForDefaultServiceAccounts constraint. - [Write the agent's permission policy from what its tools actually call](/checklist/least-privilege-policy/): Grant the agent identity only the roles its code needs; Google describes agent identity as enabling a least-privilege approach. - [Set an organization-level ceiling the agent's role cannot exceed](/checklist/set-permission-ceiling/): Use Principal Access Boundary policies to define the resources the agent's principal may access. - [Explicitly deny the agent IAM changes and destructive actions](/checklist/deny-iam-changes-and-destructive-actions/): Use an IAM deny policy that targets agent identities; Google's example denies all agents in an organization from creating roles and deleting buckets. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [Gemini Enterprise (agents hosted on Agent Runtime): identity, permissions and boundaries](/agent-platforms/gemini-enterprise/) - [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service): identity, permissions and boundaries](/agent-platforms/foundry-agent-service/) - [Self-hosted agent on AWS Lambda or Amazon EKS: identity, permissions and boundaries](/agent-platforms/self-hosted-aws/) ## Sources - docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime/setup: https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime/setup (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/deploy-an-a: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/deploy-an-agent (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-ident: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/govern/identity-delegatio: https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/identity-delegation-architecture (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/private-ser: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/private-service-connect-interface (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/general/: https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/general/vpc-service-controls (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/logging: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/logging (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/manage-agen: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/manage-agent-access (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/agent-platforms/agent-runtime-gcp/ --- # Gemini Enterprise agents run as the Agent Runtime identity, so the limits are set there Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Gemini Enterprise registers an ADK agent that runs on Agent Runtime, so the agent runs under the Agent Runtime identity described in that entry. Permissions Gemini Enterprise itself grants to a registered agent: none found on 2026-11-10. Google documents that the connection between Agent Runtime and Gemini Enterprise is compliant with VPC Service Controls. The Gemini Enterprise pages show it acting on Google Cloud resources only through ADK agents hosted on Agent Runtime, so this entry covers that path. A registered agent runs under the Agent Runtime identity. A Gemini Enterprise administrator can see the agent's identity, typically its SPIFFE ID. Permissions Gemini Enterprise itself grants to a registered agent: none found on 2026-11-10. An agent that must reach Google Cloud resources for a user needs an OAuth client, an authorization resource registered with Gemini Enterprise, and user consent. Who manages agents is set by Gemini Enterprise IAM roles; what the agent can call is set by the Agent Runtime deny and Principal Access Boundary policies. The connection to Agent Runtime is compatible with VPC Service Controls, and Cloud Audit Logs cover administrative and data-access operations. - **Status:** Documented as Gemini Enterprise. The audit logging page still refers to Cloud Audit Logs for Vertex AI Agent Builder. A page titled Agentspace: none found on 2026-11-10 in the pages read. The Gemini Enterprise pages show it acting on Google Cloud resources only through ADK agents hosted on Agent Runtime, so this entry covers that path. ([source](https://cloud.google.com/gemini/enterprise/docs/audit-logging)) - **The identity it runs as:** Gemini Enterprise registers an ADK agent that runs on Agent Runtime, so the agent runs under the Agent Runtime identity described in that entry. A Gemini Enterprise administrator can see the agent's identity, typically its SPIFFE ID, or the Agent Registry resource ID if no SPIFFE ID is published. ([source](https://cloud.google.com/gemini/enterprise/docs/agents-overview)) - **What it gets by default:** Permissions Gemini Enterprise itself grants to a registered agent: none found on 2026-11-10. Registering an ADK agent requires the Gemini Enterprise Admin role for the person doing it; the agent's own Google Cloud access comes from Agent Runtime. ([source](https://cloud.google.com/gemini/enterprise/docs/register-and-manage-an-adk-agent)) - **Tools and actions:** An agent that must reach Google Cloud resources for a user needs authorization details configured: an OAuth client, an authorization resource registered with Gemini Enterprise, and user consent. Other tool and action behavior sits in the Agent Runtime entry. ([source](https://cloud.google.com/gemini/enterprise/docs/register-and-manage-an-adk-agent)) - **The agent's permissions or the user's:** Both. ADK agents receive the user's email address from Gemini Enterprise, and an agent that must access Google Cloud resources on behalf of a user needs the authorization details configured. Agent-level access follows the Agent Runtime entry. ([source](https://cloud.google.com/gemini/enterprise/docs/register-and-manage-an-adk-agent)) - **Network reachability:** Google documents that the connection between Agent Runtime and Gemini Enterprise is compliant with VPC Service Controls. Cross-project ADK agents connect over Google Cloud private networking. ([source](https://cloud.google.com/gemini/enterprise/docs/register-and-manage-an-adk-agent)) - **What is logged:** Google documents Cloud Audit Logs for administrative and data-access operations on Gemini Enterprise (Vertex AI Agent Builder) resources, viewable in Logs Explorer, with an option to enable Data Access audit logs. Logging of what the agent does inside Google Cloud sits in the Agent Runtime entry. ([source](https://cloud.google.com/gemini/enterprise/docs/audit-logging)) ## Boundaries that apply - **Gemini Enterprise IAM roles.** Roles such as Gemini Enterprise Admin, Discovery Engine Editor and Viewer control who manages agents; users with the discoveryengine.agents.manage permission can view a private agent's details but not use it. ([source](https://cloud.google.com/gemini/enterprise/docs/agents-overview)) - **Agent sharing state.** A private agent is meant for personal use and only its creator can chat with it. ([source](https://cloud.google.com/gemini/enterprise/docs/agents-overview)) - **VPC Service Controls.** The Agent Runtime connection is compatible with VPC-SC perimeters. ([source](https://cloud.google.com/gemini/enterprise/docs/register-and-manage-an-adk-agent)) - **Agent Runtime IAM, deny, PAB.** Limits on what the agent can call in Google Cloud are the Agent Runtime boundaries. ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) ## Checks to run for Google Cloud - [Give each agent its own role or service account, not a shared one](/checklist/own-identity-per-agent/): Create dedicated service accounts for each application, or use a per-agent agent identity on Agent Runtime. - [Use short-lived role credentials and no long-lived access keys for the agent](/checklist/short-lived-credentials/): Use workload identity or an attached service account instead of service account keys, which Google describes as powerful credentials that can present a security risk. - [Limit who can pass, attach or impersonate the agent's role](/checklist/restrict-who-can-attach-role/): Grant the Service Account User role only to principals that need to attach the service account to a resource. - [Scope the role's trust to the specific agent, service account or workload](/checklist/scope-trust-to-the-agent/): In IAM policies, grant access to the principal that corresponds to one Kubernetes workload, not to the whole cluster or project. - [Do not run the agent as a default or pre-granted account](/checklist/no-default-accounts/): Do not rely on the Compute Engine default service account, which can receive the Editor role; enforce the iam.automaticIamGrantsForDefaultServiceAccounts constraint. - [Write the agent's permission policy from what its tools actually call](/checklist/least-privilege-policy/): Grant the agent identity only the roles its code needs; Google describes agent identity as enabling a least-privilege approach. - [Set an organization-level ceiling the agent's role cannot exceed](/checklist/set-permission-ceiling/): Use Principal Access Boundary policies to define the resources the agent's principal may access. - [Explicitly deny the agent IAM changes and destructive actions](/checklist/deny-iam-changes-and-destructive-actions/): Use an IAM deny policy that targets agent identities; Google's example denies all agents in an organization from creating roles and deleting buckets. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service): identity, permissions and boundaries](/agent-platforms/foundry-agent-service/) - [Self-hosted agent on AWS Lambda or Amazon EKS: identity, permissions and boundaries](/agent-platforms/self-hosted-aws/) - [Self-hosted agent on Google Cloud Run or GKE: identity, permissions and boundaries](/agent-platforms/self-hosted-gcp/) ## Sources - cloud.google.com/gemini/enterprise/docs/audit-logging: https://cloud.google.com/gemini/enterprise/docs/audit-logging (read 10 Nov 2026) - cloud.google.com/gemini/enterprise/docs/register-and-manage-an-adk-agent: https://cloud.google.com/gemini/enterprise/docs/register-and-manage-an-adk-agent (read 10 Nov 2026) - cloud.google.com/gemini/enterprise/docs/agents-overview: https://cloud.google.com/gemini/enterprise/docs/agents-overview (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-ident: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/agent-platforms/gemini-enterprise/ --- # A Microsoft Foundry agent starts with no resource permissions, and you grant each role Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Each agent gets a Microsoft Entra agent identity, a special service principal created under an agent blueprint. Microsoft documents that a newly created agent identity starts without resource permissions and that Foundry does not automatically assign RBAC roles to it; you or an administrator grant access through the target service. Each agent gets a Microsoft Entra agent identity, a special service principal created under an agent blueprint. Microsoft documents that a new agent identity starts without resource permissions and that Foundry does not assign RBAC roles to it automatically. Legacy agents can share the project-wide managed identity, which gives no per-agent separation. The service was formerly Azure AI Foundry Agent Service, and Microsoft labels agent identity acting on behalf of the user as preview. Role assignments at the agent scope are currently evaluated only for agent endpoint access, so they limit who can call an agent, not what the agent can reach. What it can reach comes from Azure RBAC on the agent identity. A Standard Setup with private networking removes public egress. Entra sign-in logs show which agent identity got a token for which resource, and on behalf of which user. - **Status:** Documented as Microsoft Foundry Agent Service; documentation URLs still use the ai-foundry path, and Microsoft documents that the Foundry RBAC roles were renamed from the Azure AI names. Microsoft labels several features preview, including agent identity acting on behalf of the user. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/overview)) - **The identity it runs as:** Each agent gets a Microsoft Entra agent identity, a special service principal created under an agent blueprint. Hosted agents get a dedicated Entra agent identity at deploy time. A project-wide managed identity also exists, and legacy agents share it, with no per-agent separation. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) - **What it gets by default:** Microsoft documents that a newly created agent identity starts without resource permissions and that Foundry does not automatically assign RBAC roles to it; you or an administrator grant access through the target service. For people, Microsoft names Foundry Agent Consumer as the least-privilege role for those who only interact with agents. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) - **Tools and actions:** For MCP and A2A tools, Agent Service passes a scoped agent identity token and the downstream resource checks the agent identity's RBAC role assignments. Tool connections can also use key-based access, the agent's or project's managed identity, or OAuth identity passthrough. OpenAPI tools support anonymous, API key or project managed identity authentication. With function calling, your own app executes the function. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) - **The agent's permissions or the user's:** Both are documented. Application-only access uses the agent's own authority; delegated access acts on behalf of the signed-in user, which needs both the agent's authorization and the user's access. Microsoft labels agent identity acting on behalf of the user as preview. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) - **Network reachability:** Microsoft documents a Standard Setup with private networking: no public egress and a delegated subnet from your virtual network. Private networking is available for prompt agents; hosted agents support bring-your-own virtual network, with each session in a VM-isolated sandbox connected to your VNet. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/virtual-networks)) - **What is logged:** Microsoft documents three sources: Entra sign-in logs (which agent identity obtained a token for which resource, including on behalf of which user), Entra audit logs (blueprint and agent identity creation and deletion), and Foundry tracing in the connected Application Insights resource for tool calls. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) ## Boundaries that apply - **Azure RBAC on the agent identity.** The downstream resource checks the agent identity's role assignments before granting or denying access. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) - **Foundry RBAC roles and agent-scope assignments.** Role assignments at the agent scope are currently evaluated only for agent endpoint access, so they limit who can call a specific agent, not what the agent can reach. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/rbac-foundry)) - **Azure Policy.** Microsoft documents Azure Policy for auditing and governing Foundry resource state. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/how-to/azure-policy)) - **Virtual network isolation.** The Standard Setup with private networking removes public egress and uses your delegated subnet. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/virtual-networks)) - **Hosted agent sandbox.** Each hosted-agent session runs in its own VM-isolated sandbox that holds no blueprint credential. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) ## Checks to run for Azure - [Give each agent its own role or service account, not a shared one](/checklist/own-identity-per-agent/): Use the per-agent Entra agent identity that Foundry creates, or a system-assigned managed identity that is tied to one container app. - [Use short-lived role credentials and no long-lived access keys for the agent](/checklist/short-lived-credentials/): Use a managed identity; Microsoft documents that you do not manage its credentials. - [Scope the role's trust to the specific agent, service account or workload](/checklist/scope-trust-to-the-agent/): Configure workload identity federation so one managed identity trusts one Kubernetes workload. - [Write the agent's permission policy from what its tools actually call](/checklist/least-privilege-policy/): Grant the least privilege needed and avoid broader roles at broader scopes. - [Set an organization-level ceiling the agent's role cannot exceed](/checklist/set-permission-ceiling/): Use deny assignments, which block listed actions even when a role assignment grants them (customer-managed deny assignments are documented as private preview), and apply governance conditions at management group scope. - [Explicitly deny the agent IAM changes and destructive actions](/checklist/deny-iam-changes-and-destructive-actions/): Use deny assignments to block specific actions even if a role assignment grants them. - [Compare the agent's granted permissions with what it used and remove the rest](/checklist/remove-unused-permissions/): Run recurring Microsoft Entra access reviews of role assignments. - [Grant roles at the narrowest scope that works](/checklist/narrowest-scope/): Assign the role at resource or resource group scope instead of subscription or management group scope. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [Self-hosted agent on AWS Lambda or Amazon EKS: identity, permissions and boundaries](/agent-platforms/self-hosted-aws/) - [Self-hosted agent on Google Cloud Run or GKE: identity, permissions and boundaries](/agent-platforms/self-hosted-gcp/) - [Self-hosted agent on Azure Container Apps or AKS: identity, permissions and boundaries](/agent-platforms/self-hosted-azure/) ## Sources - learn.microsoft.com/en-us/azure/ai-foundry/agents/overview: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/overview (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/concepts/rbac-foundry: https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/rbac-foundry (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/hosted-agents: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/hosted-agents (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/tools/openapi: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/tools/openapi (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/tools/function-calling: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/tools/function-calling (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/virtual-networks: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/virtual-networks (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/how-to/azure-policy: https://learn.microsoft.com/en-us/azure/ai-foundry/how-to/azure-policy (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/agent-platforms/foundry-agent-service/ --- # A self-hosted agent on Lambda or EKS has exactly the IAM role you give it Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. A Lambda function runs as its execution role. When you create a function in the Lambda console, AWS documents that it creates an execution role with minimal permissions (basic logging to CloudWatch Logs). AWS documents attaching a Lambda function to a VPC: private subnets for private resources, and a public subnet does not give a function internet access. A Lambda function runs as its execution role. On EKS, Pod Identity associates an IAM role with a Kubernetes service account in a namespace, and IRSA is the older OIDC-based option. AWS documents that the Lambda console creates an execution role with minimal permissions, basic logging to CloudWatch Logs. Every tool the code calls runs as that same role unless the code assumes another role. On EKS, if access to the instance metadata service is not restricted, a pod's containers can also reach the node IAM role. SCPs and permissions boundaries cap the role. CloudTrail records Lambda API calls with identity information, and by default does not log most data events. Network controls for pods in EKS: none found on 2026-11-10 in the pages read. So is propagation of an end user's permissions to the workload. - **Status:** Self-hosted pattern, not a managed agent service. The entry describes AWS's documented identity, network and logging pieces for the compute the agent code runs on. - **The identity it runs as:** A Lambda function runs as its execution role. On EKS, EKS Pod Identity associates an IAM role with a Kubernetes service account in a namespace; IAM roles for service accounts (IRSA) is the older OIDC-based option. ([source](https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html)) - **What it gets by default:** When you create a function in the Lambda console, AWS documents that it creates an execution role with minimal permissions (basic logging to CloudWatch Logs). On EKS, if IMDS access is not restricted, a pod's containers can also reach the node IAM role. ([source](https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html)) - **Tools and actions:** The agent code calls AWS APIs through an SDK, signing requests with the role's temporary credentials; every tool the code calls runs as that same role unless the code assumes another role. ([source](https://docs.aws.amazon.com/eks/latest/userguide/pod-identities.html)) - **The agent's permissions or the user's:** The workload uses its own role credentials. Propagation of an end user's permissions to the workload: none found on 2026-11-10 in the Lambda and EKS pages read. ([source](https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html)) - **Network reachability:** AWS documents attaching a Lambda function to a VPC: private subnets for private resources, and a public subnet does not give a function internet access. Network controls for pods in EKS: none found on 2026-11-10 in the pages read. ([source](https://docs.aws.amazon.com/lambda/latest/dg/configuration-vpc.html)) - **What is logged:** CloudTrail records Lambda API calls with identity information on who made each request. By default CloudTrail does not log most data events. For EKS, control plane logging can send audit logs to CloudWatch Logs. ([source](https://docs.aws.amazon.com/lambda/latest/dg/logging-using-cloudtrail.html)) ## Boundaries that apply - **IAM policy on the execution role or pod role.