Skip to content
AI Agent Access Control

30 questions, answered from the data

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.

See the ranking ยท Read the method

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.

Last reviewed 16 Nov 2026.