---
title: How do you restrict which Amazon Bedrock models agents can use?
url: https://opencontroller.lyzr.com/blog/restrict-amazon-bedrock-models-with-scps
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you restrict which Amazon Bedrock models agents can use?

**Use a service control policy (SCP) that denies Bedrock invocation for every foundation model except an approved list, applied at the organisation or OU level. Since AWS retired the Bedrock Model Access page in October 2025, serverless models are available by default, so an allow-list is now the only reliable gate. Include inference profile ARNs and every destination Region, or approved models will fail too.**

This guide covers the policy pattern, the two gotchas that break it, and how to keep the list current as agents and models change.

## What changed in Bedrock model access?

Before October 2025, an administrator had to enable each model on the Model Access page. That page acted as a soft allow-list. AWS removed it. According to the AWS Security Blog post on simplified model access:

- Serverless models are available automatically in each Region.
- The `PutFoundationModelEntitlement` permission no longer has any effect. Policies that relied on it stopped protecting you.
- Models already blocked by IAM or SCPs stay blocked.
- Anthropic models still need a one-time usage form, which can be submitted from the management account for child accounts.

If your governance relied on "nobody enabled that model", that control is gone. Any role with `bedrock:InvokeModel*` on `*` can now call new models the day they launch.

## What does an allow-list SCP look like?

AWS recommends a "block everything except" pattern using `NotResource`. A minimal version for invocation looks like this:

```json
{
  "Version": "2012-10-17",
  "Statement": [{
    "Sid": "DenyUnapprovedBedrockModels",
    "Effect": "Deny",
    "Action": [
      "bedrock:InvokeModel",
      "bedrock:InvokeModelWithResponseStream"
    ],
    "NotResource": [
      "arn:aws:bedrock:*::foundation-model/amazon.nova*",
      "arn:aws:bedrock:*::foundation-model/anthropic.claude-sonnet-4*",
      "arn:aws:bedrock:*:*:inference-profile/*",
      "arn:aws:bedrock:*:*:application-inference-profile/*"
    ]
  }]
}
```

Denying `InvokeModel` also blocks the Converse API for those models, because Converse is authorised through the same actions. That is what you want: one rule covers both APIs.

Test in a sandbox OU first. An SCP mistake blocks every role in every account below it.

## Which gotchas break the allow-list?

| Gotcha | What happens | Fix |
|---|---|---|
| Inference profiles | Cross-Region profiles send requests to foundation models in several Regions. If any destination model ARN is denied, calls fail | Allow the profile ARN and the foundation-model ARN in every Region the profile routes to |
| Global routing | Global profiles route to any supported Region, including ones you may not approve for data residency | AWS documents a Deny on `aws:RequestedRegion` equal to `unspecified` to block Global routing |
| Marketplace models | Some serverless models are sold through AWS Marketplace and subscribe automatically on first call | The caller needs `aws-marketplace:Subscribe` and `ViewSubscriptions`, or an admin enables the model once per account |
| Wildcards | `anthropic.claude*` approves every future Claude model, including ones not yet reviewed | Pin to model families or versions you have actually evaluated |

To force an agent to call a model only through a specific profile, AWS documents the `bedrock:InferenceProfileArn` condition key on the foundation-model statement.

## How do you keep the approved list current?

An SCP is only as good as the review behind it. Use this checklist.

1. **Own the list.** Name one team that approves models, with a written review: data handling, Region, cost, evaluation results.
2. **Store it as code.** Keep the SCP in version control and deploy it through your pipeline. Every change gets a pull request and a reviewer.
3. **Evaluate before approving.** Run the agents that will use the model against your test sets. See [how to test a cheaper model before switching an agent](/blog/test-a-cheaper-model-before-switching-an-agent).
4. **Watch for denied calls.** CloudTrail records `AccessDenied` on Bedrock calls. A spike after a policy change usually means a missing Region or profile ARN.
5. **Review quarterly.** Remove models nobody calls. Check [Bedrock model call auditing](/blog/audit-amazon-bedrock-model-calls) to see which models are in use.

## Is an SCP enough on its own?

No. An SCP decides which models a principal may call. It does not tell you which agent made a call, how much each agent spends, or whether outputs are grounded. It also does nothing for agents calling models outside AWS with their own keys. Pair the SCP with discovery and per-agent attribution.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, and its Posture feature flags risky permissions and configuration, such as roles with unrestricted Bedrock access. It has a runtime adapter for Bedrock AgentCore, and its LLM Gateway can route agents only to approved models with cost attributed per agent and per version. See the [control plane overview](/blog/what-is-an-ai-agent-control-plane) and [governing agents across platforms](/blog/how-to-govern-ai-agents-across-platforms), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is the Bedrock Model Access page still the way to block models?**
No. AWS retired it in October 2025. Use IAM policies and SCPs.

**Why does my SCP allow-list break cross-Region inference?**
Inference profiles route to models in several Regions. Allow the profile ARN and the foundation-model ARN in every destination Region.

**Should I use an IAM policy or an SCP?**
Both. The SCP sets the organisation-wide ceiling that account admins cannot override. IAM policies then scope each agent's role to the models it needs.

## Key facts

In October 2025 AWS retired the Amazon Bedrock Model Access page: serverless models are available automatically in each Region and the PutFoundationModelEntitlement permission no longer has any effect. AWS recommends restricting models with IAM policies and service control policies (SCPs), using an allow-list that denies Bedrock actions with NotResource for approved foundation-model ARNs. Inference profiles need both the profile ARN and the foundation-model ARN in every destination Region; the bedrock:InferenceProfileArn condition key forces calls through a profile. Some models are sold through AWS Marketplace and subscribe automatically on first invocation, which needs aws-marketplace:Subscribe. Opencontroller auto-discovers agents, models and API keys across AWS and its Posture feature detects risky permissions and configuration. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How do you audit Amazon Bedrock model calls made by AI agents?](https://opencontroller.lyzr.com/blog/audit-amazon-bedrock-model-calls.md)
- [How do you govern AI agents built on Amazon Bedrock AgentCore?](https://opencontroller.lyzr.com/blog/govern-amazon-bedrock-agentcore-agents.md)
- [How do you track Claude spend across Amazon Bedrock, Claude Platform on AWS and the Anthropic API?](https://opencontroller.lyzr.com/blog/track-claude-spend-bedrock-vs-anthropic-api.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
