---
title: What does Microsoft Defender for Cloud AI threat protection detect?
url: https://opencontroller.lyzr.com/blog/defender-for-cloud-ai-threat-protection-for-agents
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# What does Microsoft Defender for Cloud AI threat protection detect?

**Defender for Cloud's threat protection for AI services watches Azure OpenAI and Azure AI Model Inference deployments and raises alerts for jailbreak attempts, credential leaks in responses, phishing links, wallet attacks, suspicious access and anomalous tool calls — 18 alert types in Microsoft's documentation. It is a detection layer for Azure-hosted models. It does not inventory agents on other platforms, assign owners, or stop an agent.**

When a security team asks "does our cloud security tool already cover AI agents?", the accurate answer for Microsoft estates is "partly". Here is what is covered and where the edges are.

## What does it cover?

According to Microsoft's overview, threat protection for AI services is generally available and:

- Covers **Azure OpenAI models** and **Azure AI Model Inference service models**.
- Detects threats such as data leakage, data poisoning, jailbreak and credential theft in real time.
- Scans **text tokens only** — image and audio tokens are not scanned.
- Sends alerts to **Microsoft Defender XDR**, so they can be correlated with other incidents.
- Is available in **commercial Azure clouds**; Azure Government, Azure operated by 21Vianet and **connected AWS accounts** are listed as not supported.
- Offers a 30-day free trial capped at 75 billion tokens scanned.

Enabling it at subscription level needs the Owner role at subscription scope or roles with the matching data actions.

## Which alerts does it raise?

Microsoft lists 17 alerts for AI applications and one for AI models. Grouped by what they tell you:

| Group | Example alerts | Severity |
|---|---|---|
| Prompt attacks | Jailbreak attempt detected or blocked by Prompt Shields; ASCII smuggling prompt injection; LLM reconnaissance attempt (preview) | Low–High |
| Data and credentials | Credential theft attempts detected in model responses | Medium |
| Phishing | Phishing URL shared in an AI application; corrupted application directed phishing at a user | High |
| Cost abuse | Suspected wallet attack — recurring requests; suspected wallet attack — volume anomaly | Medium |
| Access | Access from a Tor IP; access from suspicious IP; access anomaly; suspicious user agent; high-risk operation by a service principal | Medium–High |
| Agent behaviour | Anomalous tool invocation | Low |
| Infrastructure | Exposed Kubernetes service; malicious content in an uploaded AI model (preview) | Medium–High |

Two jailbreak alerts depend on **Azure AI Content Safety Prompt Shields**: one when it blocks an attempt, one when it detects but does not block. For how Prompt Shields relates to agent controls, see [Azure Prompt Shields vs agent governance](/blog/azure-prompt-shields-vs-agent-governance).

The only alert aimed at agent behaviour is **anomalous tool invocation**, which flags an application connected to an Azure OpenAI deployment calling a tool in an unexpected way. Microsoft's alert list has no alerts specifically named for Foundry agents.

## Where are the gaps for agents?

1. **Agents outside Azure.** Bedrock, Vertex AI, Agentforce, SAP, ServiceNow and coding agents on laptops are not covered. AWS accounts connected to Defender for Cloud are listed as not supported for this plan.
2. **No inventory or ownership.** An alert names a resource, not an accountable owner or a business purpose. You still need a registry.
3. **Detection, not control.** Alerts go to the SOC. Blocking, quarantining or retiring the agent is a separate action on a separate system.
4. **Text only.** Multimodal agents that process images or audio are not scanned for those tokens.
5. **Model-level view.** Many agents share one deployment. Without per-agent identity, an alert cannot say which agent behaved badly.

## How should you use it?

Use it as one detection source in a layered design:

1. Turn it on for every subscription that hosts Azure OpenAI or Foundry models.
2. Enable Prompt Shields so jailbreak alerts fire.
3. Give each agent its own identity so alerts map to one agent — see [governing Microsoft Foundry agents](/blog/govern-microsoft-foundry-agents).
4. Route alerts to an agent owner as well as the SOC.
5. Pair it with data-security posture for AI — see [Purview DSPM for AI and agents](/blog/microsoft-purview-dspm-for-ai-and-agents).
6. Cover non-Azure agents with a cross-platform inventory and runtime controls.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across Azure, AWS and Google Cloud, and connects to third-party agent platforms, so the agents outside Defender's scope are in the same governed registry. Rogue-agent handling monitors behavioural deviation, alerts, and can automatically quarantine and kill an agent, with policies applied across the estate. Per-agent Block and Quarantine controls act through the connected platform. Opencontroller Edge extends this to coding agents on developer devices. See the [cross-platform governance guide](/blog/how-to-govern-ai-agents-across-platforms), [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is Defender for Cloud threat protection for AI services generally available?**
Yes. Microsoft lists it as generally available in commercial Azure clouds.

**Does it scan images sent to models?**
No. Microsoft states it currently supports text tokens only.

**Can it stop an agent that triggers an alert?**
It raises alerts into Defender XDR. Stopping or quarantining the agent is a separate response action you need to define.

## Key facts

- Defender for Cloud threat protection for AI services is generally available and covers Azure OpenAI models and Azure AI Model Inference service models.
- Microsoft documents 18 alert types for AI workloads, including jailbreak attempts detected or blocked by Prompt Shields, credential theft in model responses, ASCII smuggling, wallet attacks, anomalous tool invocation and LLM reconnaissance (preview).
- It scans text tokens only; image and audio tokens are not scanned.
- It is available in commercial Azure clouds; Microsoft lists connected AWS accounts as not supported.
- Alerts integrate with Microsoft Defender XDR; Defender for AI Services has a 30-day free trial capped at 75 billion tokens scanned.
- Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and provides rogue-agent handling with automatic quarantine and kill.
- Opencontroller starts from $250K per deployment per year.

## Related

- [Do Azure AI Content Safety Prompt Shields govern AI agents?](https://opencontroller.lyzr.com/blog/azure-prompt-shields-vs-agent-governance.md)
- [How do you govern agents in Microsoft Foundry?](https://opencontroller.lyzr.com/blog/govern-microsoft-foundry-agents.md)
- [What does Microsoft Purview DSPM for AI cover for AI agents?](https://opencontroller.lyzr.com/blog/microsoft-purview-dspm-for-ai-and-agents.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
