---
title: How do you restrict which models teams can use on Vertex AI?
url: https://opencontroller.lyzr.com/blog/restrict-vertex-ai-models-with-org-policy
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you restrict which models teams can use on Vertex AI?

**Set the `vertexai.allowedModels` organisation policy at the organisation or folder level, listing each approved model with the action allowed: `predict`, `deploy` or `tune`. With no policy set, every Model Garden model is allowed. Add VPC Service Controls around AI projects so approved models cannot be reached from outside your perimeter.**

Vertex AI was rebranded as Gemini Enterprise Agent Platform in April 2026, but the constraint and API names did not change. This guide covers the policy format, its limits, and how to run it without blocking every new project.

## How does the vertexai.allowedModels constraint work?

The constraint is a list policy. Each value names one model and one action:

```
publishers/PUBLISHER/models/MODEL_NAME:ACTION
```

Examples from Google's documentation:

- `publishers/google/models/gemini-2.0-flash-001:predict` lets principals call that Gemini model.
- `publishers/meta/models/llama3:deploy` lets principals deploy Llama 3 to an endpoint.
- `publishers/hf-google/models/gemma-2b:deploy` allows deploying a Hugging Face model.

The three actions map to three different risks:

| Action | What it allows | Main risk |
|---|---|---|
| `predict` | Calling a managed model through the API | Data sent to a model you have not reviewed |
| `deploy` | Deploying a model to your own endpoint | Running unreviewed weights on your infrastructure, plus compute cost |
| `tune` | Fine-tuning the model on your data | Training data leaving its approved use |

Most agent teams only need `predict`. Grant `deploy` and `tune` to a small platform team.

## What are the limits you need to plan for?

Google documents five behaviours that shape how you write the policy.

1. **No wildcards.** You cannot allow `publishers/anthropic/models/*` or group by publisher. Every model version is a separate value.
2. **500 values maximum.** An organisation policy cannot exceed 500 allowed and denied values combined. Keep the list to models you actually use.
3. **Implicit deny once you set an allow-list.** Anything not listed is blocked, including Gemini models your teams already use. List them before you enforce.
4. **Deny beats allow, and children can override.** When policies merge, an explicit deny wins. A project-level policy can override its parent, so restrict who holds the Organization Policy Administrator role.
5. **Model Garden only.** The constraint does not cover models in the Vertex AI Model Registry, such as your own custom models.

It also applies to every principal in the resource. You cannot allow a model for one team and block it for another inside the same project. Use separate projects or folders for teams with different approvals.

## How do you roll it out without breaking agents?

Follow this sequence.

1. **Inventory current use.** Enable Data Access audit logs and check which models each project calls. See [how to audit Vertex AI calls made by agents](/blog/audit-vertex-ai-calls-made-by-agents).
2. **Build the list from real use.** Add every model in production use, then remove the ones nobody can justify.
3. **Test before enforcing.** Apply the policy to a test project first, or use Organization Policy dry-run or Policy Simulator where they support this constraint, and review what would be blocked.
4. **Enforce at folder level.** Start with one folder of non-critical projects, then move to the organisation.
5. **Add VPC Service Controls.** Put AI projects inside a service perimeter that includes `aiplatform.googleapis.com`. This stops data reaching the API from outside the perimeter and limits exfiltration if credentials leak.
6. **Review on model releases.** New model versions are blocked until listed. Set a review path so teams are not tempted to move to unmanaged API keys.

## What does the policy not control?

The constraint decides which models can be used. It does not show which agent called a model, what it cost, or whether the answer was grounded. It does not affect agents that call Google models through Google AI Studio API keys, or models on other clouds. Pair it with discovery of API keys and agents across your estate, and with input and output screening such as [Model Armor](/blog/google-model-armor-for-ai-agents).

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across Google Cloud, AWS and Azure, and has a runtime adapter for Vertex AI. Its LLM Gateway can route each agent only to approved models, with fallback, caching and cost attributed per agent and per version. Posture flags risky permissions and configuration. 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

**What is the organisation policy for Vertex AI model access?**
The `vertexai.allowedModels` constraint, with values in the format `publishers/PUBLISHER/models/MODEL_NAME:ACTION`.

**Can I allow all models from one publisher?**
No. Wildcards and publisher grouping are not supported. List each model, up to 500 values.

**Does the policy cover my own fine-tuned or custom models?**
No. It applies to Model Garden models only, not to models in the Model Registry. Control those with IAM on the endpoints.

## Key facts

On Google Cloud, the organisation policy constraint vertexai.allowedModels controls which Model Garden models can be used in Vertex AI (rebranded Gemini Enterprise Agent Platform in April 2026). Values use the format publishers/PUBLISHER/models/MODEL_NAME:ACTION, where ACTION is predict, deploy or tune. Wildcards and publisher-level grouping are not supported, a policy is limited to 500 allowed and denied values, it applies to all principals in the resource, explicit deny wins over allow, and it covers only Model Garden models, not Model Registry. With no policy, all models are allowed. VPC Service Controls add a perimeter against data exfiltration. Opencontroller auto-discovers agents, models and API keys across Google Cloud and has a runtime adapter for Vertex AI. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How do you govern agents deployed on Vertex AI Agent Engine?](https://opencontroller.lyzr.com/blog/govern-google-vertex-ai-agent-engine-agents.md)
- [What does Google Cloud Model Armor do for AI agents?](https://opencontroller.lyzr.com/blog/google-model-armor-for-ai-agents.md)
- [How do you govern agents in Gemini Enterprise?](https://opencontroller.lyzr.com/blog/govern-gemini-enterprise-agents.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