** The role's permissions policy is the main limit on what the agent can do in the account. ([source](https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html)) - **Pod Identity trust policy conditions.** Session tags let the trust policy restrict which service accounts, namespaces and clusters can use the role. ([source](https://docs.aws.amazon.com/eks/latest/userguide/pod-id-role.html)) - **SCPs and permissions boundaries.** AWS documents both as maximum-permission controls that apply to IAM entities. ([source](https://docs.aws.amazon.com/lambda/latest/dg/security-iam.html)) - **IMDS restriction.** With IMDS access restricted, a pod's containers can retrieve credentials only for the role tied to their service account. ([source](https://docs.aws.amazon.com/eks/latest/userguide/pod-identities.html)) ## Checks to run for AWS - [Give each agent its own role or service account, not a shared one](/checklist/own-identity-per-agent/): Use one Lambda execution role per function, or one EKS Pod Identity association per service account, so each agent maps to its own IAM role. - [Use short-lived role credentials and no long-lived access keys for the agent](/checklist/short-lived-credentials/): Have the workload use temporary credentials through an IAM role rather than IAM user access keys. - [Limit who can pass, attach or impersonate the agent's role](/checklist/restrict-who-can-attach-role/): Grant iam:PassRole only for the specific roles a principal needs to pass to a service. - [Scope the role's trust to the specific agent, service account or workload](/checklist/scope-trust-to-the-agent/): Use trust policy conditions (for Pod Identity, session tags for service account, namespace and cluster; for services, aws:SourceArn and aws:SourceAccount). - [Do not run the agent as a default or pre-granted account](/checklist/no-default-accounts/): Replace the console-created default role with a role you wrote; AWS documents the Lambda console default as an execution role with minimal permissions (basic logging). - [Write the agent's permission policy from what its tools actually call](/checklist/least-privilege-policy/): Start from the documented minimum policy and narrow it; AWS says to use broad deployment permissions only for development and testing. - [Set an organization-level ceiling the agent's role cannot exceed](/checklist/set-permission-ceiling/): Apply SCPs, which specify the maximum available permissions for IAM roles in the organization, and permissions boundaries on the agent's role. - [Explicitly deny the agent IAM changes and destructive actions](/checklist/deny-iam-changes-and-destructive-actions/): Use an SCP or permissions boundary whose maximum excludes IAM write actions and destructive calls on protected resources. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [Self-hosted agent on Google Cloud Run or GKE: identity, permissions and boundaries](/agent-platforms/self-hosted-gcp/) - [Self-hosted agent on Azure Container Apps or AKS: identity, permissions and boundaries](/agent-platforms/self-hosted-azure/) - [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic): identity, permissions and boundaries](/agent-platforms/bedrock-agents/) ## Sources - docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html: https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html (read 10 Nov 2026) - docs.aws.amazon.com/eks/latest/userguide/pod-identities.html: https://docs.aws.amazon.com/eks/latest/userguide/pod-identities.html (read 10 Nov 2026) - docs.aws.amazon.com/eks/latest/userguide/iam-roles-for-service-accounts.html: https://docs.aws.amazon.com/eks/latest/userguide/iam-roles-for-service-accounts.html (read 10 Nov 2026) - docs.aws.amazon.com/lambda/latest/dg/configuration-vpc.html: https://docs.aws.amazon.com/lambda/latest/dg/configuration-vpc.html (read 10 Nov 2026) - docs.aws.amazon.com/lambda/latest/dg/logging-using-cloudtrail.html: https://docs.aws.amazon.com/lambda/latest/dg/logging-using-cloudtrail.html (read 10 Nov 2026) - docs.aws.amazon.com/eks/latest/userguide/control-plane-logs.html: https://docs.aws.amazon.com/eks/latest/userguide/control-plane-logs.html (read 10 Nov 2026) - docs.aws.amazon.com/eks/latest/userguide/pod-id-role.html: https://docs.aws.amazon.com/eks/latest/userguide/pod-id-role.html (read 10 Nov 2026) - docs.aws.amazon.com/lambda/latest/dg/security-iam.html: https://docs.aws.amazon.com/lambda/latest/dg/security-iam.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/agent-platforms/self-hosted-aws/ --- # A self-hosted agent on Cloud Run or GKE runs as a service account, and the default one might hold the Editor role Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Cloud Run runs the service as a service account (the Compute Engine default service account if you do not choose one). Google documents that, depending on organization policy, the Compute Engine default service account might automatically be granted the Editor role on the project, and that the constraint iam.automaticIamGrantsForDefaultServiceAccounts is enforced by default for organizations created after May 3, 2024. Cloud Run runs the service as a service account, the Compute Engine default service account if you do not choose one. Google documents that, depending on organization policy, that default account might automatically be granted the Editor role on the project. The iam.automaticIamGrantsForDefaultServiceAccounts constraint is enforced by default for organizations created after May 3, 2024. On GKE, Workload Identity Federation assigns distinct IAM identities to each application. In GKE, all Pods in a cluster can communicate with each other until network policy is enforced. Cloud Run reaches your VPC through Direct VPC egress or a connector. IAM deny policies and Principal Access Boundary policies set limits whatever the allow policies say. Admin Activity audit logs cannot be disabled; Data Access audit logs must be explicitly enabled. - **Status:** Self-hosted pattern, not a managed agent service. - **The identity it runs as:** Cloud Run runs the service as a service account (the Compute Engine default service account if you do not choose one). On GKE, Workload Identity Federation for GKE assigns distinct IAM identities to each application in the cluster. ([source](https://cloud.google.com/run/docs/securing/service-identity)) - **What it gets by default:** Google documents that, depending on organization policy, the Compute Engine default service account might automatically be granted the Editor role on the project, and that the constraint iam.automaticIamGrantsForDefaultServiceAccounts is enforced by default for organizations created after May 3, 2024. ([source](https://cloud.google.com/run/docs/securing/service-identity)) - **Tools and actions:** The agent code calls Google Cloud APIs as its service account or Kubernetes workload principal; Google documents using IAM policies to grant Kubernetes workloads access to specific Google Cloud APIs. ([source](https://cloud.google.com/kubernetes-engine/docs/concepts/workload-identity)) - **The agent's permissions or the user's:** The workload uses its own service account or workload identity. Propagation of an end user's permissions to the workload: none found on 2026-11-10 in the Cloud Run and GKE pages read. ([source](https://cloud.google.com/kubernetes-engine/docs/concepts/workload-identity)) - **Network reachability:** Cloud Run can send egress traffic to a VPC network with Direct VPC egress (recommended) or a VPC connector, giving access to resources with internal IP addresses. In GKE, by default all Pods in a cluster can communicate with each other until network policy is enforced. ([source](https://cloud.google.com/run/docs/configuring/connecting-vpc)) - **What is logged:** Google documents Admin Activity audit logs that cannot be disabled, and Data Access audit logs that must be explicitly enabled to receive them. ([source](https://cloud.google.com/kubernetes-engine/docs/how-to/audit-logging)) ## Boundaries that apply - **IAM roles on the service account.** The roles granted to the service account decide what the agent can reach. ([source](https://cloud.google.com/iam/docs/best-practices-service-accounts)) - **Organization policy iam.automaticIamGrantsForDefaultServiceAccounts.** Google recommends enforcing this constraint to stop the automatic Editor grant on default service accounts. ([source](https://cloud.google.com/run/docs/securing/service-identity)) - **IAM deny policies.** Deny policies set guardrails on access to Google Cloud resources regardless of allow policies. ([source](https://cloud.google.com/iam/docs/deny-overview)) - **Principal Access Boundary policies.** PAB policies define the resources a principal is eligible to access. ([source](https://cloud.google.com/iam/docs/principal-access-boundary-policies)) ## Checks to run for Google Cloud - [Give each agent its own role or service account, not a shared one](/checklist/own-identity-per-agent/): Create dedicated service accounts for each application, or use a per-agent agent identity on Agent Runtime. - [Use short-lived role credentials and no long-lived access keys for the agent](/checklist/short-lived-credentials/): Use workload identity or an attached service account instead of service account keys, which Google describes as powerful credentials that can present a security risk. - [Limit who can pass, attach or impersonate the agent's role](/checklist/restrict-who-can-attach-role/): Grant the Service Account User role only to principals that need to attach the service account to a resource. - [Scope the role's trust to the specific agent, service account or workload](/checklist/scope-trust-to-the-agent/): In IAM policies, grant access to the principal that corresponds to one Kubernetes workload, not to the whole cluster or project. - [Do not run the agent as a default or pre-granted account](/checklist/no-default-accounts/): Do not rely on the Compute Engine default service account, which can receive the Editor role; enforce the iam.automaticIamGrantsForDefaultServiceAccounts constraint. - [Write the agent's permission policy from what its tools actually call](/checklist/least-privilege-policy/): Grant the agent identity only the roles its code needs; Google describes agent identity as enabling a least-privilege approach. - [Set an organization-level ceiling the agent's role cannot exceed](/checklist/set-permission-ceiling/): Use Principal Access Boundary policies to define the resources the agent's principal may access. - [Explicitly deny the agent IAM changes and destructive actions](/checklist/deny-iam-changes-and-destructive-actions/): Use an IAM deny policy that targets agent identities; Google's example denies all agents in an organization from creating roles and deleting buckets. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [Self-hosted agent on Azure Container Apps or AKS: identity, permissions and boundaries](/agent-platforms/self-hosted-azure/) - [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic): identity, permissions and boundaries](/agent-platforms/bedrock-agents/) - [Amazon Bedrock AgentCore (Runtime, Gateway, Identity): identity, permissions and boundaries](/agent-platforms/bedrock-agentcore/) ## Sources - cloud.google.com/run/docs/securing/service-identity: https://cloud.google.com/run/docs/securing/service-identity (read 10 Nov 2026) - cloud.google.com/kubernetes-engine/docs/concepts/workload-identity: https://cloud.google.com/kubernetes-engine/docs/concepts/workload-identity (read 10 Nov 2026) - cloud.google.com/run/docs/configuring/connecting-vpc: https://cloud.google.com/run/docs/configuring/connecting-vpc (read 10 Nov 2026) - cloud.google.com/kubernetes-engine/docs/how-to/network-policy: https://cloud.google.com/kubernetes-engine/docs/how-to/network-policy (read 10 Nov 2026) - cloud.google.com/kubernetes-engine/docs/how-to/audit-logging: https://cloud.google.com/kubernetes-engine/docs/how-to/audit-logging (read 10 Nov 2026) - cloud.google.com/iam/docs/best-practices-service-accounts: https://cloud.google.com/iam/docs/best-practices-service-accounts (read 10 Nov 2026) - cloud.google.com/iam/docs/deny-overview: https://cloud.google.com/iam/docs/deny-overview (read 10 Nov 2026) - cloud.google.com/iam/docs/principal-access-boundary-policies: https://cloud.google.com/iam/docs/principal-access-boundary-policies (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/agent-platforms/self-hosted-gcp/ --- # A self-hosted agent on Container Apps or AKS acts as a managed identity, and its network is open by default Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. A container app can have a system-assigned managed identity (tied to the app) or user-assigned identities (standalone resources). Microsoft documents that for some resources you must configure role assignments for the app's managed identity, otherwise calls to services such as Key Vault and SQL Database are rejected even with a valid token. A container app can have a system-assigned managed identity tied to the app, or user-assigned identities that are standalone resources. On AKS, Microsoft Entra Workload ID gives pods an Entra identity through a projected service account token. Microsoft documents that you must configure role assignments for the identity, or calls to services such as Key Vault and SQL Database are rejected. A default role grant: none found on 2026-11-10. By default Container Apps is publicly accessible and can reach only internet-accessible endpoints; supplying your own virtual network adds control over outbound traffic. In AKS, all pods can send and receive traffic without limitations by default. Azure Policy, RBAC role assignments and Kubernetes network policy bound the agent. Container Apps send console and system logs to Log Analytics. - **Status:** Self-hosted pattern, not a managed agent service. - **The identity it runs as:** A container app can have a system-assigned managed identity (tied to the app) or user-assigned identities (standalone resources). On AKS, Microsoft Entra Workload ID uses a projected Kubernetes service account token and OIDC federation to give pods an Entra identity. ([source](https://learn.microsoft.com/en-us/azure/container-apps/managed-identity)) - **What it gets by default:** Microsoft documents that for some resources you must configure role assignments for the app's managed identity, otherwise calls to services such as Key Vault and SQL Database are rejected even with a valid token. A default role grant: none found on 2026-11-10. ([source](https://learn.microsoft.com/en-us/azure/container-apps/managed-identity)) - **Tools and actions:** Workloads on AKS need Entra application credentials or managed identities to reach Entra-protected resources such as Key Vault and Microsoft Graph; calls run as that identity. ([source](https://learn.microsoft.com/en-us/azure/aks/workload-identity-overview)) - **The agent's permissions or the user's:** The workload uses its own managed or workload identity. Propagation of an end user's permissions to the workload: none found on 2026-11-10 in the Container Apps and AKS pages read. ([source](https://learn.microsoft.com/en-us/azure/container-apps/managed-identity)) - **Network reachability:** By default Container Apps is publicly accessible and can communicate only with internet-accessible endpoints; supplying your own virtual network adds control over outbound traffic and access to private endpoints. In AKS, all pods can send and receive traffic without limitations by default. ([source](https://learn.microsoft.com/en-us/azure/container-apps/networking)) - **What is logged:** Container Apps send console logs (from your app) and system logs (from the service) to Log Analytics. For AKS, the Azure Monitor activity log collects some data automatically and control plane logs are resource logs sent through a diagnostic setting. ([source](https://learn.microsoft.com/en-us/azure/container-apps/log-monitoring)) ## Boundaries that apply - **Azure RBAC role assignments.** You use role-based access control to grant specific permissions to the managed identity. ([source](https://learn.microsoft.com/en-us/azure/container-apps/managed-identity)) - **Azure Policy.** Azure Policy enforces organizational standards and assesses compliance at scale. ([source](https://learn.microsoft.com/en-us/azure/governance/policy/overview)) - **Deny assignments.** A deny assignment attaches deny actions to a principal at a scope; Microsoft notes customer-managed deny assignments are in private preview. ([source](https://learn.microsoft.com/en-us/azure/role-based-access-control/deny-assignments)) - **Kubernetes network policy on AKS.** Network policies define rules for traffic between pods. ([source](https://learn.microsoft.com/en-us/azure/aks/use-network-policies)) ## Checks to run for Azure - [Give each agent its own role or service account, not a shared one](/checklist/own-identity-per-agent/): Use the per-agent Entra agent identity that Foundry creates, or a system-assigned managed identity that is tied to one container app. - [Use short-lived role credentials and no long-lived access keys for the agent](/checklist/short-lived-credentials/): Use a managed identity; Microsoft documents that you do not manage its credentials. - [Scope the role's trust to the specific agent, service account or workload](/checklist/scope-trust-to-the-agent/): Configure workload identity federation so one managed identity trusts one Kubernetes workload. - [Write the agent's permission policy from what its tools actually call](/checklist/least-privilege-policy/): Grant the least privilege needed and avoid broader roles at broader scopes. - [Set an organization-level ceiling the agent's role cannot exceed](/checklist/set-permission-ceiling/): Use deny assignments, which block listed actions even when a role assignment grants them (customer-managed deny assignments are documented as private preview), and apply governance conditions at management group scope. - [Explicitly deny the agent IAM changes and destructive actions](/checklist/deny-iam-changes-and-destructive-actions/): Use deny assignments to block specific actions even if a role assignment grants them. - [Compare the agent's granted permissions with what it used and remove the rest](/checklist/remove-unused-permissions/): Run recurring Microsoft Entra access reviews of role assignments. - [Grant roles at the narrowest scope that works](/checklist/narrowest-scope/): Assign the role at resource or resource group scope instead of subscription or management group scope. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic): identity, permissions and boundaries](/agent-platforms/bedrock-agents/) - [Amazon Bedrock AgentCore (Runtime, Gateway, Identity): identity, permissions and boundaries](/agent-platforms/bedrock-agentcore/) - [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder): identity, permissions and boundaries](/agent-platforms/agent-runtime-gcp/) ## Sources - learn.microsoft.com/en-us/azure/container-apps/managed-identity: https://learn.microsoft.com/en-us/azure/container-apps/managed-identity (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/aks/workload-identity-overview: https://learn.microsoft.com/en-us/azure/aks/workload-identity-overview (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/container-apps/networking: https://learn.microsoft.com/en-us/azure/container-apps/networking (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/aks/use-network-policies: https://learn.microsoft.com/en-us/azure/aks/use-network-policies (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/container-apps/log-monitoring: https://learn.microsoft.com/en-us/azure/container-apps/log-monitoring (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/aks/monitor-aks: https://learn.microsoft.com/en-us/azure/aks/monitor-aks (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/governance/policy/overview: https://learn.microsoft.com/en-us/azure/governance/policy/overview (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/role-based-access-control/deny-assignments: https://learn.microsoft.com/en-us/azure/role-based-access-control/deny-assignments (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/agent-platforms/self-hosted-azure/ --- # The Agent Boundary Checklist: 30 checks before an agent goes to production Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. The Agent Boundary Checklist has 30 checks in seven sections: identity, permissions, data, network, tools, change control and logging. Each check says why it matters and gives the AWS, Google Cloud and Azure way to do it, with a documentation quote, so an approval meeting can mark each one done or open. Use it as the agenda of the approval meeting. Take the sections in order and ask the engineer to show the answer for each check in the cloud the agent runs in: the role, the ceiling policy, the endpoint, the log query. A check with no evidence is open. A check the cloud has no documented answer for is marked none found on 2026-11-10, and the meeting decides whether to accept the risk. The checks are about the cloud boundary around the agent, not about what the model says. Prompt and output filtering is a different job for a different tool, and this list does not cover it. Count the open checks at the end, record who owns each, and approve the agent only when the identity, permissions and logging sections are closed. ## Identity An agent should run as one identity of its own, with short-lived credentials and no long-lived keys. The checks cover a separate role or service account per agent, who may pass or attach that role, trust that names the specific workload, and avoiding default or pre-granted accounts. If two agents share an identity, one agent's access is every agent's access, and the logs cannot say which one acted. 