---
title: How do you restrict which models teams can deploy in Microsoft Foundry?
url: https://opencontroller.lyzr.com/blog/restrict-models-in-microsoft-foundry-with-azure-policy
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you restrict which models teams can deploy in Microsoft Foundry?

**Assign the built-in Azure Policy "Foundry model deployments should only use approved models" with the effect set to Deny, listing approved publishers or model asset IDs. Add the companion "eligibility requirements" policy to block preview models. Assign both at subscription or management-group scope so project owners cannot bypass them, and control who can deploy with RBAC.**

Microsoft Foundry lets any user with deploy rights pick from a catalog of thousands of models. Without a policy, the approved-model list lives in a wiki. This guide shows the two built-in policies, their parameters, and the details that trip teams up.

## Which built-in policies control Foundry deployments?

Microsoft documents two generally available built-in definitions.

| Policy | Parameters | What it blocks |
|---|---|---|
| Foundry model deployments should only use approved models (formerly "Cognitive Services Deployments should only use approved Registry Models") | `effect`, `allowedPublishers`, `allowedAssetIds` | Any deployment whose model is not from an allowed publisher or asset ID |
| Foundry model deployments should meet eligibility requirements | `effect`, `onlyAllowDirectFromAzure`, `denyPreviewModels` | Models that are not "Direct from Azure" and models in Preview, when those flags are true |

The rename did not change the policy definition ID, so existing assignments keep working.

An example parameter file for the approval policy:

```json
{
  "effect": { "value": "Deny" },
  "allowedPublishers": { "value": ["OpenAI"] },
  "allowedAssetIds": {
    "value": ["azureml://registries/azure-openai/models/gpt-5/"]
  }
}
```

An Audit effect only reports. It will not stop a deployment.

## What details break the policy?

Five behaviours from Microsoft's documentation matter in practice.

1. **Asset IDs match as prefixes.** `.../models/gpt-5` without a trailing slash also matches GPT-5.2 and GPT-5.4. Add a trailing slash to approve one model, or the full versioned ID to approve one version.
2. **Values are case-sensitive.** `openai` is not `OpenAI`. Copy publisher names and IDs exactly.
3. **Model router needs every model approved.** A model router deployment passes only if the router and every model it can route to are allowed. With publisher approval, include `Microsoft` and each routed publisher. Through ARM or CLI, one non-compliant model fails the whole deployment; the portal can deploy a compliant subset.
4. **It is not instant.** Wait at least 15 minutes after assignment. Compliance reports can take up to 24 hours to show non-compliant resources.
5. **It works at deployment time.** The catalog still shows every model; the Deploy button is disabled for denied ones. Check compliance results for deployments that existed before the policy.

## Who should be allowed to deploy models?

Policy limits what can be deployed. RBAC limits who can deploy. Use both.

- **Assign policies** with Owner or Resource Policy Contributor at the target scope. Keep that group small.
- **Separate deployers from users.** Most agent builders need to call an existing deployment, not create new ones. Give them a data-plane role, not deployment rights.
- **Use Entra ID, not keys.** Disable local key auth so every call carries an identity. See [replacing Azure OpenAI keys with Entra ID](/blog/replace-azure-openai-keys-with-entra-id).
- **Manage policy as code.** Store assignments and parameter files in version control so each change to the approved list has a reviewer.

## What does the policy not cover?

An approved-model policy answers "can this model be deployed here?". It does not tell you which agents use a deployment, what each agent costs, whether outputs are grounded, or whether an agent built elsewhere is calling an unapproved model through its own API key. It also covers only Azure. Agents on AWS, Google Cloud or SaaS platforms need their own controls. See [governing AI agents in a Microsoft-first enterprise](/blog/govern-ai-agents-in-a-microsoft-first-enterprise) for the wider picture.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across Azure, AWS and Google Cloud, and has a runtime adapter for Microsoft Foundry. Posture detects risky permissions and configuration, and the LLM Gateway can route agents only to approved models, with cost attributed per agent and per version. Policies are defined as code with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. Read 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

**Which Azure Policy restricts models in Microsoft Foundry?**
The built-in "Foundry model deployments should only use approved models", with allowed publishers or asset IDs and the effect set to Deny.

**Does the policy hide models from the catalog?**
No. Models stay visible, but deployment is blocked for denied models.

**How do I approve one model without approving its newer versions?**
Use a trailing slash on the asset ID, or the full versioned ID, because IDs match as prefixes.

## Key facts

Microsoft Foundry has two generally available built-in Azure Policy definitions for model deployments: "Foundry model deployments should only use approved models" (formerly "Cognitive Services Deployments should only use approved Registry Models"; parameters effect, allowedPublishers, allowedAssetIds) and "Foundry model deployments should meet eligibility requirements" (parameters onlyAllowDirectFromAzure, denyPreviewModels). Asset IDs match as prefixes and are case-sensitive; a trailing slash limits a match to one model. Assignments take at least 15 minutes to apply and are evaluated at deployment time. Assigning them requires Owner or Resource Policy Contributor. Opencontroller auto-discovers agents, models and API keys across Azure, has a runtime adapter for Microsoft Foundry, and Posture detects risky configuration. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How do you govern agents in Microsoft Foundry?](https://opencontroller.lyzr.com/blog/govern-microsoft-foundry-agents.md)
- [How do you replace Azure OpenAI API keys with Entra ID for AI agents?](https://opencontroller.lyzr.com/blog/replace-azure-openai-keys-with-entra-id.md)
- [How do you govern AI agents in a Microsoft-first enterprise?](https://opencontroller.lyzr.com/blog/govern-ai-agents-in-a-microsoft-first-enterprise.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
