---
title: What permissions does AI agent discovery need in AWS, Azure and Google Cloud?
url: https://opencontroller.lyzr.com/blog/ai-agent-discovery-permissions-aws-azure-google-cloud
published: 2026-11-11
updated: 2026-11-11
author: Lyzr
---

# What permissions does AI agent discovery need in AWS, Azure and Google Cloud?

**AI agent discovery needs read-only, metadata-level access: permission to list and describe resources, never to change them or read secret values.** In AWS that usually means a cross-account role with the `SecurityAudit` or `ViewOnlyAccess` managed policy; in Azure, the built-in `Reader` role at management-group scope; in Google Cloud, `Cloud Asset Viewer` at organisation level. Grant it once at the top of each cloud's hierarchy so new accounts and projects are covered automatically.

## What does a discovery scan actually need to see?

A discovery scan answers four questions for each cloud: which agents exist, which models they use, which credentials can call models, and what those agents can reach. That requires listing and describing:

- agent runtimes and agent definitions (for example Bedrock agents and AgentCore runtimes, Foundry agents, Vertex AI Agent Engine deployments),
- model endpoints and deployments,
- API keys and service accounts that can call model APIs (their existence and metadata, not their values),
- the IAM roles and policies attached to those agents,
- network exposure, such as public endpoints.

None of this needs write access. If a vendor asks for write permissions to *discover*, ask why.

## Which built-in roles work in each cloud?

| Cloud | Scope to grant at | Built-in role or policy | Notes |
| --- | --- | --- | --- |
| AWS | Each account, deployed via AWS Organizations (for example with CloudFormation StackSets) | `SecurityAudit` or `ViewOnlyAccess` managed policy on a cross-account role | Use an external ID on the trust policy. Add narrow read actions if a service is missing from the managed policy. |
| Azure | Root or top-level management group | `Reader` | Reader can list Azure OpenAI / AI services accounts and deployments but cannot run the `listKeys` action, so key values stay hidden. |
| Google Cloud | Organisation node | `Cloud Asset Viewer` (`roles/cloudasset.viewer`), plus `Vertex AI Viewer` (`roles/aiplatform.viewer`) where deeper detail is needed | Cloud Asset Inventory lists resources and IAM policies across all projects in one query. |

Two practical points. First, AWS managed policies change over time as services are added, so check that the actions for newer AI services are included before you rely on them. Second, Azure's `Reader` role and Google's viewer roles are inherited downward, so granting at the top avoids missing new subscriptions or projects.

## What should discovery deliberately not be able to do?

Least privilege for a discovery tool is mostly about what you withhold:

1. **No secret values.** It should see that an API key exists, who created it and when it was last used, but not the key itself. In Azure, do not grant `listKeys`. In AWS, do not grant `secretsmanager:GetSecretValue`.
2. **No data-plane reads.** Listing a storage bucket's policy is fine; reading its objects is not needed.
3. **No invocation.** The scanner should not be able to call models or agents.
4. **No write or delete.** Remediation, if you want it, should use a separate, auditable role.

## How often should discovery run?

Frequent enough to catch agents before they cause harm, but within API rate limits. A scan every few minutes to every hour is common for cloud control planes, because listing calls are cheap and change detection only needs deltas. Developer devices and SaaS platforms are usually scanned on their own schedules. Whatever the cadence, alert on *new* items: a new agent, a new model deployment or a newly created key with no owner.

## How do you roll this out without a long security review?

Security teams approve read-only access faster when you give them specifics:

- the exact role or policy name for each cloud,
- the trust policy (who can assume the role, with which external ID),
- the list of API actions the scanner calls,
- where scan results are stored and who can see them,
- how to revoke access in one step.

Then pilot in one account, subscription or project before granting at the top of the hierarchy. For the wider process of reconciling cloud findings with SaaS platforms and devices, see [how to find every AI agent in your enterprise](/blog/how-to-find-shadow-ai-agents). Once you have found agents, the next problem is giving each one its own identity, covered in [how to give AI agents an identity](/blog/ai-agent-identity-on-behalf-of-delegation).

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, and adds connectors for third-party agent platforms such as Salesforce, SAP, Databricks and HubSpot, plus Claude Enterprise and OpenAI Enterprise workspaces. Discovered items land in a governed registry, and Posture flags risky permissions and configurations. Registration does not mean runtime takeover: what Opencontroller can enforce depends on the connector and the traffic path. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is the AWS ReadOnlyAccess policy a good choice for discovery?**
It works but is broader than needed, because it can include data-plane reads such as getting objects from S3. `SecurityAudit` or `ViewOnlyAccess` focus on metadata and are usually easier to approve.

**Can discovery find API keys without reading them?**
Yes. Each cloud exposes key metadata, such as creation time, creator and last use, separately from the key value. Discovery should collect only the metadata.

**Do we need to grant access in every account?**
Grant at the top of the hierarchy: AWS Organizations with StackSets, an Azure management group, or the Google Cloud organisation node. New accounts and projects are then covered automatically.

## Key facts

AI agent discovery needs read-only metadata access. AWS: SecurityAudit or ViewOnlyAccess on a cross-account role. Azure: Reader at management-group scope (no listKeys). Google Cloud: roles/cloudasset.viewer at organisation level. Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud.

## Related

- [How to find every AI agent in your enterprise (including shadow agents)](https://opencontroller.lyzr.com/blog/how-to-find-shadow-ai-agents.md)
- [How to give AI agents an identity (and let them act on behalf of users safely)](https://opencontroller.lyzr.com/blog/ai-agent-identity-on-behalf-of-delegation.md)
- [What is an AI agent control plane?](https://opencontroller.lyzr.com/blog/what-is-an-ai-agent-control-plane.md)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