1. [Give each agent its own role or service account, not a shared one](/checklist/own-identity-per-agent/). A shared identity means one agent's access is every agent's access, and logs cannot say which agent acted. A separate identity lets you grant, review and revoke access agent by agent. 2. [Use short-lived role credentials and no long-lived access keys for the agent](/checklist/short-lived-credentials/). A key stored in code, a config file or a container image stays valid until someone rotates it, so a leak gives an attacker lasting access. Temporary credentials from a role expire on their own. 3. [Limit who can pass, attach or impersonate the agent's role](/checklist/restrict-who-can-attach-role/). Anyone who can attach the agent's role to a new resource can borrow its access without ever being granted it. Limiting that right keeps the agent's permissions from spreading. 4. [Scope the role's trust to the specific agent, service account or workload](/checklist/scope-trust-to-the-agent/). A role that any workload in an account or cluster can assume is a role every compromised workload can use. Tying trust to one agent keeps a break-in elsewhere from reaching it. 5. [Do not run the agent as a default or pre-granted account](/checklist/no-default-accounts/). Default accounts are created for convenience and can carry broad roles that nobody chose for the agent. An agent that inherits one gets access no one reviewed. ## Permissions These checks decide how much the identity can do. Write the policy from what the tools actually call, set an organization-level ceiling the role cannot exceed, and deny IAM changes and destructive actions outright. Then compare what was granted with what was used, and grant at the narrowest scope that works. The ceiling matters most, because it still holds when someone later widens the agent's own policy by mistake. 1. [Write the agent's permission policy from what its tools actually call](/checklist/least-privilege-policy/). Broad starter policies make every tool call possible, including ones nobody planned. A policy limited to the real calls limits what a wrong or hijacked call can do. 2. [Set an organization-level ceiling the agent's role cannot exceed](/checklist/set-permission-ceiling/). A role policy can be edited by whoever manages that account, so it is a weak limit by itself. A ceiling set higher in the hierarchy holds even if the role's own policy is widened. 3. [Explicitly deny the agent IAM changes and destructive actions](/checklist/deny-iam-changes-and-destructive-actions/). An agent that can edit roles or delete data can widen its own access or cause damage no review would have approved. An explicit deny wins over any allow added later by mistake. 4. [Compare the agent's granted permissions with what it used and remove the rest](/checklist/remove-unused-permissions/). Permissions granted during a pilot tend to stay long after the agent stops using them. Each unused permission is access an attacker could use for nothing the agent needs. 5. [Grant roles at the narrowest scope that works](/checklist/narrowest-scope/). A role granted on a whole account, project or subscription covers every resource in it, including ones created later. Granting on specific resources keeps new resources out of the agent's reach by default. ## Data An agent is only as dangerous as the data it can read and move. These checks keep encryption keys under your control, block public access on storage the agent writes to, put the data stores inside a perimeter that stops copying data out, and limit access with tags or conditions instead of resource lists alone. Run them against the stores the agent's tools touch, not against every store in the account. 1. [Keep the encryption keys for agent data under your control and limit who can use them](/checklist/control-the-keys/). If the agent's role can use a key, it can read whatever that key protects. Customer-managed keys let you cut off access by changing the key policy or disabling the key. 2. [Block public access on storage the agent can write to](/checklist/block-public-exposure/). An agent that can change a bucket setting, or write to a bucket that already allows public reads, can expose data by accident. Account-level public access blocks stop that even when a policy says otherwise. 3. [Put data stores the agent reads inside a perimeter that blocks copying data out](/checklist/data-perimeter/). An agent allowed to read data can usually also send it somewhere else unless something stops the transfer. A perimeter turns an allowed read into a read that cannot leave. 4. [Limit data access with tags, attributes or conditions, not only resource lists](/checklist/condition-based-data-access/). Resource lists go stale as new data stores appear. Conditions on tags or attributes keep access tied to what the data is, so a new store does not become reachable by default. ## Network The network decides where the agent's traffic can go and what can call it. The checks run agent compute on private networks with no direct path to the internet, allow outbound traffic only to needed destinations, reach cloud APIs through private endpoints, and keep agent code from reading node or instance credentials. Defaults differ by cloud: Container Apps is public by default, AKS pods are open by default, and a Foundry Standard Setup removes public egress. 1. [Run agent compute on private networks with no direct path to the internet](/checklist/private-network-no-direct-internet/). Compute with a direct internet path can send data out or pull in instructions from anywhere. A private network makes every outside path an explicit choice. 2. [Allow outbound traffic only to the destinations the agent needs](/checklist/allowlist-egress/). Open outbound access lets a hijacked agent reach any server to send data or fetch tools. A short allow list makes unexpected destinations fail and show up in logs. 3. [Reach cloud APIs and data stores through private endpoints](/checklist/private-endpoints-for-cloud-apis/). Calls to cloud services over public addresses can be made from anywhere that holds valid credentials. Private endpoints keep the path inside your network, so stolen credentials are harder to use from outside. 4. [Keep the agent's code from reading node, instance or other workloads' credentials](/checklist/credential-isolation/). If agent code can read the host's metadata service, it can take the host's credentials, which are often broader than the agent's own. Isolating credentials keeps one compromised agent from becoming a compromised node. ## Tools A tool is where an agent's words become actions. These checks give each tool function its own least-privilege role, restrict which principals can invoke each tool endpoint, cap how much the agent can run and spend, and decide for every tool whether the action runs with the agent's authority or the signed-in user's. That last decision is the one reviewers most often find undocumented, so ask for it per tool. 1. [Give each tool function its own least-privilege role instead of the agent's role](/checklist/tool-own-role/). A tool that runs with the agent's full role can do everything the agent can, even if the tool needs one action. A narrow role per tool limits damage when one tool is abused or buggy. 2. [Restrict which principals can invoke each tool endpoint](/checklist/restrict-tool-invokers/). A tool endpoint that accepts calls from any caller can be used without the agent. Limiting invokers means only the intended agent can trigger the action. 3. [Cap how much the agent can run and spend](/checklist/cap-run-and-spend/). A looping or hijacked agent can start thousands of runs and a large bill before anyone notices. Limits on concurrency and budget alerts shrink the damage window. 4. [Decide for each tool whether the action runs with the agent's authority or the signed-in user's](/checklist/decide-user-or-agent-authority/). If a tool runs with the agent's authority, every user can reach whatever the agent can reach. If it runs with the user's authority, the agent can only do what that user may. ## Change control These checks keep the boundary in place after launch. Run agents in their own account, project or subscription, tag every agent resource with an owner and environment and enforce the tags, test permission and policy changes in a dry run before enforcing them, and remove the agent's identity and access when the agent is retired. For AWS the retirement step has no documented answer: none found on 2026-11-10. 1. [Run agents in their own account, project or subscription](/checklist/dedicated-account-project-subscription/). Agents placed beside production workloads share their blast radius, quotas and admin rights. A dedicated container lets you apply stricter policy and limits and remove it cleanly. 2. [Tag every agent resource with an owner and environment, and enforce the tags](/checklist/tag-agent-resources/). Untagged agent resources cannot be traced to an owner when something goes wrong. Tags also let policy and cost controls target agent resources specifically. 3. [Test permission and policy changes in a dry run before enforcing them](/checklist/test-before-enforcing/). A new deny or boundary can break an agent's tools in ways nobody saw coming, and a mistaken allow can open access without anyone noticing. Testing first shows the effect before it reaches production. 4. [Remove the agent's identity and access when the agent is retired](/checklist/retire-identity-with-agent/). Leftover identities from retired agents keep their permissions and nobody is watching them. Tying the identity to the agent's lifecycle removes the access with the agent. ## Logging If an incident happens, the logs must show which agent did what. The checks turn on organization-wide audit logging of control-plane actions, log reads and writes to the data the agent touches, make every logged action traceable to one agent identity, and alert on denied calls and on changes to the agent's role or policies. Data events and resource logs are off by default in all three clouds until you enable them. 1. [Turn on organization-wide audit logging of control-plane actions](/checklist/control-plane-audit-logs/). Without audit logs there is no record of which role or setting the agent changed, or who changed the agent's access. A log kept outside the agent's reach is the base for every investigation. 2. [Log reads and writes to the data the agent touches](/checklist/log-data-access/). Control-plane logs show configuration changes, not which records the agent read. Data-level logs are what show whether the agent touched something it should not have. 3. [Make every logged action traceable to the specific agent identity](/checklist/attribute-actions-to-the-agent/). If logs only show a shared role or a generic account, you cannot tell which agent did what. Per-agent attribution turns a log line into a decision about one agent. 4. [Alert on denied calls and on changes to the agent's role or policies](/checklist/alert-on-denied-and-role-changes/). A string of denied calls or an edit to the agent's role is often the first visible sign of misuse. Without an alert, these show up only after the damage. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [What an AI agent can access in AWS](/clouds/aws/) - [What an AI agent can access in Google Cloud](/clouds/gcp/) - [What an AI agent can access in Azure](/clouds/azure/) ## Sources - docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html: https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html (read 10 Nov 2026) - docs.aws.amazon.com/eks/latest/userguide/pod-identities.html: https://docs.aws.amazon.com/eks/latest/userguide/pod-identities.html (read 10 Nov 2026) - cloud.google.com/iam/docs/best-practices-service-accounts: https://cloud.google.com/iam/docs/best-practices-service-accounts (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-ident: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/container-apps/managed-identity: https://learn.microsoft.com/en-us/azure/container-apps/managed-identity (read 10 Nov 2026) - docs.aws.amazon.com/IAM/latest/UserGuide/best-practices.html: https://docs.aws.amazon.com/IAM/latest/UserGuide/best-practices.html (read 10 Nov 2026) - cloud.google.com/iam/docs/workload-identity-federation: https://cloud.google.com/iam/docs/workload-identity-federation (read 10 Nov 2026) - learn.microsoft.com/en-us/entra/identity/managed-identities-azure-resources/over: https://learn.microsoft.com/en-us/entra/identity/managed-identities-azure-resources/overview (read 10 Nov 2026) - docs.aws.amazon.com/IAM/latest/UserGuide/id_roles_use_passrole.html: https://docs.aws.amazon.com/IAM/latest/UserGuide/id_roles_use_passrole.html (read 10 Nov 2026) - cloud.google.com/iam/docs/service-account-permissions: https://cloud.google.com/iam/docs/service-account-permissions (read 10 Nov 2026) - docs.aws.amazon.com/eks/latest/userguide/pod-id-role.html: https://docs.aws.amazon.com/eks/latest/userguide/pod-id-role.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/ --- # Give each agent its own role or service account, not a shared one Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 1 of 30: Give each agent its own role or service account, not a shared one. A shared identity means one agent's access is every agent's access, and logs cannot say which agent acted. A separate identity lets you grant, review and revoke access agent by agent. ## Why it matters A shared identity means one agent's access is every agent's access, and logs cannot say which agent acted. A separate identity lets you grant, review and revoke access agent by agent. ## How, in each cloud - **AWS:** Use one Lambda execution role per function, or one EKS Pod Identity association per service account, so each agent maps to its own IAM role. - **Google Cloud:** Create dedicated service accounts for each application, or use a per-agent agent identity on Agent Runtime. - **Azure:** Use the per-agent Entra agent identity that Foundry creates, or a system-assigned managed identity that is tied to one container app. ## What the documentation says - "A Lambda Event Function's execution role is an AWS Identity and Access Management (IAM) role that grants the function permission to access" ([source](https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html)) - "Each EKS Pod Identity association maps a role to a service account in a namespace in the specified cluster." ([source](https://docs.aws.amazon.com/eks/latest/userguide/pod-identities.html)) - "Create dedicated service accounts for each application, and avoid using" ([source](https://cloud.google.com/iam/docs/best-practices-service-accounts)) - "Agent identity provides a per-agent identity that enables a least-privilege" ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) - "Per-agent identities let you scope permissions and audit activity to individual agents." ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) - "A system-assigned identity is tied to your container app and is deleted when your container app is deleted." ([source](https://learn.microsoft.com/en-us/azure/container-apps/managed-identity)) - [Next: Use short-lived role credentials and no long-lived access keys for the agent](/checklist/short-lived-credentials/) ## Related pages - [Use short-lived role credentials and no long-lived access keys for the agent: the check, and how in AWS, Google Cloud and Azure](/checklist/short-lived-credentials/) - [Limit who can pass, attach or impersonate the agent's role: the check, and how in AWS, Google Cloud and Azure](/checklist/restrict-who-can-attach-role/) - [Scope the role's trust to the specific agent, service account or workload: the check, and how in AWS, Google Cloud and Azure](/checklist/scope-trust-to-the-agent/) ## Sources - docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html: https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html (read 10 Nov 2026) - docs.aws.amazon.com/eks/latest/userguide/pod-identities.html: https://docs.aws.amazon.com/eks/latest/userguide/pod-identities.html (read 10 Nov 2026) - cloud.google.com/iam/docs/best-practices-service-accounts: https://cloud.google.com/iam/docs/best-practices-service-accounts (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-ident: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/container-apps/managed-identity: https://learn.microsoft.com/en-us/azure/container-apps/managed-identity (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/own-identity-per-agent/ --- # Use short-lived role credentials and no long-lived access keys for the agent Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 2 of 30: Use short-lived role credentials and no long-lived access keys for the agent. A key stored in code, a config file or a container image stays valid until someone rotates it, so a leak gives an attacker lasting access. Temporary credentials from a role expire on their own. ## Why it matters A key stored in code, a config file or a container image stays valid until someone rotates it, so a leak gives an attacker lasting access. Temporary credentials from a role expire on their own. ## How, in each cloud - **AWS:** Have the workload use temporary credentials through an IAM role rather than IAM user access keys. - **Google Cloud:** Use workload identity or an attached service account instead of service account keys, which Google describes as powerful credentials that can present a security risk. - **Azure:** Use a managed identity; Microsoft documents that you do not manage its credentials. ## What the documentation says - "Require workloads to use temporary credentials with IAM roles to access AWS" ([source](https://docs.aws.amazon.com/IAM/latest/UserGuide/best-practices.html)) - "security burden associated with service account keys." ([source](https://cloud.google.com/iam/docs/workload-identity-federation)) - "You don't need to manage credentials." ([source](https://learn.microsoft.com/en-us/entra/identity/managed-identities-azure-resources/overview)) - [Previous: Give each agent its own role or service account, not a shared one](/checklist/own-identity-per-agent/) - [Next: Limit who can pass, attach or impersonate the agent's role](/checklist/restrict-who-can-attach-role/) ## Related pages - [Limit who can pass, attach or impersonate the agent's role: the check, and how in AWS, Google Cloud and Azure](/checklist/restrict-who-can-attach-role/) - [Scope the role's trust to the specific agent, service account or workload: the check, and how in AWS, Google Cloud and Azure](/checklist/scope-trust-to-the-agent/) - [Do not run the agent as a default or pre-granted account: the check, and how in AWS, Google Cloud and Azure](/checklist/no-default-accounts/) ## Sources - docs.aws.amazon.com/IAM/latest/UserGuide/best-practices.html: https://docs.aws.amazon.com/IAM/latest/UserGuide/best-practices.html (read 10 Nov 2026) - cloud.google.com/iam/docs/workload-identity-federation: https://cloud.google.com/iam/docs/workload-identity-federation (read 10 Nov 2026) - learn.microsoft.com/en-us/entra/identity/managed-identities-azure-resources/over: https://learn.microsoft.com/en-us/entra/identity/managed-identities-azure-resources/overview (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/short-lived-credentials/ --- # Limit who can pass, attach or impersonate the agent's role Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 3 of 30: Limit who can pass, attach or impersonate the agent's role. Anyone who can attach the agent's role to a new resource can borrow its access without ever being granted it. Limiting that right keeps the agent's permissions from spreading. ## Why it matters Anyone who can attach the agent's role to a new resource can borrow its access without ever being granted it. Limiting that right keeps the agent's permissions from spreading. ## How, in each cloud - **AWS:** Grant iam:PassRole only for the specific roles a principal needs to pass to a service. - **Google Cloud:** Grant the Service Account User role only to principals that need to attach the service account to a resource. - **Azure:** none found on 2026-11-10 ## What the documentation says - "pass a role to an AWS service, you must grant the PassRole permission to the" ([source](https://docs.aws.amazon.com/IAM/latest/UserGuide/id_roles_use_passrole.html)) - "The Service Account User role ( roles/iam.serviceAccountUser ) lets a principal" ([source](https://cloud.google.com/iam/docs/service-account-permissions)) - [Previous: Use short-lived role credentials and no long-lived access keys for the agent](/checklist/short-lived-credentials/) - [Next: Scope the role's trust to the specific agent, service account or workload](/checklist/scope-trust-to-the-agent/) ## Related pages - [Scope the role's trust to the specific agent, service account or workload: the check, and how in AWS, Google Cloud and Azure](/checklist/scope-trust-to-the-agent/) - [Do not run the agent as a default or pre-granted account: the check, and how in AWS, Google Cloud and