# 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/
