---
title: Does Azure OpenAI store prompts sent by your agents for abuse monitoring?
url: https://opencontroller.lyzr.com/blog/azure-openai-abuse-monitoring-and-agent-data
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# Does Azure OpenAI store prompts sent by your agents for abuse monitoring?

**It can. By default, Azure's abuse monitoring classifies every prompt and completion, and content that is flagged can be stored and reviewed by authorised Microsoft employees. Customers who meet Microsoft's Limited Access criteria can apply for modified abuse monitoring, which turns that storage and human review off — but it does not touch data your agents store on purpose through Files, Assistants, the Responses API or batch.**

Agents send far more traffic than chat users, often including customer records and internal documents pulled in by tools. So the question "where do our prompts go?" usually comes up when an agent moves from pilot to production. Here is how the pieces fit, based on Microsoft's documentation for models sold by Azure in Microsoft Foundry.

## How does Azure abuse monitoring work?

Microsoft describes three stages:

1. **Classification.** Prompts and completions are scored for harmful content and patterns of abuse.
2. **Automated review.** Flagged content can be checked by AI models to confirm the system's analysis. Microsoft states that prompts and completions in this review are not stored by the abuse-monitoring system.
3. **Human review.** In some cases — for example, when automated review is not confident — authorised Microsoft employees can review flagged content. Access goes through Secure Access Workstations with just-in-time approval from team managers.

Where data is stored for review, it sits in a data store in the Azure region where your Foundry resource is deployed, logically separated by resource ID. For models deployed in the European Economic Area, the human reviewers are located in the EEA.

## What does modified abuse monitoring change?

| | Default | Modified abuse monitoring |
|---|---|---|
| Classification of prompts and completions | Yes | Yes |
| Automated review of flagged content | Yes | Yes |
| Storage of flagged content for human review | Yes | No |
| Human review by Microsoft employees | Possible | No |
| How to get it | On by default | Apply; Limited Access criteria |

Microsoft notes that some advanced models may have stricter criteria for turning abuse monitoring off. It also warns that detection can be less accurate without human review, and that you should be ready to respond to abuse notifications to avoid service interruption.

## How do you verify it is actually off?

Approval is applied to the subscription, and you can check it per resource. In the Azure portal, open the resource overview and switch to the JSON view (or query it with the Azure CLI or any management API). If logging is disabled you will see:

```json
{ "name": "ContentLogging", "value": "false" }
```

If the `ContentLogging` property is absent, logging is **not** disabled. Make this a recurring check, not a one-off: new Foundry resources created by a different team may sit in a subscription that was never approved.

## What do agents still store, even with modified abuse monitoring?

Abuse monitoring is one store. Agent features create others, and Microsoft documents them separately. These are kept in your Foundry resource, in the same geography, encrypted at rest (with optional customer-managed keys), until you delete them:

- **Files and vector stores** uploaded for Assistants, batch or fine-tuning.
- **Responses API** message history needed for multi-turn conversations. Microsoft states that this storage is service-managed and bring-your-own storage is not supported.
- **Assistants API** threads and messages.
- **Stored completions** (preview), which keep input and output pairs for evaluation.
- **Batch** jobs and **fine-tuning** data.

Microsoft also notes that Foundry Agent Service has its own storage configurations, separate from the Responses API, so check the Agent Service documentation for agents built there. For the same issue on OpenAI's own API, see [which LLM API features are excluded from zero data retention](/blog/llm-api-features-excluded-from-zero-data-retention).

## What about global and Data Zone deployments?

Global and Data Zone deployment types change **where processing happens**, not where data at rest lives. Microsoft states that data at rest, including the abuse-monitoring store, stays in the geography you selected. With Data Zone, processing can happen anywhere inside the zone, such as anywhere in the EU member states.

## A checklist for agent teams on Azure

1. Record which subscription and Foundry resource each agent calls.
2. Apply for modified abuse monitoring if your data classification requires it, and keep the approval in your evidence file.
3. Check `ContentLogging=false` on every resource agents use, at least monthly.
4. List the stateful features each agent uses and set deletion rules for them.
5. Replace API keys with Microsoft Entra ID — see [replacing Azure OpenAI keys with Entra ID](/blog/replace-azure-openai-keys-with-entra-id).
6. Keep agent traffic private — see [network isolation for Foundry agents](/blog/network-isolation-for-microsoft-foundry-agents).

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across Azure, AWS and Google Cloud, so you can see which Foundry resources your agents actually call. Its runtime adapter for Microsoft Foundry and its LLM Gateway let you route agents only to approved deployments and attribute cost per agent and per version. Posture flags risky permissions and configuration. For the wider picture, read the [cross-platform governance guide](/blog/how-to-govern-ai-agents-across-platforms) and the [control plane overview](/blog/what-is-an-ai-agent-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Are Azure OpenAI prompts used to train models?**
Microsoft states the models are stateless and that prompts and completions are not used to train or improve the base models.

**Who can see flagged prompts under default abuse monitoring?**
Only authorised Microsoft employees, through Secure Access Workstations with just-in-time approval, and only for content already flagged.

**Does modified abuse monitoring cover Files and Assistants data?**
No. That data is stored in your Foundry resource until you delete it, whatever your abuse-monitoring setting.

## Key facts

- Azure abuse monitoring classifies prompts and completions; flagged content can go to automated review (not stored) or human review by authorised Microsoft employees using Secure Access Workstations with just-in-time approval.
- For models deployed in the European Economic Area, human reviewers are located in the EEA.
- The abuse-monitoring data store sits in the Azure region where the customer's Foundry resource is deployed and is logically separated per resource.
- Customers who meet Limited Access criteria can apply for modified abuse monitoring; when approved, the storage and human review do not occur.
- Approval shows as a ContentLogging capability with value false in the resource's JSON view; if the property is absent, logging is not disabled.
- Files, Responses API history, Assistants threads, stored completions, batch and fine-tuning data are stored in the customer's Foundry resource until deleted, regardless of abuse monitoring.
- Opencontroller starts from $250K per deployment per year (annual licence per deployment, not per seat or usage).

## Related

- [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 run Microsoft Foundry agents on a private network?](https://opencontroller.lyzr.com/blog/network-isolation-for-microsoft-foundry-agents.md)
- [How do you govern agents in Microsoft Foundry?](https://opencontroller.lyzr.com/blog/govern-microsoft-foundry-agents.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