Azure](/checklist/no-default-accounts/) - [Write the agent's permission policy from what its tools actually call: the check, and how in AWS, Google Cloud and Azure](/checklist/least-privilege-policy/) ## Sources - docs.aws.amazon.com/IAM/latest/UserGuide/id_roles_use_passrole.html: https://docs.aws.amazon.com/IAM/latest/UserGuide/id_roles_use_passrole.html (read 10 Nov 2026) - cloud.google.com/iam/docs/service-account-permissions: https://cloud.google.com/iam/docs/service-account-permissions (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/restrict-who-can-attach-role/ --- # Scope the role's trust to the specific agent, service account or workload Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 4 of 30: Scope the role's trust to the specific agent, service account or workload. A role that any workload in an account or cluster can assume is a role every compromised workload can use. Tying trust to one agent keeps a break-in elsewhere from reaching it. ## Why it matters A role that any workload in an account or cluster can assume is a role every compromised workload can use. Tying trust to one agent keeps a break-in elsewhere from reaching it. ## How, in each cloud - **AWS:** Use trust policy conditions (for Pod Identity, session tags for service account, namespace and cluster; for services, aws:SourceArn and aws:SourceAccount). - **Google Cloud:** In IAM policies, grant access to the principal that corresponds to one Kubernetes workload, not to the whole cluster or project. - **Azure:** Configure workload identity federation so one managed identity trusts one Kubernetes workload. ## What the documentation says - "You can use these tags in the condition keys in the trust policy to restrict which service accounts, namespaces, and clusters can use this role." ([source](https://docs.aws.amazon.com/eks/latest/userguide/pod-id-role.html)) - "We recommend using aws:SourceArn , aws:SourceAccount , aws:SourceOrgID , or aws:SourceOrgPaths in your resource policies wherever" ([source](https://docs.aws.amazon.com/IAM/latest/UserGuide/confused-deputy.html)) - "In your IAM policy, you refer to a Kubernetes resource by using" ([source](https://cloud.google.com/kubernetes-engine/docs/concepts/workload-identity)) - "Establish a trust relationship between your user-assigned managed identity or app in Microsoft Entra ID and a Kubernetes workload" ([source](https://learn.microsoft.com/en-us/entra/workload-id/workload-identity-federation)) - [Previous: Limit who can pass, attach or impersonate the agent's role](/checklist/restrict-who-can-attach-role/) - [Next: Do not run the agent as a default or pre-granted account](/checklist/no-default-accounts/) ## Related pages - [Do not run the agent as a default or pre-granted account: the check, and how in AWS, Google Cloud and Azure](/checklist/no-default-accounts/) - [Write the agent's permission policy from what its tools actually call: the check, and how in AWS, Google Cloud and Azure](/checklist/least-privilege-policy/) - [Set an organization-level ceiling the agent's role cannot exceed: the check, and how in AWS, Google Cloud and Azure](/checklist/set-permission-ceiling/) ## Sources - docs.aws.amazon.com/eks/latest/userguide/pod-id-role.html: https://docs.aws.amazon.com/eks/latest/userguide/pod-id-role.html (read 10 Nov 2026) - docs.aws.amazon.com/IAM/latest/UserGuide/confused-deputy.html: https://docs.aws.amazon.com/IAM/latest/UserGuide/confused-deputy.html (read 10 Nov 2026) - cloud.google.com/kubernetes-engine/docs/concepts/workload-identity: https://cloud.google.com/kubernetes-engine/docs/concepts/workload-identity (read 10 Nov 2026) - learn.microsoft.com/en-us/entra/workload-id/workload-identity-federation: https://learn.microsoft.com/en-us/entra/workload-id/workload-identity-federation (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/scope-trust-to-the-agent/ --- # Do not run the agent as a default or pre-granted account Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 5 of 30: Do not run the agent as a default or pre-granted account. Default accounts are created for convenience and can carry broad roles that nobody chose for the agent. An agent that inherits one gets access no one reviewed. ## Why it matters Default accounts are created for convenience and can carry broad roles that nobody chose for the agent. An agent that inherits one gets access no one reviewed. ## How, in each cloud - **AWS:** Replace the console-created default role with a role you wrote; AWS documents the Lambda console default as an execution role with minimal permissions (basic logging). - **Google Cloud:** Do not rely on the Compute Engine default service account, which can receive the Editor role; enforce the iam.automaticIamGrantsForDefaultServiceAccounts constraint. - **Azure:** none found on 2026-11-10 ## What the documentation says - "By default, Lambda creates an execution role with minimal permissions when you create a function in the Lambda" ([source](https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html)) - "Editor role ( roles/editor ) on the project by default." ([source](https://cloud.google.com/iam/docs/best-practices-service-accounts)) - "automatically be granted the Editor role on your" ([source](https://cloud.google.com/run/docs/securing/service-identity)) - [Previous: Scope the role's trust to the specific agent, service account or workload](/checklist/scope-trust-to-the-agent/) - [Next: Write the agent's permission policy from what its tools actually call](/checklist/least-privilege-policy/) ## Related pages - [Write the agent's permission policy from what its tools actually call: the check, and how in AWS, Google Cloud and Azure](/checklist/least-privilege-policy/) - [Set an organization-level ceiling the agent's role cannot exceed: the check, and how in AWS, Google Cloud and Azure](/checklist/set-permission-ceiling/) - [Explicitly deny the agent IAM changes and destructive actions: the check, and how in AWS, Google Cloud and Azure](/checklist/deny-iam-changes-and-destructive-actions/) ## Sources - docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html: https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html (read 10 Nov 2026) - cloud.google.com/iam/docs/best-practices-service-accounts: https://cloud.google.com/iam/docs/best-practices-service-accounts (read 10 Nov 2026) - cloud.google.com/run/docs/securing/service-identity: https://cloud.google.com/run/docs/securing/service-identity (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/no-default-accounts/ --- # Write the agent's permission policy from what its tools actually call Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 6 of 30: Write the agent's permission policy from what its tools actually call. Broad starter policies make every tool call possible, including ones nobody planned. A policy limited to the real calls limits what a wrong or hijacked call can do. ## Why it matters Broad starter policies make every tool call possible, including ones nobody planned. A policy limited to the real calls limits what a wrong or hijacked call can do. ## How, in each cloud - **AWS:** Start from the documented minimum policy and narrow it; AWS says to use broad deployment permissions only for development and testing. - **Google Cloud:** Grant the agent identity only the roles its code needs; Google describes agent identity as enabling a least-privilege approach. - **Azure:** Grant the least privilege needed and avoid broader roles at broader scopes. ## What the documentation says - "Use broad deployment permissions only for development and testing." ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html)) - "Agent identity provides a per-agent identity that enables a least-privilege" ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) - "When planning your access control strategy, it's a best practice to grant users the least privilege to get their work done." ([source](https://learn.microsoft.com/en-us/azure/role-based-access-control/best-practices)) - [Previous: Do not run the agent as a default or pre-granted account](/checklist/no-default-accounts/) - [Next: Set an organization-level ceiling the agent's role cannot exceed](/checklist/set-permission-ceiling/) ## Related pages - [Set an organization-level ceiling the agent's role cannot exceed: the check, and how in AWS, Google Cloud and Azure](/checklist/set-permission-ceiling/) - [Explicitly deny the agent IAM changes and destructive actions: the check, and how in AWS, Google Cloud and Azure](/checklist/deny-iam-changes-and-destructive-actions/) - [Compare the agent's granted permissions with what it used and remove the rest: the check, and how in AWS, Google Cloud and Azure](/checklist/remove-unused-permissions/) ## Sources - docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-permissions.html (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-ident: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/role-based-access-control/best-practices: https://learn.microsoft.com/en-us/azure/role-based-access-control/best-practices (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/least-privilege-policy/ --- # Set an organization-level ceiling the agent's role cannot exceed Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 7 of 30: Set an organization-level ceiling the agent's role cannot exceed. A role policy can be edited by whoever manages that account, so it is a weak limit by itself. A ceiling set higher in the hierarchy holds even if the role's own policy is widened. ## Why it matters A role policy can be edited by whoever manages that account, so it is a weak limit by itself. A ceiling set higher in the hierarchy holds even if the role's own policy is widened. ## How, in each cloud - **AWS:** Apply SCPs, which specify the maximum available permissions for IAM roles in the organization, and permissions boundaries on the agent's role. - **Google Cloud:** Use Principal Access Boundary policies to define the resources the agent's principal may access. - **Azure:** Use deny assignments, which block listed actions even when a role assignment grants them (customer-managed deny assignments are documented as private preview), and apply governance conditions at management group scope. ## What the documentation says - "specify the maximum available permissions for the IAM users and IAM roles in your organization." ([source](https://docs.aws.amazon.com/organizations/latest/userguide/orgs_manage_policies_scps.html)) - "policy to set the maximum permissions that an identity-based policy can grant to an IAM" ([source](https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_boundaries.html)) - "Principal Access Boundary (PAB) policies let you define the resources that" ([source](https://cloud.google.com/iam/docs/principal-access-boundary-policies)) - "Customer-managed deny assignments are currently in private preview." ([source](https://learn.microsoft.com/en-us/azure/role-based-access-control/deny-assignments)) - "above subscriptions. When you organize subscriptions into management groups, the governance conditions that you apply" ([source](https://learn.microsoft.com/en-us/azure/governance/management-groups/overview)) - [Previous: Write the agent's permission policy from what its tools actually call](/checklist/least-privilege-policy/) - [Next: Explicitly deny the agent IAM changes and destructive actions](/checklist/deny-iam-changes-and-destructive-actions/) ## Related pages - [Explicitly deny the agent IAM changes and destructive actions: the check, and how in AWS, Google Cloud and Azure](/checklist/deny-iam-changes-and-destructive-actions/) - [Compare the agent's granted permissions with what it used and remove the rest: the check, and how in AWS, Google Cloud and Azure](/checklist/remove-unused-permissions/) - [Grant roles at the narrowest scope that works: the check, and how in AWS, Google Cloud and Azure](/checklist/narrowest-scope/) ## Sources - docs.aws.amazon.com/organizations/latest/userguide/orgs_manage_policies_scps.htm: https://docs.aws.amazon.com/organizations/latest/userguide/orgs_manage_policies_scps.html (read 10 Nov 2026) - docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_boundaries.html: https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_boundaries.html (read 10 Nov 2026) - cloud.google.com/iam/docs/principal-access-boundary-policies: https://cloud.google.com/iam/docs/principal-access-boundary-policies (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/role-based-access-control/deny-assignments: https://learn.microsoft.com/en-us/azure/role-based-access-control/deny-assignments (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/governance/management-groups/overview: https://learn.microsoft.com/en-us/azure/governance/management-groups/overview (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/set-permission-ceiling/ --- # Explicitly deny the agent IAM changes and destructive actions Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 8 of 30: Explicitly deny the agent IAM changes and destructive actions. An agent that can edit roles or delete data can widen its own access or cause damage no review would have approved. An explicit deny wins over any allow added later by mistake. ## Why it matters An agent that can edit roles or delete data can widen its own access or cause damage no review would have approved. An explicit deny wins over any allow added later by mistake. ## How, in each cloud - **AWS:** Use an SCP or permissions boundary whose maximum excludes IAM write actions and destructive calls on protected resources. - **Google Cloud:** Use an IAM deny policy that targets agent identities; Google's example denies all agents in an organization from creating roles and deleting buckets. - **Azure:** Use deny assignments to block specific actions even if a role assignment grants them. ## What the documentation says - "policy to set the maximum permissions that an identity-based policy can grant to an IAM" ([source](https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_boundaries.html)) - "Identity and Access Management (IAM) deny policies let you set guardrails on access to" ([source](https://cloud.google.com/iam/docs/deny-overview)) - "// Deny policy (deny all agents across the org from ability to create or delete buckets)" ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) - "Customer-managed deny assignments are currently in private preview." ([source](https://learn.microsoft.com/en-us/azure/role-based-access-control/deny-assignments)) - [Previous: Set an organization-level ceiling the agent's role cannot exceed](/checklist/set-permission-ceiling/) - [Next: Compare the agent's granted permissions with what it used and remove the rest](/checklist/remove-unused-permissions/) ## Related pages - [Compare the agent's granted permissions with what it used and remove the rest: the check, and how in AWS, Google Cloud and Azure](/checklist/remove-unused-permissions/) - [Grant roles at the narrowest scope that works: the check, and how in AWS, Google Cloud and Azure](/checklist/narrowest-scope/) - [Keep the encryption keys for agent data under your control and limit who can use them: the check, and how in AWS, Google Cloud and Azure](/checklist/control-the-keys/) ## Sources - docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_boundaries.html: https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_boundaries.html (read 10 Nov 2026) - cloud.google.com/iam/docs/deny-overview: https://cloud.google.com/iam/docs/deny-overview (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-ident: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/role-based-access-control/deny-assignments: https://learn.microsoft.com/en-us/azure/role-based-access-control/deny-assignments (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/deny-iam-changes-and-destructive-actions/ --- # Compare the agent's granted permissions with what it used and remove the rest Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 9 of 30: Compare the agent's granted permissions with what it used and remove the rest. Permissions granted during a pilot tend to stay long after the agent stops using them. Each unused permission is access an attacker could use for nothing the agent needs. ## Why it matters Permissions granted during a pilot tend to stay long after the agent stops using them. Each unused permission is access an attacker could use for nothing the agent needs. ## How, in each cloud - **AWS:** Create an IAM Access Analyzer unused access analyzer to find unused access. - **Google Cloud:** Use IAM role recommendations to identify and remove excess permissions. - **Azure:** Run recurring Microsoft Entra access reviews of role assignments. ## What the documentation says - "IAM Access Analyzer unused access analyzers help identify unused access in your" ([source](https://docs.aws.amazon.com/IAM/latest/UserGuide/what-is-access-analyzer.html)) - "Role recommendations help you identify and remove excess permissions from your" ([source](https://cloud.google.com/iam/docs/recommender-overview)) - "enable organizations to efficiently manage group memberships, access to enterprise applications, and role assignments." ([source](https://learn.microsoft.com/en-us/entra/id-governance/access-reviews-overview)) - [Previous: Explicitly deny the agent IAM changes and destructive actions](/checklist/deny-iam-changes-and-destructive-actions/) - [Next: Grant roles at the narrowest scope that works](/checklist/narrowest-scope/) ## Related pages - [Grant roles at the narrowest scope that works: the check, and how in AWS, Google Cloud and Azure](/checklist/narrowest-scope/) - [Keep the encryption keys for agent data under your control and limit who can use them: the check, and how in AWS, Google Cloud and Azure](/checklist/control-the-keys/) - [Block public access on storage the agent can write to: the check, and how in AWS, Google Cloud and Azure](/checklist/block-public-exposure/) ## Sources - docs.aws.amazon.com/IAM/latest/UserGuide/what-is-access-analyzer.html: https://docs.aws.amazon.com/IAM/latest/UserGuide/what-is-access-analyzer.html (read 10 Nov 2026) - cloud.google.com/iam/docs/recommender-overview: https://cloud.google.com/iam/docs/recommender-overview (read 10 Nov 2026) - learn.microsoft.com/en-us/entra/id-governance/access-reviews-overview: https://learn.microsoft.com/en-us/entra/id-governance/access-reviews-overview (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/remove-unused-permissions/ --- # Grant roles at the narrowest scope that works Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 10 of 30: Grant roles at the narrowest scope that works. A role granted on a whole account, project or subscription covers every resource in it, including ones created later. Granting on specific resources keeps new resources out of the agent's reach by default. ## Why it matters A role granted on a whole account, project or subscription covers every resource in it, including ones created later. Granting on specific resources keeps new resources out of the agent's reach by default. ## How, in each cloud - **AWS:** Name specific resource ARNs in the policy; AWS's Bedrock guidance says to replace the wildcard with specific agent IDs once they exist. - **Google Cloud:** Add IAM Conditions so a role binding applies only when conditions are met. - **Azure:** Assign the role at resource or resource group scope instead of subscription or management group scope. ## What the documentation says - "As a best practice for security purposes, replace the * with specific agent IDs after you have created them." ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html)) - "You can use IAM Conditions to define and enforce conditional," ([source](https://cloud.google.com/iam/docs/conditions-overview)) - "Scope is the set of resources that the access applies to." ([source](https://learn.microsoft.com/en-us/azure/role-based-access-control/overview)) - [Previous: Compare the agent's granted permissions with what it used and remove the rest](/checklist/remove-unused-permissions/) - [Next: Keep the encryption keys for agent data under your control and limit who can use them](/checklist/control-the-keys/) ## Related pages - [Keep the encryption keys for agent data under your control and limit who can use them: the check, and how in AWS, Google Cloud and Azure](/checklist/control-the-keys/) - [Block public access on storage the agent can write to: the check, and how in AWS, Google Cloud and Azure](/checklist/block-public-exposure/) - [Put data stores the agent reads inside a perimeter that blocks copying data out: the check, and how in AWS, Google Cloud and Azure](/checklist/data-perimeter/) ## Sources - docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html (read 10 Nov 2026) - cloud.google.com/iam/docs/conditions-overview: https://cloud.google.com/iam/docs/conditions-overview (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/role-based-access-control/overview: https://learn.microsoft.com/en-us/azure/role-based-access-control/overview (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/narrowest-scope/ --- # Keep the encryption keys for agent data under your control and limit who can use them Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 11 of 30: Keep the encryption keys for agent data under your control and limit who can use them. If the agent's role can use a key, it can read whatever that key protects. Customer-managed keys let you cut off access by changing the key policy or disabling the key. ## Why it matters If the agent's role can use a key, it can read whatever that key protects. Customer-managed keys let you cut off access by changing the key policy or disabling the key. ## How, in each cloud - **AWS:** Use a KMS key policy, which AWS calls the primary way to control access to a KMS key. - **Google Cloud:** Use Cloud KMS customer-managed encryption keys (CMEK) and grant key access only to the identities that need it. - **Azure:** Use customer-managed keys for Azure Storage held in an Azure key store. ## What the documentation says - "A key policy is a resource policy for an AWS KMS key. Key policies are the primary way to" ([source](https://docs.aws.amazon.com/kms/latest/developerguide/key-policies.html)) - "customer-managed encryption keys (CMEK). Using Cloud KMS CMEK gives you" ([source](https://cloud.google.com/kms/docs/cmek)) - "You can use your own encryption key to protect the data in your storage account." ([source](https://learn.microsoft.com/en-us/azure/storage/common/customer-managed-keys-overview)) - [Previous: Grant roles at the narrowest scope that works](/checklist/narrowest-scope/) - [Next: Block public access on storage the agent can write to](/checklist/block-public-exposure/) ## Related pages - [Block public access on storage the agent can write to: the check, and how in AWS, Google Cloud and Azure](/checklist/block-public-exposure/) - [Put data stores the agent reads inside a perimeter that blocks copying data out: the check, and how in AWS, Google Cloud and Azure](/checklist/data-perimeter/) - [Limit data access with tags, attributes or conditions, not only resource lists: the check, and how in AWS, Google Cloud and Azure](/checklist/condition-based-data-access/) ## Sources - docs.aws.amazon.com/kms/latest/developerguide/key-policies.html: https://docs.aws.amazon.com/kms/latest/developerguide/key-policies.html (read 10 Nov 2026) - cloud.google.com/kms/docs/cmek: https://cloud.google.com/kms/docs/cmek (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/storage/common/customer-managed-keys-overview: https://learn.microsoft.com/en-us/azure/storage/common/customer-managed-keys-overview (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/control-the-keys/ --- # Block public access on storage the agent can write to Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 12 of 30: Block public access on storage the agent can write to. An agent that can change a bucket setting, or write to a bucket that already allows public reads, can expose data by accident. Account-level public access blocks stop that even when a policy says otherwise. ## Why it matters An agent that can change a bucket setting, or write to a bucket that already allows public reads, can expose data by accident. Account-level public access blocks stop that even when a policy says otherwise. ## How, in each cloud - **AWS:** Turn on S3 Block Public Access settings. - **Google Cloud:** Enable public access prevention on buckets or through the organization policy constraint. - **Azure:** Disable anonymous access on storage accounts; Microsoft recommends this. ## What the documentation says - "The Amazon S3 Block Public Access feature provides settings for access points, buckets, accounts, and" ([source](https://docs.aws.amazon.com/AmazonS3/latest/userguide/access-control-block-public-access.html)) - "Public access prevention protects Cloud Storage buckets and objects" ([source](https://cloud.google.com/storage/docs/public-access-prevention)) - "We recommend that you disable anonymous access for optimal security." ([source](https://learn.microsoft.com/en-us/azure/storage/blobs/anonymous-read-access-prevent)) - [Previous: Keep the encryption keys for agent data under your control and limit who can use them](/checklist/control-the-keys/) - [Next: Put data stores the agent reads inside a perimeter that blocks copying data out](/checklist/data-perimeter/) ## Related pages - [Put data stores the agent reads inside a perimeter that blocks copying data out: the check, and how in AWS, Google Cloud and Azure](/checklist/data-perimeter/) - [Limit data access with tags, attributes or conditions, not only resource lists: the check, and how in AWS, Google Cloud and Azure](/checklist/condition-based-data-access/) - [Run agent compute on private networks with no direct path to the internet: the check, and how in AWS, Google Cloud and Azure](/checklist/private-network-no-direct-internet/) ## Sources - docs.aws.amazon.com/AmazonS3/latest/userguide/access-control-block-public-access: https://docs.aws.amazon.com/AmazonS3/latest/userguide/access-control-block-public-access.html (read 10 Nov 2026) - cloud.google.com/storage/docs/public-access-prevention: https://cloud.google.com/storage/docs/public-access-prevention (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/storage/blobs/anonymous-read-access-prevent: https://learn.microsoft.com/en-us/azure/storage/blobs/anonymous-read-access-prevent (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/block-public-exposure/ --- # Put data stores the agent reads inside a perimeter that blocks copying data out Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 13 of 30: Put data stores the agent reads inside a perimeter that blocks copying data out. An agent allowed to read data can usually also send it somewhere else unless something stops the transfer. A perimeter turns an allowed read into a read that cannot leave. ## Why it matters An agent allowed to read data can usually also send it somewhere else unless something stops the transfer. A perimeter turns an allowed read into a read that cannot leave. ## How, in each cloud - **AWS:** Build a data perimeter with identity, resource and network controls, including VPC endpoint policies. - **Google Cloud:** Use VPC Service Controls, which Google documents as protecting against the risk of data exfiltration. - **Azure:** Use virtual network service endpoint policies to allow data exfiltration only to specific Azure Storage accounts. ## What the documentation says - "Many organizations want to implement perimeter controls to help protect against unintended" ([source](https://docs.aws.amazon.com/whitepapers/latest/building-a-data-perimeter-on-aws/building-a-data-perimeter-on-aws.html)) - "An endpoint policy is a resource-based policy that you attach to a VPC endpoint to control" ([source](https://docs.aws.amazon.com/vpc/latest/privatelink/vpc-endpoints-access.html)) - "VPC Service Controls helps protect against the risk of data exfiltration" ([source](https://cloud.google.com/vpc-service-controls/docs/overview)) - "allow data exfiltration to only specific Azure Storage accounts" ([source](https://learn.microsoft.com/en-us/azure/virtual-network/virtual-network-service-endpoint-policies-overview)) - [Previous: Block public access on storage the agent can write to](/checklist/block-public-exposure/) - [Next: Limit data access with tags, attributes or conditions, not only resource lists](/checklist/condition-based-data-access/) ## Related pages - [Limit data access with tags, attributes or conditions, not only resource lists: the check, and how in AWS, Google Cloud and Azure](/checklist/condition-based-data-access/) - [Run agent compute on private networks with no direct path to the internet: the check, and how in AWS, Google Cloud and Azure](/checklist/private-network-no-direct-internet/) - [Allow outbound traffic only to the destinations the agent needs: the check, and how in AWS, Google Cloud and Azure](/checklist/allowlist-egress/) ## Sources - docs.aws.amazon.com/whitepapers/latest/building-a-data-perimeter-on-aws/building: https://docs.aws.amazon.com/whitepapers/latest/building-a-data-perimeter-on-aws/building-a-data-perimeter-on-aws.html (read 10 Nov 2026) - docs.aws.amazon.com/vpc/latest/privatelink/vpc-endpoints-access.html: https://docs.aws.amazon.com/vpc/latest/privatelink/vpc-endpoints-access.html (read 10 Nov 2026) - cloud.google.com/vpc-service-controls/docs/overview: https://cloud.google.com/vpc-service-controls/docs/overview (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/virtual-network/virtual-network-service-endpoint: https://learn.microsoft.com/en-us/azure/virtual-network/virtual-network-service-endpoint-policies-overview (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/data-perimeter/ --- # Limit data access with tags, attributes or conditions, not only resource lists Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 14 of 30: Limit data access with tags, attributes or conditions, not only resource lists. Resource lists go stale as new data stores appear. Conditions on tags or attributes keep access tied to what the data is, so a new store does not become reachable by default. ## Why it matters Resource lists go stale as new data stores appear. Conditions on tags or attributes keep access tied to what the data is, so a new store does not become reachable by default. ## How, in each cloud - **AWS:** Use attribute-based access control with tags on IAM entities and resources. - **Google Cloud:** Use IAM Conditions on role bindings. - **Azure:** Use Azure role assignment conditions (Azure ABAC) that filter a role's permissions by attributes. ## What the documentation says - "based on attributes. AWS calls these attributes tags . You" ([source](https://docs.aws.amazon.com/IAM/latest/UserGuide/introduction_attribute-based-access-control.html)) - "You can use IAM Conditions to define and enforce conditional," ([source](https://cloud.google.com/iam/docs/conditions-overview)) - "Azure ABAC builds on Azure RBAC by adding role assignment conditions based on attributes in the context of specific actions." ([source](https://learn.microsoft.com/en-us/azure/role-based-access-control/conditions-overview)) - [Previous: Put data stores the agent reads inside a perimeter that blocks copying data out](/checklist/data-perimeter/) - [Next: Run agent compute on private networks with no direct path to the internet](/checklist/private-network-no-direct-internet/) ## Related pages - [Run agent compute on private networks with no direct path to the internet: the check, and how in AWS, Google Cloud and Azure](/checklist/private-network-no-direct-internet/) - [Allow outbound traffic only to the destinations the agent needs: the check, and how in AWS, Google Cloud and Azure](/checklist/allowlist-egress/) - [Reach cloud APIs and data stores through private endpoints: the check, and how in AWS, Google Cloud and Azure](/checklist/private-endpoints-for-cloud-apis/) ## Sources - docs.aws.amazon.com/IAM/latest/UserGuide/introduction_attribute-based-access-con: https://docs.aws.amazon.com/IAM/latest/UserGuide/introduction_attribute-based-access-control.html (read 10 Nov 2026) - cloud.google.com/iam/docs/conditions-overview: https://cloud.google.com/iam/docs/conditions-overview (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/role-based-access-control/conditions-overview: https://learn.microsoft.com/en-us/azure/role-based-access-control/conditions-overview (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/condition-based-data-access/ --- # Run agent compute on private networks with no direct path to the internet Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 15 of 30: Run agent compute on private networks with no direct path to the internet. Compute with a direct internet path can send data out or pull in instructions from anywhere. A private network makes every outside path an explicit choice. ## Why it matters Compute with a direct internet path can send data out or pull in instructions from anywhere. A private network makes every outside path an explicit choice. ## How, in each cloud - **AWS:** Attach functions to private subnets; AWS says a public subnet does not give a Lambda function internet access. - **Google Cloud:** Use a Private Service Connect interface so Agent Runtime traffic stays inside Google's network and never crosses the public internet. - **Azure:** Use the Foundry Standard Setup with private networking, which Microsoft documents as having no public egress. ## What the documentation says - "To access private resources, connect your function to private subnets." ([source](https://docs.aws.amazon.com/lambda/latest/dg/configuration-vpc.html)) - "All traffic between your agent and your VPC travels securely within Google's network, never traversing the public internet." ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/private-service-connect-interface)) - "No public egress : Foundational infrastructure provides the right authentication and security for your agents and tools, without requiring trusted service bypass." ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/virtual-networks)) - [Previous: Limit data access with tags, attributes or conditions, not only resource lists](/checklist/condition-based-data-access/) - [Next: Allow outbound traffic only to the destinations the agent needs](/checklist/allowlist-egress/) ## Related pages - [Allow outbound traffic only to the destinations the agent needs: the check, and how in AWS, Google Cloud and Azure](/checklist/allowlist-egress/) - [Reach cloud APIs and data stores through private endpoints: the check, and how in AWS, Google Cloud and Azure](/checklist/private-endpoints-for-cloud-apis/) - [Keep the agent's code from reading node, instance or other workloads' credentials: the check, and how in AWS, Google Cloud and Azure](/checklist/credential-isolation/) ## Sources - docs.aws.amazon.com/lambda/latest/dg/configuration-vpc.html: https://docs.aws.amazon.com/lambda/latest/dg/configuration-vpc.html (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/private-ser: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/private-service-connect-interface (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/virtual-networks: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/virtual-networks (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/private-network-no-direct-internet/ --- # Allow outbound traffic only to the destinations the agent needs Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 16 of 30: Allow outbound traffic only to the destinations the agent needs. Open outbound access lets a hijacked agent reach any server to send data or fetch tools. A short allow list makes unexpected destinations fail and show up in logs. ## Why it matters Open outbound access lets a hijacked agent reach any server to send data or fetch tools. A short allow list makes unexpected destinations fail and show up in logs. ## How, in each cloud - **AWS:** Use security group rules and AWS Network Firewall to filter traffic at the perimeter of the VPC. - **Google Cloud:** Use VPC firewall rules to allow or deny connections, including egress. - **Azure:** Use network security groups to allow or deny outbound traffic, and Azure Firewall for central filtering. ## What the documentation says - "The security group acts as a virtual firewall." ([source](https://docs.aws.amazon.com/vpc/latest/userguide/vpc-security-groups.html)) - "(Amazon VPC). With Network Firewall, you can filter traffic at the perimeter of your VPC." ([source](https://docs.aws.amazon.com/network-firewall/latest/developerguide/what-is-aws-network-firewall.html)) - "VPC firewall rules let you allow or deny connections to or" ([source](https://cloud.google.com/firewall/docs/firewalls)) - "You can use an Azure network security group to filter network traffic between Azure resources in Azure virtual networks." ([source](https://learn.microsoft.com/en-us/azure/virtual-network/network-security-groups-overview)) - "Azure Firewall is a cloud-native, intelligent network firewall security service" ([source](https://learn.microsoft.com/en-us/azure/firewall/overview)) - [Previous: Run agent compute on private networks with no direct path to the internet](/checklist/private-network-no-direct-internet/) - [Next: Reach cloud APIs and data stores through private endpoints](/checklist/private-endpoints-for-cloud-apis/) ## Related pages - [Reach cloud APIs and data stores through private endpoints: the check, and how in AWS, Google Cloud and Azure](/checklist/private-endpoints-for-cloud-apis/) - [Keep the agent's code from reading node, instance or other workloads' credentials: the check, and how in AWS, Google Cloud and Azure](/checklist/credential-isolation/) - [Give each tool function its own least-privilege role instead of the agent's role: the check, and how in AWS, Google Cloud and Azure](/checklist/tool-own-role/) ## Sources - docs.aws.amazon.com/vpc/latest/userguide/vpc-security-groups.html: https://docs.aws.amazon.com/vpc/latest/userguide/vpc-security-groups.html (read 10 Nov 2026) - docs.aws.amazon.com/network-firewall/latest/developerguide/what-is-aws-network-f: https://docs.aws.amazon.com/network-firewall/latest/developerguide/what-is-aws-network-firewall.html (read 10 Nov 2026) - cloud.google.com/firewall/docs/firewalls: https://cloud.google.com/firewall/docs/firewalls (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/virtual-network/network-security-groups-overview: https://learn.microsoft.com/en-us/azure/virtual-network/network-security-groups-overview (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/firewall/overview: https://learn.microsoft.com/en-us/azure/firewall/overview (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/allowlist-egress/ --- # Reach cloud APIs and data stores through private endpoints Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 17 of 30: Reach cloud APIs and data stores through private endpoints. Calls to cloud services over public addresses can be made from anywhere that holds valid credentials. Private endpoints keep the path inside your network, so stolen credentials are harder to use from outside. ## Why it matters Calls to cloud services over public addresses can be made from anywhere that holds valid credentials. Private endpoints keep the path inside your network, so stolen credentials are harder to use from outside. ## How, in each cloud - **AWS:** Create VPC interface endpoints with AWS PrivateLink. - **Google Cloud:** Use Private Google Access so VMs with only internal IP addresses can reach Google APIs. - **Azure:** Use private endpoints, which put a service on a private IP address from your virtual network. ## What the documentation says - "creating a VPC interface endpoint with AWS PrivateLink to establish a private connection to your data." ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/usingVPC.html)) - "VM instances that only have internal IP addresses (no external IP addresses) can" ([source](https://cloud.google.com/vpc/docs/private-google-access)) - "A private endpoint is a network interface that uses a private IP address from your virtual network." ([source](https://learn.microsoft.com/en-us/azure/private-link/private-endpoint-overview)) - [Previous: Allow outbound traffic only to the destinations the agent needs](/checklist/allowlist-egress/) - [Next: Keep the agent's code from reading node, instance or other workloads' credentials](/checklist/credential-isolation/) ## Related pages - [Keep the agent's code from reading node, instance or other workloads' credentials: the check, and how in AWS, Google Cloud and Azure](/checklist/credential-isolation/) - [Give each tool function its own least-privilege role instead of the agent's role: the check, and how in AWS, Google Cloud and Azure](/checklist/tool-own-role/) - [Restrict which principals can invoke each tool endpoint: the check, and how in AWS, Google Cloud and Azure](/checklist/restrict-tool-invokers/) ## Sources - docs.aws.amazon.com/bedrock/latest/userguide/usingVPC.html: https://docs.aws.amazon.com/bedrock/latest/userguide/usingVPC.html (read 10 Nov 2026) - cloud.google.com/vpc/docs/private-google-access: https://cloud.google.com/vpc/docs/private-google-access (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/private-link/private-endpoint-overview: https://learn.microsoft.com/en-us/azure/private-link/private-endpoint-overview (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/private-endpoints-for-cloud-apis/ --- # Keep the agent's code from reading node, instance or other workloads' credentials Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 18 of 30: Keep the agent's code from reading node, instance or other workloads' credentials. If agent code can read the host's metadata service, it can take the host's credentials, which are often broader than the agent's own. Isolating credentials keeps one compromised agent from becoming a compromised node. ## Why it matters If agent code can read the host's metadata service, it can take the host's credentials, which are often broader than the agent's own. Isolating credentials keeps one compromised agent from becoming a compromised node. ## How, in each cloud - **AWS:** Restrict access to the EC2 Instance Metadata Service for pods so containers can only retrieve credentials for their own role. - **Google Cloud:** Use Workload Identity Federation for GKE, which protects the metadata that metadata concealment covered. - **Azure:** Run hosted agents in per-session VM-isolated sandboxes, which Microsoft documents as holding no blueprint credential. ## What the documentation says - "A container never has access to credentials that are used by other containers in other Pods." ([source](https://docs.aws.amazon.com/eks/latest/userguide/pod-identities.html)) - "The sensitive metadata protected by metadata concealment is also protected by" ([source](https://cloud.google.com/kubernetes-engine/docs/concepts/workload-identity)) - "Each hosted-agent session runs in its own VM-isolated sandbox" ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) - [Previous: Reach cloud APIs and data stores through private endpoints](/checklist/private-endpoints-for-cloud-apis/) - [Next: Give each tool function its own least-privilege role instead of the agent's role](/checklist/tool-own-role/) ## Related pages - [Give each tool function its own least-privilege role instead of the agent's role: the check, and how in AWS, Google Cloud and Azure](/checklist/tool-own-role/) - [Restrict which principals can invoke each tool endpoint: the check, and how in AWS, Google Cloud and Azure](/checklist/restrict-tool-invokers/) - [Cap how much the agent can run and spend: the check, and how in AWS, Google Cloud and Azure](/checklist/cap-run-and-spend/) ## Sources - docs.aws.amazon.com/eks/latest/userguide/pod-identities.html: https://docs.aws.amazon.com/eks/latest/userguide/pod-identities.html (read 10 Nov 2026) - cloud.google.com/kubernetes-engine/docs/concepts/workload-identity: https://cloud.google.com/kubernetes-engine/docs/concepts/workload-identity (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/credential-isolation/ --- # Give each tool function its own least-privilege role instead of the agent's role Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 19 of 30: Give each tool function its own least-privilege role instead of the agent's role. A tool that runs with the agent's full role can do everything the agent can, even if the tool needs one action. A narrow role per tool limits damage when one tool is abused or buggy. ## Why it matters A tool that runs with the agent's full role can do everything the agent can, even if the tool needs one action. A narrow role per tool limits damage when one tool is abused or buggy. ## How, in each cloud - **AWS:** Action group Lambda functions run under their own execution role, and Bedrock reaches them through a resource-based policy. - **Google Cloud:** Create dedicated service accounts for each application, so each tool service can have its own account. - **Azure:** Use user-assigned managed identities per workload and control which identities are available to the container app. ## What the documentation says - "Lambda automatically assumes your execution role when you invoke your function." ([source](https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html)) - "For your agent to use a Lambda function, you must attach a resource-based policy to the function to provide permissions for the agent." ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-lambda.html)) - "Create dedicated service accounts for each application, and avoid using" ([source](https://cloud.google.com/iam/docs/best-practices-service-accounts)) - "you can control which managed identities are available to your container app during the init and main phases" ([source](https://learn.microsoft.com/en-us/azure/container-apps/managed-identity)) - [Previous: Keep the agent's code from reading node, instance or other workloads' credentials](/checklist/credential-isolation/) - [Next: Restrict which principals can invoke each tool endpoint](/checklist/restrict-tool-invokers/) ## Related pages - [Restrict which principals can invoke each tool endpoint: the check, and how in AWS, Google Cloud and Azure](/checklist/restrict-tool-invokers/) - [Cap how much the agent can run and spend: the check, and how in AWS, Google Cloud and Azure](/checklist/cap-run-and-spend/) - [Decide for each tool whether the action runs with the agent's authority or the signed-in user's: the check, and how in AWS, Google Cloud and Azure](/checklist/decide-user-or-agent-authority/) ## Sources - docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html: https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/agents-lambda.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-lambda.html (read 10 Nov 2026) - cloud.google.com/iam/docs/best-practices-service-accounts: https://cloud.google.com/iam/docs/best-practices-service-accounts (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/container-apps/managed-identity: https://learn.microsoft.com/en-us/azure/container-apps/managed-identity (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/tool-own-role/ --- # Restrict which principals can invoke each tool endpoint Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 20 of 30: Restrict which principals can invoke each tool endpoint. A tool endpoint that accepts calls from any caller can be used without the agent. Limiting invokers means only the intended agent can trigger the action. In AWS: Attach a resource-based policy on the tool's Lambda function that allows only the Bedrock service principal for one agent ARN. ## Why it matters A tool endpoint that accepts calls from any caller can be used without the agent. Limiting invokers means only the intended agent can trigger the action. ## How, in each cloud - **AWS:** Attach a resource-based policy on the tool's Lambda function that allows only the Bedrock service principal for one agent ARN. - **Google Cloud:** Grant the Cloud Run Invoker role only to the agent's service account. - **Azure:** Assign Foundry roles at agent scope to grant endpoint access to one agent without granting it for all agents in the project. ## What the documentation says - "Resource-based policy to allow Amazon Bedrock to invoke an action group Lambda function" ([source](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html)) - "Cloud Run Service Invoker ( roles/run.servicesInvoker )" ([source](https://cloud.google.com/run/docs/securing/managing-access)) - "so use this scope to grant access to a specific agent's endpoints without granting endpoint access to all agents in the project" ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/rbac-foundry)) - [Previous: Give each tool function its own least-privilege role instead of the agent's role](/checklist/tool-own-role/) - [Next: Cap how much the agent can run and spend](/checklist/cap-run-and-spend/) ## Related pages - [Cap how much the agent can run and spend: the check, and how in AWS, Google Cloud and Azure](/checklist/cap-run-and-spend/) - [Decide for each tool whether the action runs with the agent's authority or the signed-in user's: the check, and how in AWS, Google Cloud and Azure](/checklist/decide-user-or-agent-authority/) - [Run agents in their own account, project or subscription: the check, and how in AWS, Google Cloud and Azure](/checklist/dedicated-account-project-subscription/) ## Sources - docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html (read 10 Nov 2026) - cloud.google.com/run/docs/securing/managing-access: https://cloud.google.com/run/docs/securing/managing-access (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/concepts/rbac-foundry: https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/rbac-foundry (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/restrict-tool-invokers/ --- # Cap how much the agent can run and spend Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 21 of 30: Cap how much the agent can run and spend. A looping or hijacked agent can start thousands of runs and a large bill before anyone notices. Limits on concurrency and budget alerts shrink the damage window. ## Why it matters A looping or hijacked agent can start thousands of runs and a large bill before anyone notices. Limits on concurrency and budget alerts shrink the damage window. ## How, in each cloud - **AWS:** Set Lambda reserved concurrency and AWS Budgets alerts. - **Google Cloud:** Set a maximum number of Cloud Run instances and Cloud Billing budget alerts. - **Azure:** Set maximum replicas on the container app and Cost Management budgets; Microsoft notes that budget thresholds trigger notifications but do not stop consumption. ## What the documentation says - "This sets both the maximum and minimum number of concurrent instances allocated to your Event Function." ([source](https://docs.aws.amazon.com/lambda/latest/dg/configuration-concurrency.html)) - "and receive alerts when costs approach or exceed your defined threshold." ([source](https://docs.aws.amazon.com/cost-management/latest/userguide/budgets-managing-costs.html)) - "This page describes how to set the maximum number of instances that" ([source](https://cloud.google.com/run/docs/configuring/max-instances)) - "After you've set a budget amount, you set budget alert threshold rules that" ([source](https://cloud.google.com/billing/docs/how-to/budgets)) - "Limits define the minimum and maximum possible number of replicas per revision as your container app scales." ([source](https://learn.microsoft.com/en-us/azure/container-apps/scale-app)) - "Resources aren't affected, and your consumption isn't stopped." ([source](https://learn.microsoft.com/en-us/azure/cost-management-billing/costs/tutorial-acm-create-budgets)) - [Previous: Restrict which principals can invoke each tool endpoint](/checklist/restrict-tool-invokers/) - [Next: Decide for each tool whether the action runs with the agent's authority or the signed-in user's](/checklist/decide-user-or-agent-authority/) ## Related pages - [Decide for each tool whether the action runs with the agent's authority or the signed-in user's: the check, and how in AWS, Google Cloud and Azure](/checklist/decide-user-or-agent-authority/) - [Run agents in their own account, project or subscription: the check, and how in AWS, Google Cloud and Azure](/checklist/dedicated-account-project-subscription/) - [Tag every agent resource with an owner and environment, and enforce the tags: the check, and how in AWS, Google Cloud and Azure](/checklist/tag-agent-resources/) ## Sources - docs.aws.amazon.com/lambda/latest/dg/configuration-concurrency.html: https://docs.aws.amazon.com/lambda/latest/dg/configuration-concurrency.html (read 10 Nov 2026) - docs.aws.amazon.com/cost-management/latest/userguide/budgets-managing-costs.html: https://docs.aws.amazon.com/cost-management/latest/userguide/budgets-managing-costs.html (read 10 Nov 2026) - cloud.google.com/run/docs/configuring/max-instances: https://cloud.google.com/run/docs/configuring/max-instances (read 10 Nov 2026) - cloud.google.com/billing/docs/how-to/budgets: https://cloud.google.com/billing/docs/how-to/budgets (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/container-apps/scale-app: https://learn.microsoft.com/en-us/azure/container-apps/scale-app (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/cost-management-billing/costs/tutorial-acm-creat: https://learn.microsoft.com/en-us/azure/cost-management-billing/costs/tutorial-acm-create-budgets (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/cap-run-and-spend/ --- # Decide for each tool whether the action runs with the agent's authority or the signed-in user's Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 22 of 30: Decide for each tool whether the action runs with the agent's authority or the signed-in user's. If a tool runs with the agent's authority, every user can reach whatever the agent can reach. If it runs with the user's authority, the agent can only do what that user may. ## Why it matters If a tool runs with the agent's authority, every user can reach whatever the agent can reach. If it runs with the user's authority, the agent can only do what that user may. ## How, in each cloud - **AWS:** AgentCore Gateway lets each target use the gateway role, the caller's credentials, or a token exchange that carries the user and agent identities. - **Google Cloud:** Google describes separate user, agent and delegated identities; delegated calls are governed by both the agent's policies and the user's permissions. - **Azure:** Foundry documents application-only access (agent's own authority) and delegated access (on behalf of the signed-in user, in preview). ## What the documentation says - "for a new, scoped access token that targets a downstream resource." ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html)) - "acts on behalf of the initiating human user using scoped 3-legged OAuth" ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/identity-delegation-architecture)) - "Application-only access : The agent acts under its own authority, without a signed-in user." ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) - "Delegated access : The agent acts on behalf of the signed-in user." ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) - [Previous: Cap how much the agent can run and spend](/checklist/cap-run-and-spend/) - [Next: Run agents in their own account, project or subscription](/checklist/dedicated-account-project-subscription/) ## Related pages - [Run agents in their own account, project or subscription: the check, and how in AWS, Google Cloud and Azure](/checklist/dedicated-account-project-subscription/) - [Tag every agent resource with an owner and environment, and enforce the tags: the check, and how in AWS, Google Cloud and Azure](/checklist/tag-agent-resources/) - [Test permission and policy changes in a dry run before enforcing them: the check, and how in AWS, Google Cloud and Azure](/checklist/test-before-enforcing/) ## Sources - docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/govern/identity-delegatio: https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/identity-delegation-architecture (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/decide-user-or-agent-authority/ --- # Run agents in their own account, project or subscription Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 23 of 30: Run agents in their own account, project or subscription. Agents placed beside production workloads share their blast radius, quotas and admin rights. A dedicated container lets you apply stricter policy and limits and remove it cleanly. ## Why it matters Agents placed beside production workloads share their blast radius, quotas and admin rights. A dedicated container lets you apply stricter policy and limits and remove it cleanly. ## How, in each cloud - **AWS:** Use separate AWS accounts to isolate workloads. - **Google Cloud:** Use a separate project, grouped under folders in the resource hierarchy. - **Azure:** Use a separate subscription, grouped under management groups. ## What the documentation says - "Using multiple AWS accounts to help isolate and manage your business" ([source](https://docs.aws.amazon.com/whitepapers/latest/organizing-your-aws-environment/organizing-your-aws-environment.html)) - "projects. Structuring your resources this way helps you centralize asset" ([source](https://cloud.google.com/resource-manager/docs/cloud-platform-resource-hierarchy)) - "Subscriptions are a unit of management, billing, and scale within Azure." ([source](https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ready/landing-zone/design-area/resource-org-subscriptions)) - [Previous: Decide for each tool whether the action runs with the agent's authority or the signed-in user's](/checklist/decide-user-or-agent-authority/) - [Next: Tag every agent resource with an owner and environment, and enforce the tags](/checklist/tag-agent-resources/) ## Related pages - [Tag every agent resource with an owner and environment, and enforce the tags: the check, and how in AWS, Google Cloud and Azure](/checklist/tag-agent-resources/) - [Test permission and policy changes in a dry run before enforcing them: the check, and how in AWS, Google Cloud and Azure](/checklist/test-before-enforcing/) - [Remove the agent's identity and access when the agent is retired: the check, and how in AWS, Google Cloud and Azure](/checklist/retire-identity-with-agent/) ## Sources - docs.aws.amazon.com/whitepapers/latest/organizing-your-aws-environment/organizin: https://docs.aws.amazon.com/whitepapers/latest/organizing-your-aws-environment/organizing-your-aws-environment.html (read 10 Nov 2026) - cloud.google.com/resource-manager/docs/cloud-platform-resource-hierarchy: https://cloud.google.com/resource-manager/docs/cloud-platform-resource-hierarchy (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/cloud-adoption-framework/ready/landing-zone/desi: https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ready/landing-zone/design-area/resource-org-subscriptions (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/dedicated-account-project-subscription/ --- # Tag every agent resource with an owner and environment, and enforce the tags Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 24 of 30: Tag every agent resource with an owner and environment, and enforce the tags. Untagged agent resources cannot be traced to an owner when something goes wrong. Tags also let policy and cost controls target agent resources specifically. ## Why it matters Untagged agent resources cannot be traced to an owner when something goes wrong. Tags also let policy and cost controls target agent resources specifically. ## How, in each cloud - **AWS:** Attach tags to IAM entities and resources; AWS documents tags as the attributes for attribute-based access control. - **Google Cloud:** Attach tags, which are key-value pairs, to resources. - **Azure:** Use Azure Policy to govern tags. ## What the documentation says - "based on attributes. AWS calls these attributes tags . You" ([source](https://docs.aws.amazon.com/IAM/latest/UserGuide/introduction_attribute-based-access-control.html)) - "Tags are key-value pairs that you can attach to your" ([source](https://cloud.google.com/resource-manager/docs/tags/tags-overview)) - "tags can be the basis for applying your business policies with Azure Policy" ([source](https://learn.microsoft.com/en-us/azure/governance/policy/tutorials/govern-tags)) - [Previous: Run agents in their own account, project or subscription](/checklist/dedicated-account-project-subscription/) - [Next: Test permission and policy changes in a dry run before enforcing them](/checklist/test-before-enforcing/) ## Related pages - [Test permission and policy changes in a dry run before enforcing them: the check, and how in AWS, Google Cloud and Azure](/checklist/test-before-enforcing/) - [Remove the agent's identity and access when the agent is retired: the check, and how in AWS, Google Cloud and Azure](/checklist/retire-identity-with-agent/) - [Turn on organization-wide audit logging of control-plane actions: the check, and how in AWS, Google Cloud and Azure](/checklist/control-plane-audit-logs/) ## Sources - docs.aws.amazon.com/IAM/latest/UserGuide/introduction_attribute-based-access-con: https://docs.aws.amazon.com/IAM/latest/UserGuide/introduction_attribute-based-access-control.html (read 10 Nov 2026) - cloud.google.com/resource-manager/docs/tags/tags-overview: https://cloud.google.com/resource-manager/docs/tags/tags-overview (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/governance/policy/tutorials/govern-tags: https://learn.microsoft.com/en-us/azure/governance/policy/tutorials/govern-tags (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/tag-agent-resources/ --- # Test permission and policy changes in a dry run before enforcing them Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 25 of 30: Test permission and policy changes in a dry run before enforcing them. A new deny or boundary can break an agent's tools in ways nobody saw coming, and a mistaken allow can open access without anyone noticing. Testing first shows the effect before it reaches production. ## Why it matters A new deny or boundary can break an agent's tools in ways nobody saw coming, and a mistaken allow can open access without anyone noticing. Testing first shows the effect before it reaches production. ## How, in each cloud - **AWS:** Use the IAM policy simulator to test identity-based policies. - **Google Cloud:** Use organization policies in dry-run mode to monitor how a policy would behave. - **Azure:** Assign Azure Policy with enforcementMode DoNotEnforce. ## What the documentation says - "With the IAM policy simulator, you can test identity-based policies," ([source](https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_testing-policies.html)) - "This page shows how to use an organization policy in dry-run mode to monitor how" ([source](https://cloud.google.com/resource-manager/docs/organization-policy/dry-run-policy)) - "The enforcementMode property provides customers the ability to control the outcome of a policy on existing resources without initiating the policy effect" ([source](https://learn.microsoft.com/en-us/azure/governance/policy/concepts/assignment-structure)) - [Previous: Tag every agent resource with an owner and environment, and enforce the tags](/checklist/tag-agent-resources/) - [Next: Remove the agent's identity and access when the agent is retired](/checklist/retire-identity-with-agent/) ## Related pages - [Remove the agent's identity and access when the agent is retired: the check, and how in AWS, Google Cloud and Azure](/checklist/retire-identity-with-agent/) - [Turn on organization-wide audit logging of control-plane actions: the check, and how in AWS, Google Cloud and Azure](/checklist/control-plane-audit-logs/) - [Log reads and writes to the data the agent touches: the check, and how in AWS, Google Cloud and Azure](/checklist/log-data-access/) ## Sources - docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_testing-policies.html: https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_testing-policies.html (read 10 Nov 2026) - cloud.google.com/resource-manager/docs/organization-policy/dry-run-policy: https://cloud.google.com/resource-manager/docs/organization-policy/dry-run-policy (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/governance/policy/concepts/assignment-structure: https://learn.microsoft.com/en-us/azure/governance/policy/concepts/assignment-structure (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/test-before-enforcing/ --- # Remove the agent's identity and access when the agent is retired Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 26 of 30: Remove the agent's identity and access when the agent is retired. Leftover identities from retired agents keep their permissions and nobody is watching them. Tying the identity to the agent's lifecycle removes the access with the agent. ## Why it matters Leftover identities from retired agents keep their permissions and nobody is watching them. Tying the identity to the agent's lifecycle removes the access with the agent. ## How, in each cloud - **AWS:** none found on 2026-11-10 - **Google Cloud:** Use agent identity, which Google documents as tied to the lifecycle of the agent. - **Azure:** Use a system-assigned managed identity, which is deleted when the container app is deleted. ## What the documentation says - "is tied to the lifecycle of the agent, making agent identity a more" ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) - "A system-assigned identity is tied to your container app and is deleted when your container app is deleted." ([source](https://learn.microsoft.com/en-us/azure/container-apps/managed-identity)) - [Previous: Test permission and policy changes in a dry run before enforcing them](/checklist/test-before-enforcing/) - [Next: Turn on organization-wide audit logging of control-plane actions](/checklist/control-plane-audit-logs/) ## Related pages - [Turn on organization-wide audit logging of control-plane actions: the check, and how in AWS, Google Cloud and Azure](/checklist/control-plane-audit-logs/) - [Log reads and writes to the data the agent touches: the check, and how in AWS, Google Cloud and Azure](/checklist/log-data-access/) - [Make every logged action traceable to the specific agent identity: the check, and how in AWS, Google Cloud and Azure](/checklist/attribute-actions-to-the-agent/) ## Sources - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-ident: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/container-apps/managed-identity: https://learn.microsoft.com/en-us/azure/container-apps/managed-identity (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/retire-identity-with-agent/ --- # Turn on organization-wide audit logging of control-plane actions Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 27 of 30: Turn on organization-wide audit logging of control-plane actions. Without audit logs there is no record of which role or setting the agent changed, or who changed the agent's access. A log kept outside the agent's reach is the base for every investigation. ## Why it matters Without audit logs there is no record of which role or setting the agent changed, or who changed the agent's access. A log kept outside the agent's reach is the base for every investigation. ## How, in each cloud - **AWS:** Create an organization trail in CloudTrail. - **Google Cloud:** Rely on Admin Activity audit logs, which cannot be disabled. - **Azure:** Use the Azure Monitor activity log, which records management operations for Azure resources. ## What the documentation says - "and apply it to an organization, making it an organization trail. Organization trails log" ([source](https://docs.aws.amazon.com/awscloudtrail/latest/userguide/creating-trail-organization.html)) - "You can't disable Admin Activity audit logs." ([source](https://cloud.google.com/kubernetes-engine/docs/how-to/audit-logging)) - "Azure Monitor records management operations for your Azure resources through the activity log feature." ([source](https://learn.microsoft.com/en-us/azure/azure-monitor/essentials/activity-log)) - [Previous: Remove the agent's identity and access when the agent is retired](/checklist/retire-identity-with-agent/) - [Next: Log reads and writes to the data the agent touches](/checklist/log-data-access/) ## Related pages - [Log reads and writes to the data the agent touches: the check, and how in AWS, Google Cloud and Azure](/checklist/log-data-access/) - [Make every logged action traceable to the specific agent identity: the check, and how in AWS, Google Cloud and Azure](/checklist/attribute-actions-to-the-agent/) - [Alert on denied calls and on changes to the agent's role or policies: the check, and how in AWS, Google Cloud and Azure](/checklist/alert-on-denied-and-role-changes/) ## Sources - docs.aws.amazon.com/awscloudtrail/latest/userguide/creating-trail-organization.h: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/creating-trail-organization.html (read 10 Nov 2026) - cloud.google.com/kubernetes-engine/docs/how-to/audit-logging: https://cloud.google.com/kubernetes-engine/docs/how-to/audit-logging (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/azure-monitor/essentials/activity-log: https://learn.microsoft.com/en-us/azure/azure-monitor/essentials/activity-log (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/control-plane-audit-logs/ --- # Log reads and writes to the data the agent touches Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 28 of 30: Log reads and writes to the data the agent touches. Control-plane logs show configuration changes, not which records the agent read. Data-level logs are what show whether the agent touched something it should not have. In AWS: Configure CloudTrail data events, which trails do not log by default. ## Why it matters Control-plane logs show configuration changes, not which records the agent read. Data-level logs are what show whether the agent touched something it should not have. ## How, in each cloud - **AWS:** Configure CloudTrail data events, which trails do not log by default. - **Google Cloud:** Enable Data Access audit logs, which are disabled by default for most services. - **Azure:** Create a diagnostic setting to collect resource logs, which are not collected until you do. ## What the documentation says - "By default, trails and event data stores do not log data events." ([source](https://docs.aws.amazon.com/awscloudtrail/latest/userguide/logging-data-events-with-cloudtrail.html)) - "Data Access audit logs are disabled by default for all services except some" ([source](https://cloud.google.com/logging/docs/audit/configure-data-access)) - "Resource logs aren't collected and stored until you create a diagnostic setting" ([source](https://learn.microsoft.com/en-us/azure/storage/blobs/monitor-blob-storage)) - [Previous: Turn on organization-wide audit logging of control-plane actions](/checklist/control-plane-audit-logs/) - [Next: Make every logged action traceable to the specific agent identity](/checklist/attribute-actions-to-the-agent/) ## Related pages - [Make every logged action traceable to the specific agent identity: the check, and how in AWS, Google Cloud and Azure](/checklist/attribute-actions-to-the-agent/) - [Alert on denied calls and on changes to the agent's role or policies: the check, and how in AWS, Google Cloud and Azure](/checklist/alert-on-denied-and-role-changes/) - [Give each agent its own role or service account, not a shared one: the check, and how in AWS, Google Cloud and Azure](/checklist/own-identity-per-agent/) ## Sources - docs.aws.amazon.com/awscloudtrail/latest/userguide/logging-data-events-with-clou: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/logging-data-events-with-cloudtrail.html (read 10 Nov 2026) - cloud.google.com/logging/docs/audit/configure-data-access: https://cloud.google.com/logging/docs/audit/configure-data-access (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/storage/blobs/monitor-blob-storage: https://learn.microsoft.com/en-us/azure/storage/blobs/monitor-blob-storage (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/log-data-access/ --- # Make every logged action traceable to the specific agent identity Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 29 of 30: Make every logged action traceable to the specific agent identity. If logs only show a shared role or a generic account, you cannot tell which agent did what. Per-agent attribution turns a log line into a decision about one agent. ## Why it matters If logs only show a shared role or a generic account, you cannot tell which agent did what. Per-agent attribution turns a log line into a decision about one agent. ## How, in each cloud - **AWS:** CloudTrail records the type of IAM identity that made each request in the userIdentity element. - **Google Cloud:** Agent identity appears in logs across Google Cloud services, and for user-delegated flows logs show both the user and the agent. - **Azure:** Entra sign-in logs show which agent identity obtained a token for which resource, including on behalf of which user. ## What the documentation says - "element contains details about the type of IAM identity that made the request, and which" ([source](https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-event-reference-user-identity.html)) - "show both the user identity and the agent identity." ([source](https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity)) - "These logs show which agent identity obtained a token for which resource" ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity)) - [Previous: Log reads and writes to the data the agent touches](/checklist/log-data-access/) - [Next: Alert on denied calls and on changes to the agent's role or policies](/checklist/alert-on-denied-and-role-changes/) ## Related pages - [Alert on denied calls and on changes to the agent's role or policies: the check, and how in AWS, Google Cloud and Azure](/checklist/alert-on-denied-and-role-changes/) - [Give each agent its own role or service account, not a shared one: the check, and how in AWS, Google Cloud and Azure](/checklist/own-identity-per-agent/) - [Use short-lived role credentials and no long-lived access keys for the agent: the check, and how in AWS, Google Cloud and Azure](/checklist/short-lived-credentials/) ## Sources - docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-event-reference-us: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/cloudtrail-event-reference-user-identity.html (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-ident: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/attribute-actions-to-the-agent/ --- # Alert on denied calls and on changes to the agent's role or policies Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. Check 30 of 30: Alert on denied calls and on changes to the agent's role or policies. A string of denied calls or an edit to the agent's role is often the first visible sign of misuse. Without an alert, these show up only after the damage. ## Why it matters A string of denied calls or an edit to the agent's role is often the first visible sign of misuse. Without an alert, these show up only after the damage. ## How, in each cloud - **AWS:** Send CloudTrail events to CloudWatch Logs and alarm on matching events. - **Google Cloud:** Create log-based alerting policies that notify you when a specific message appears in logs. - **Azure:** Create Azure Monitor alert rules on log data. ## What the documentation says - "Complete the following steps to configure CloudTrail with CloudWatch Logs to monitor your trail logs and" ([source](https://docs.aws.amazon.com/awscloudtrail/latest/userguide/monitor-cloudtrail-log-files-with-cloudwatch-logs.html)) - "You can configure an alerting policy to notify you whenever a specific message" ([source](https://cloud.google.com/logging/docs/alerting/log-based-alerts)) - "You can alert on any metric or log data source in the Azure Monitor data platform." ([source](https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/alerts-overview)) - [Previous: Make every logged action traceable to the specific agent identity](/checklist/attribute-actions-to-the-agent/) ## Related pages - [Give each agent its own role or service account, not a shared one: the check, and how in AWS, Google Cloud and Azure](/checklist/own-identity-per-agent/) - [Use short-lived role credentials and no long-lived access keys for the agent: the check, and how in AWS, Google Cloud and Azure](/checklist/short-lived-credentials/) - [Limit who can pass, attach or impersonate the agent's role: the check, and how in AWS, Google Cloud and Azure](/checklist/restrict-who-can-attach-role/) ## Sources - docs.aws.amazon.com/awscloudtrail/latest/userguide/monitor-cloudtrail-log-files-: https://docs.aws.amazon.com/awscloudtrail/latest/userguide/monitor-cloudtrail-log-files-with-cloudwatch-logs.html (read 10 Nov 2026) - cloud.google.com/logging/docs/alerting/log-based-alerts: https://cloud.google.com/logging/docs/alerting/log-based-alerts (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/azure-monitor/alerts/alerts-overview: https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/alerts-overview (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/checklist/alert-on-denied-and-role-changes/ --- # What an AI agent can access in AWS, and the controls that bound it Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. In AWS, an AI agent runs as an IAM role, and can reach whatever that identity's permissions and the network allow. The map covers 3 ways to run one: Amazon Bedrock Agents; Amazon Bedrock AgentCore; Self-hosted agent on AWS Lambda or Amazon EKS. In AWS the agent is an IAM role. Amazon Bedrock Agents Classic runs as a service role, Amazon Bedrock AgentCore Runtime as an execution role, and a Lambda or EKS agent as its function or pod role. The checklist's AWS answers put the limits on that role: an SCP or permissions boundary as the ceiling (ab-07), an explicit deny of IAM changes and destructive actions (ab-08), and trust policy conditions that name the agent (ab-04). Around the role sit the network and the tools. The checklist uses private subnets, security groups and AWS Network Firewall for egress (ab-15, ab-16), VPC interface endpoints for AWS APIs (ab-17), and a separate role and resource-based policy for each tool function (ab-19, ab-20). Logging is an organization trail plus data events, which trails do not log by default (ab-27, ab-28). | Agent platform | Runs as | Network | | --- | --- | --- | | [Amazon Bedrock Agents (now Amazon Bedrock Agents Classic)](/agent-platforms/bedrock-agents/) | AWS documents that the agent runs as an IAM service role that Amazon Bedrock assumes. | AWS documents interface endpoints (AWS PrivateLink) for the bedrock-agent and bedrock-agent-runtime APIs, with an endpoint policy you can attach. | | [Amazon Bedrock AgentCore (Runtime, Gateway, Identity)](/agent-platforms/bedrock-agentcore/) | Runtime runs the agent under an IAM execution role that AgentCore assumes. | AWS documents VPC connectivity for Runtime and the built-in Code Interpreter and Browser tools: AgentCore creates network interfaces in your VPC through a service-linked role. | | [Self-hosted agent on AWS Lambda or Amazon EKS](/agent-platforms/self-hosted-aws/) | A Lambda function runs as its execution role. | AWS documents attaching a Lambda function to a VPC: private subnets for private resources, and a public subnet does not give a function internet access. | ## The checklist, AWS column - [Give each agent its own role or service account, not a shared one](/checklist/own-identity-per-agent/): Use one Lambda execution role per function, or one EKS Pod Identity association per service account, so each agent maps to its own IAM role. - [Use short-lived role credentials and no long-lived access keys for the agent](/checklist/short-lived-credentials/): Have the workload use temporary credentials through an IAM role rather than IAM user access keys. - [Limit who can pass, attach or impersonate the agent's role](/checklist/restrict-who-can-attach-role/): Grant iam:PassRole only for the specific roles a principal needs to pass to a service. - [Scope the role's trust to the specific agent, service account or workload](/checklist/scope-trust-to-the-agent/): Use trust policy conditions (for Pod Identity, session tags for service account, namespace and cluster; for services, aws:SourceArn and aws:SourceAccount). - [Do not run the agent as a default or pre-granted account](/checklist/no-default-accounts/): Replace the console-created default role with a role you wrote; AWS documents the Lambda console default as an execution role with minimal permissions (basic logging). - [Write the agent's permission policy from what its tools actually call](/checklist/least-privilege-policy/): Start from the documented minimum policy and narrow it; AWS says to use broad deployment permissions only for development and testing. - [Set an organization-level ceiling the agent's role cannot exceed](/checklist/set-permission-ceiling/): Apply SCPs, which specify the maximum available permissions for IAM roles in the organization, and permissions boundaries on the agent's role. - [Explicitly deny the agent IAM changes and destructive actions](/checklist/deny-iam-changes-and-destructive-actions/): Use an SCP or permissions boundary whose maximum excludes IAM write actions and destructive calls on protected resources. - [Compare the agent's granted permissions with what it used and remove the rest](/checklist/remove-unused-permissions/): Create an IAM Access Analyzer unused access analyzer to find unused access. - [Grant roles at the narrowest scope that works](/checklist/narrowest-scope/): Name specific resource ARNs in the policy; AWS's Bedrock guidance says to replace the wildcard with specific agent IDs once they exist. - [Keep the encryption keys for agent data under your control and limit who can use them](/checklist/control-the-keys/): Use a KMS key policy, which AWS calls the primary way to control access to a KMS key. - [Block public access on storage the agent can write to](/checklist/block-public-exposure/): Turn on S3 Block Public Access settings. - [Put data stores the agent reads inside a perimeter that blocks copying data out](/checklist/data-perimeter/): Build a data perimeter with identity, resource and network controls, including VPC endpoint policies. - [Limit data access with tags, attributes or conditions, not only resource lists](/checklist/condition-based-data-access/): Use attribute-based access control with tags on IAM entities and resources. - [Run agent compute on private networks with no direct path to the internet](/checklist/private-network-no-direct-internet/): Attach functions to private subnets; AWS says a public subnet does not give a Lambda function internet access. - [Allow outbound traffic only to the destinations the agent needs](/checklist/allowlist-egress/): Use security group rules and AWS Network Firewall to filter traffic at the perimeter of the VPC. - [Reach cloud APIs and data stores through private endpoints](/checklist/private-endpoints-for-cloud-apis/): Create VPC interface endpoints with AWS PrivateLink. - [Keep the agent's code from reading node, instance or other workloads' credentials](/checklist/credential-isolation/): Restrict access to the EC2 Instance Metadata Service for pods so containers can only retrieve credentials for their own role. - [Give each tool function its own least-privilege role instead of the agent's role](/checklist/tool-own-role/): Action group Lambda functions run under their own execution role, and Bedrock reaches them through a resource-based policy. - [Restrict which principals can invoke each tool endpoint](/checklist/restrict-tool-invokers/): Attach a resource-based policy on the tool's Lambda function that allows only the Bedrock service principal for one agent ARN. - [Cap how much the agent can run and spend](/checklist/cap-run-and-spend/): Set Lambda reserved concurrency and AWS Budgets alerts. - [Decide for each tool whether the action runs with the agent's authority or the signed-in user's](/checklist/decide-user-or-agent-authority/): AgentCore Gateway lets each target use the gateway role, the caller's credentials, or a token exchange that carries the user and agent identities. - [Run agents in their own account, project or subscription](/checklist/dedicated-account-project-subscription/): Use separate AWS accounts to isolate workloads. - [Tag every agent resource with an owner and environment, and enforce the tags](/checklist/tag-agent-resources/): Attach tags to IAM entities and resources; AWS documents tags as the attributes for attribute-based access control. - [Test permission and policy changes in a dry run before enforcing them](/checklist/test-before-enforcing/): Use the IAM policy simulator to test identity-based policies. - [Remove the agent's identity and access when the agent is retired](/checklist/retire-identity-with-agent/): none found on 2026-11-10 - [Turn on organization-wide audit logging of control-plane actions](/checklist/control-plane-audit-logs/): Create an organization trail in CloudTrail. - [Log reads and writes to the data the agent touches](/checklist/log-data-access/): Configure CloudTrail data events, which trails do not log by default. - [Make every logged action traceable to the specific agent identity](/checklist/attribute-actions-to-the-agent/): CloudTrail records the type of IAM identity that made each request in the userIdentity element. - [Alert on denied calls and on changes to the agent's role or policies](/checklist/alert-on-denied-and-role-changes/): Send CloudTrail events to CloudWatch Logs and alarm on matching events. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [What an AI agent can access in Google Cloud](/clouds/gcp/) - [What an AI agent can access in Azure](/clouds/azure/) - [About](/about/) ## Sources - docs.aws.amazon.com/bedrock/latest/userguide/agents.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/agents-classic-maintenance-mode.htm: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-classic-maintenance-mode.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-permissions.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/agents-lambda.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-lambda.html (read 10 Nov 2026) - docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html: https://docs.aws.amazon.com/lambda/latest/dg/lambda-intro-execution-role.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/agents-how.html: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-how.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/vpc-interface-endpoints.html: https://docs.aws.amazon.com/bedrock/latest/userguide/vpc-interface-endpoints.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/usingVPC.html: https://docs.aws.amazon.com/bedrock/latest/userguide/usingVPC.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/logging-using-cloudtrail.html: https://docs.aws.amazon.com/bedrock/latest/userguide/logging-using-cloudtrail.html (read 10 Nov 2026) - docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html: https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/clouds/aws/ --- # What an AI agent can access in Google Cloud, and the controls that bound it Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. In Google Cloud, an AI agent runs as a service account or a per-agent identity, and can reach whatever that identity's permissions and the network allow. The map covers 3 ways to run one: Agent Runtime on Gemini Enterprise Agent Platform; Gemini Enterprise; Self-hosted agent on Google Cloud Run or GKE. In Google Cloud the agent is an agent identity on Agent Runtime, or a service account on Cloud Run or GKE. Agent Runtime lets you choose between a per-agent identity and a shared service account, and the checklist asks for the per-agent one (ab-01). The default Compute Engine service account might hold the Editor role, so ab-05 enforces the constraint that stops that grant. The ceiling is a Principal Access Boundary policy (ab-07) and an IAM deny policy that targets agent identities (ab-08). VPC Service Controls and Private Service Connect bound data and egress (ab-13, ab-15). Cloud Audit Logs record Admin Activity always, while Data Access logs are off by default for most services and must be enabled (ab-27, ab-28). | Agent platform | Runs as | Network | | --- | --- | --- | | [Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine, Agent Builder)](/agent-platforms/agent-runtime-gcp/) | Google documents three choices: a per-agent IAM agent identity (recommended, a principal tied to the individual agent); the Google-managed AI Platform Reasoning Engine Service Agent used by default; or a custom service account. | Google documents that the agent runs in a Google-managed network without access to your VPC unless you configure a Private Service Connect interface, which gives private egress to your VPC and on-premises resources. | | [Gemini Enterprise (agents hosted on Agent Runtime)](/agent-platforms/gemini-enterprise/) | Gemini Enterprise registers an ADK agent that runs on Agent Runtime, so the agent runs under the Agent Runtime identity described in that entry. | Google documents that the connection between Agent Runtime and Gemini Enterprise is compliant with VPC Service Controls. | | [Self-hosted agent on Google Cloud Run or GKE](/agent-platforms/self-hosted-gcp/) | Cloud Run runs the service as a service account (the Compute Engine default service account if you do not choose one). | Cloud Run can send egress traffic to a VPC network with Direct VPC egress (recommended) or a VPC connector, giving access to resources with internal IP addresses. | ## The checklist, Google Cloud column - [Give each agent its own role or service account, not a shared one](/checklist/own-identity-per-agent/): Create dedicated service accounts for each application, or use a per-agent agent identity on Agent Runtime. - [Use short-lived role credentials and no long-lived access keys for the agent](/checklist/short-lived-credentials/): Use workload identity or an attached service account instead of service account keys, which Google describes as powerful credentials that can present a security risk. - [Limit who can pass, attach or impersonate the agent's role](/checklist/restrict-who-can-attach-role/): Grant the Service Account User role only to principals that need to attach the service account to a resource. - [Scope the role's trust to the specific agent, service account or workload](/checklist/scope-trust-to-the-agent/): In IAM policies, grant access to the principal that corresponds to one Kubernetes workload, not to the whole cluster or project. - [Do not run the agent as a default or pre-granted account](/checklist/no-default-accounts/): Do not rely on the Compute Engine default service account, which can receive the Editor role; enforce the iam.automaticIamGrantsForDefaultServiceAccounts constraint. - [Write the agent's permission policy from what its tools actually call](/checklist/least-privilege-policy/): Grant the agent identity only the roles its code needs; Google describes agent identity as enabling a least-privilege approach. - [Set an organization-level ceiling the agent's role cannot exceed](/checklist/set-permission-ceiling/): Use Principal Access Boundary policies to define the resources the agent's principal may access. - [Explicitly deny the agent IAM changes and destructive actions](/checklist/deny-iam-changes-and-destructive-actions/): Use an IAM deny policy that targets agent identities; Google's example denies all agents in an organization from creating roles and deleting buckets. - [Compare the agent's granted permissions with what it used and remove the rest](/checklist/remove-unused-permissions/): Use IAM role recommendations to identify and remove excess permissions. - [Grant roles at the narrowest scope that works](/checklist/narrowest-scope/): Add IAM Conditions so a role binding applies only when conditions are met. - [Keep the encryption keys for agent data under your control and limit who can use them](/checklist/control-the-keys/): Use Cloud KMS customer-managed encryption keys (CMEK) and grant key access only to the identities that need it. - [Block public access on storage the agent can write to](/checklist/block-public-exposure/): Enable public access prevention on buckets or through the organization policy constraint. - [Put data stores the agent reads inside a perimeter that blocks copying data out](/checklist/data-perimeter/): Use VPC Service Controls, which Google documents as protecting against the risk of data exfiltration. - [Limit data access with tags, attributes or conditions, not only resource lists](/checklist/condition-based-data-access/): Use IAM Conditions on role bindings. - [Run agent compute on private networks with no direct path to the internet](/checklist/private-network-no-direct-internet/): Use a Private Service Connect interface so Agent Runtime traffic stays inside Google's network and never crosses the public internet. - [Allow outbound traffic only to the destinations the agent needs](/checklist/allowlist-egress/): Use VPC firewall rules to allow or deny connections, including egress. - [Reach cloud APIs and data stores through private endpoints](/checklist/private-endpoints-for-cloud-apis/): Use Private Google Access so VMs with only internal IP addresses can reach Google APIs. - [Keep the agent's code from reading node, instance or other workloads' credentials](/checklist/credential-isolation/): Use Workload Identity Federation for GKE, which protects the metadata that metadata concealment covered. - [Give each tool function its own least-privilege role instead of the agent's role](/checklist/tool-own-role/): Create dedicated service accounts for each application, so each tool service can have its own account. - [Restrict which principals can invoke each tool endpoint](/checklist/restrict-tool-invokers/): Grant the Cloud Run Invoker role only to the agent's service account. - [Cap how much the agent can run and spend](/checklist/cap-run-and-spend/): Set a maximum number of Cloud Run instances and Cloud Billing budget alerts. - [Decide for each tool whether the action runs with the agent's authority or the signed-in user's](/checklist/decide-user-or-agent-authority/): Google describes separate user, agent and delegated identities; delegated calls are governed by both the agent's policies and the user's permissions. - [Run agents in their own account, project or subscription](/checklist/dedicated-account-project-subscription/): Use a separate project, grouped under folders in the resource hierarchy. - [Tag every agent resource with an owner and environment, and enforce the tags](/checklist/tag-agent-resources/): Attach tags, which are key-value pairs, to resources. - [Test permission and policy changes in a dry run before enforcing them](/checklist/test-before-enforcing/): Use organization policies in dry-run mode to monitor how a policy would behave. - [Remove the agent's identity and access when the agent is retired](/checklist/retire-identity-with-agent/): Use agent identity, which Google documents as tied to the lifecycle of the agent. - [Turn on organization-wide audit logging of control-plane actions](/checklist/control-plane-audit-logs/): Rely on Admin Activity audit logs, which cannot be disabled. - [Log reads and writes to the data the agent touches](/checklist/log-data-access/): Enable Data Access audit logs, which are disabled by default for most services. - [Make every logged action traceable to the specific agent identity](/checklist/attribute-actions-to-the-agent/): Agent identity appears in logs across Google Cloud services, and for user-delegated flows logs show both the user and the agent. - [Alert on denied calls and on changes to the agent's role or policies](/checklist/alert-on-denied-and-role-changes/): Create log-based alerting policies that notify you when a specific message appears in logs. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [What an AI agent can access in Azure](/clouds/azure/) - [About](/about/) - [Editorial policy](/editorial-policy/) ## Sources - docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime/setup: https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime/setup (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/deploy-an-a: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/deploy-an-agent (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-ident: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/agent-identity (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/govern/identity-delegatio: https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/identity-delegation-architecture (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/private-ser: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/private-service-connect-interface (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/general/: https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/general/vpc-service-controls (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/logging: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/logging (read 10 Nov 2026) - docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/manage-agen: https://docs.cloud.google.com/gemini-enterprise-agent-platform/scale/runtime/manage-agent-access (read 10 Nov 2026) - cloud.google.com/gemini/enterprise/docs/audit-logging: https://cloud.google.com/gemini/enterprise/docs/audit-logging (read 10 Nov 2026) - cloud.google.com/gemini/enterprise/docs/register-and-manage-an-adk-agent: https://cloud.google.com/gemini/enterprise/docs/register-and-manage-an-adk-agent (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/clouds/gcp/ --- # What an AI agent can access in Azure, and the controls that bound it Facts read 10 Nov 2026. Page reviewed 16 Nov 2026. In Azure, an AI agent runs as a managed identity or an Entra agent identity, and can reach whatever that identity's permissions and the network allow. The map covers 2 ways to run one: Microsoft Foundry Agent Service; Self-hosted agent on Azure Container Apps or AKS. In Azure the agent is a Microsoft Entra agent identity in Foundry Agent Service, or a managed identity on Container Apps and AKS. Foundry gives a new agent identity no resource permissions, so access is whatever you assign with Azure RBAC. The checklist asks for assignment at resource or resource group scope rather than subscription scope (ab-10) and deny assignments as the ceiling (ab-07, ab-08), noting that customer-managed deny assignments are in private preview. Network and logs need deliberate setup. Foundry's Standard Setup with private networking removes public egress (ab-15); Container Apps and AKS are open by default. Entra sign-in logs show which agent identity got a token and for which user (ab-29), and resource logs are not collected until you create a diagnostic setting (ab-28). Restricting who can attach a managed identity: none found on 2026-11-10 (ab-03). | Agent platform | Runs as | Network | | --- | --- | --- | | [Microsoft Foundry Agent Service (formerly Azure AI Foundry Agent Service)](/agent-platforms/foundry-agent-service/) | Each agent gets a Microsoft Entra agent identity, a special service principal created under an agent blueprint. | Microsoft documents a Standard Setup with private networking: no public egress and a delegated subnet from your virtual network. | | [Self-hosted agent on Azure Container Apps or AKS](/agent-platforms/self-hosted-azure/) | A container app can have a system-assigned managed identity (tied to the app) or user-assigned identities (standalone resources). | By default Container Apps is publicly accessible and can communicate only with internet-accessible endpoints; supplying your own virtual network adds control over outbound traffic and access to private endpoints. | ## The checklist, Azure column - [Give each agent its own role or service account, not a shared one](/checklist/own-identity-per-agent/): Use the per-agent Entra agent identity that Foundry creates, or a system-assigned managed identity that is tied to one container app. - [Use short-lived role credentials and no long-lived access keys for the agent](/checklist/short-lived-credentials/): Use a managed identity; Microsoft documents that you do not manage its credentials. - [Limit who can pass, attach or impersonate the agent's role](/checklist/restrict-who-can-attach-role/): none found on 2026-11-10 - [Scope the role's trust to the specific agent, service account or workload](/checklist/scope-trust-to-the-agent/): Configure workload identity federation so one managed identity trusts one Kubernetes workload. - [Do not run the agent as a default or pre-granted account](/checklist/no-default-accounts/): none found on 2026-11-10 - [Write the agent's permission policy from what its tools actually call](/checklist/least-privilege-policy/): Grant the least privilege needed and avoid broader roles at broader scopes. - [Set an organization-level ceiling the agent's role cannot exceed](/checklist/set-permission-ceiling/): Use deny assignments, which block listed actions even when a role assignment grants them (customer-managed deny assignments are documented as private preview), and apply governance conditions at management group scope. - [Explicitly deny the agent IAM changes and destructive actions](/checklist/deny-iam-changes-and-destructive-actions/): Use deny assignments to block specific actions even if a role assignment grants them. - [Compare the agent's granted permissions with what it used and remove the rest](/checklist/remove-unused-permissions/): Run recurring Microsoft Entra access reviews of role assignments. - [Grant roles at the narrowest scope that works](/checklist/narrowest-scope/): Assign the role at resource or resource group scope instead of subscription or management group scope. - [Keep the encryption keys for agent data under your control and limit who can use them](/checklist/control-the-keys/): Use customer-managed keys for Azure Storage held in an Azure key store. - [Block public access on storage the agent can write to](/checklist/block-public-exposure/): Disable anonymous access on storage accounts; Microsoft recommends this. - [Put data stores the agent reads inside a perimeter that blocks copying data out](/checklist/data-perimeter/): Use virtual network service endpoint policies to allow data exfiltration only to specific Azure Storage accounts. - [Limit data access with tags, attributes or conditions, not only resource lists](/checklist/condition-based-data-access/): Use Azure role assignment conditions (Azure ABAC) that filter a role's permissions by attributes. - [Run agent compute on private networks with no direct path to the internet](/checklist/private-network-no-direct-internet/): Use the Foundry Standard Setup with private networking, which Microsoft documents as having no public egress. - [Allow outbound traffic only to the destinations the agent needs](/checklist/allowlist-egress/): Use network security groups to allow or deny outbound traffic, and Azure Firewall for central filtering. - [Reach cloud APIs and data stores through private endpoints](/checklist/private-endpoints-for-cloud-apis/): Use private endpoints, which put a service on a private IP address from your virtual network. - [Keep the agent's code from reading node, instance or other workloads' credentials](/checklist/credential-isolation/): Run hosted agents in per-session VM-isolated sandboxes, which Microsoft documents as holding no blueprint credential. - [Give each tool function its own least-privilege role instead of the agent's role](/checklist/tool-own-role/): Use user-assigned managed identities per workload and control which identities are available to the container app. - [Restrict which principals can invoke each tool endpoint](/checklist/restrict-tool-invokers/): Assign Foundry roles at agent scope to grant endpoint access to one agent without granting it for all agents in the project. - [Cap how much the agent can run and spend](/checklist/cap-run-and-spend/): Set maximum replicas on the container app and Cost Management budgets; Microsoft notes that budget thresholds trigger notifications but do not stop consumption. - [Decide for each tool whether the action runs with the agent's authority or the signed-in user's](/checklist/decide-user-or-agent-authority/): Foundry documents application-only access (agent's own authority) and delegated access (on behalf of the signed-in user, in preview). - [Run agents in their own account, project or subscription](/checklist/dedicated-account-project-subscription/): Use a separate subscription, grouped under management groups. - [Tag every agent resource with an owner and environment, and enforce the tags](/checklist/tag-agent-resources/): Use Azure Policy to govern tags. - [Test permission and policy changes in a dry run before enforcing them](/checklist/test-before-enforcing/): Assign Azure Policy with enforcementMode DoNotEnforce. - [Remove the agent's identity and access when the agent is retired](/checklist/retire-identity-with-agent/): Use a system-assigned managed identity, which is deleted when the container app is deleted. - [Turn on organization-wide audit logging of control-plane actions](/checklist/control-plane-audit-logs/): Use the Azure Monitor activity log, which records management operations for Azure resources. - [Log reads and writes to the data the agent touches](/checklist/log-data-access/): Create a diagnostic setting to collect resource logs, which are not collected until you do. - [Make every logged action traceable to the specific agent identity](/checklist/attribute-actions-to-the-agent/): Entra sign-in logs show which agent identity obtained a token for which resource, including on behalf of which user. - [Alert on denied calls and on changes to the agent's role or policies](/checklist/alert-on-denied-and-role-changes/): Create Azure Monitor alert rules on log data. **Where Act is behind.** Act draws the boundary at the access layer (cloud IAM and network controls, and the gateways or MCP servers in front of agents); it does not inspect prompts or model outputs, which needs a different tool. Your team applies the boundaries it proposes, and its customer evidence for AI boundaries is less mature than for its core access control. ## Related pages - [About](/about/) - [Editorial policy](/editorial-policy/) - [Corrections](/corrections/) ## Sources - learn.microsoft.com/en-us/azure/ai-foundry/agents/overview: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/overview (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/concepts/rbac-foundry: https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/rbac-foundry (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/agent-identity (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/hosted-agents: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/concepts/hosted-agents (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/tools/openapi: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/tools/openapi (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/tools/function-calling: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/tools/function-calling (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/virtual-networks: https://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/virtual-networks (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/ai-foundry/how-to/azure-policy: https://learn.microsoft.com/en-us/azure/ai-foundry/how-to/azure-policy (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/container-apps/managed-identity: https://learn.microsoft.com/en-us/azure/container-apps/managed-identity (read 10 Nov 2026) - learn.microsoft.com/en-us/azure/aks/workload-identity-overview: https://learn.microsoft.com/en-us/azure/aks/workload-identity-overview (read 10 Nov 2026) Last reviewed 16 Nov 2026. Canonical: https://aiagentaccesscontrol.com/clouds/azure/