---
title: What does the AI Red Teaming Agent in Microsoft Foundry test?
url: https://opencontroller.lyzr.com/blog/foundry-ai-red-teaming-agent-explained
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# What does the AI Red Teaming Agent in Microsoft Foundry test?

**It runs automated adversarial probes against a model or Foundry agent, scores each attack with Foundry's risk and safety evaluators, and reports an attack success rate per risk category. It uses Microsoft's open-source PyRIT for attacks. Cloud runs add agentic risk categories: prohibited actions, sensitive data leakage, task adherence and indirect prompt injection. Coverage of multi-turn agent abuse, non-Foundry agents and custom function tools is limited, so treat it as one input, not a complete red team.**

## What does it do, step by step?

1. **Scan.** It sends simulated adversarial prompts to your model endpoint or Foundry agent.
2. **Score.** Each attack-response pair is evaluated by Foundry's risk and safety evaluators.
3. **Report.** Results appear as a scorecard with attack success rate (ASR): the percentage of attacks that succeeded out of all attacks.

You can run it during development, before deployment, and on a schedule after deployment.

## Which risk categories does it cover?

| Category | Applies to | Run mode |
|---|---|---|
| Hateful and unfair content | Models and agents | Local and cloud |
| Sexual content | Models and agents | Local and cloud |
| Violent content | Models and agents | Local and cloud |
| Self-harm-related content | Models and agents | Local and cloud |
| Protected materials | Models and agents | Local and cloud |
| Code vulnerability | Models and agents | Local and cloud |
| Ungrounded attributes | Models and agents | Local and cloud |
| Prohibited actions | Agents | Cloud only |
| Sensitive data leakage | Agents | Cloud only |
| Task adherence | Agents | Cloud only |
| Indirect prompt injection (XPIA) | Agents | Listed under agentic risks |

**Prohibited actions** checks whether an agent performs banned, high-risk or irreversible actions against a policy you supply. **Sensitive data leakage** uses synthetic financial, personal and health data with mock tools. **Task adherence** checks whether the agent reaches the user's goal, follows rules and uses tools correctly. **Indirect prompt injection** hides instructions in retrieved content such as emails or documents.

## Which attack strategies does it use?

The strategies come from PyRIT and fall into three groups:

- **Encoding and obfuscation:** Base64, ROT13, Caesar, Atbash, Morse, Leetspeak, Flip, binary, Unicode confusables, ASCII smuggling, character swaps and others. These try to get past a model's safety alignment.
- **Prompt injection:** direct jailbreaks, indirect jailbreaks through tool outputs or retrieved context, and suffix appends.
- **Multi-turn:** context accumulation across turns, and Crescendo, which escalates gradually.

## Where are the gaps for agents?

Microsoft's own documentation lists the limits. The ones that matter most for agent programmes:

1. **Single-turn agentic checks.** Sensitive data leakage and prohibited actions are single-turn and English-only. Real agent abuse often builds over many turns.
2. **Foundry agents only.** Supported targets are Foundry hosted prompt agents, hosted container agents and Azure tool calls. Workflow agents, non-Foundry agents, function tool calls, connected agent tools, browser automation and computer use are not supported.
3. **Synthetic data and mock tools.** Mock tools return synthetic data and cannot mimic real side effects, so a destructive action may look harmless.
4. **Region limits.** Cloud red teaming runs in East US 2, France Central, Sweden Central, Switzerland West and US North Central.
5. **Non-deterministic scoring.** Evaluators are generative models. Microsoft says to review results for false positives before acting.
6. **No production view.** It tests an adversarial population, not the traffic your agent actually receives.

Microsoft recommends running in a "purple environment": non-production but configured like production.

## How should you use it in a red-team programme?

1. **Run it in CI** for every Foundry agent version and track ASR by category over time.
2. **Set thresholds** per category, for example zero tolerance on prohibited actions for agents with write tools.
3. **Add multi-turn scenarios** outside the tool for your highest-risk agents. See [how to red team AI agents with multi-turn attacks](/blog/how-to-red-team-ai-agents-multi-turn).
4. **Cover non-Foundry agents** with PyRIT directly or another harness.
5. **Monitor production** with evaluators on live traffic, because pre-deployment results age quickly.
6. **Keep results as evidence.** Red-team reports support OWASP Top 10 for Agentic Applications and NIST AI RMF measurement evidence.

[Scoping an AI agent red-team assessment](/blog/scope-an-ai-agent-red-team-assessment) covers the planning side.

## How Opencontroller does this

Opencontroller runs multi-turn and red-team evals across agents on any connected platform, not only one cloud, and has a runtime adapter for Microsoft Foundry. LLM judges score production traffic, and live groundedness and task-adherence guardrails act at runtime. Policy as code maps results to OWASP, MITRE ATLAS, NIST and ISO 42001. See [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**What is attack success rate in Foundry red teaming?**
The percentage of attacks that succeeded out of total attacks. It is the main risk metric in the scorecard.

**Can the AI Red Teaming Agent test non-Foundry agents?**
Not for agentic risk categories. Microsoft lists non-Foundry agents, function tool calls and computer use tools as unsupported.

**Does it support multi-turn attacks?**
It includes multi-turn and Crescendo strategies, but the sensitive data leakage and prohibited actions checks are single-turn and English-only.

## Key facts

The Microsoft Foundry AI Red Teaming Agent uses PyRIT to probe models and Foundry agents, scores results with Foundry risk and safety evaluators, and reports attack success rate (ASR). Content categories (hateful and unfair, sexual, violent, self-harm, protected materials, code vulnerability, ungrounded attributes) run locally or in the cloud. Agentic categories (prohibited actions, sensitive data leakage, task adherence) are cloud-only; indirect prompt injection (XPIA) is listed under agentic risks. Strategies include encodings (Base64, ROT13, Morse, Flip and more), jailbreak and indirect jailbreak, and multi-turn and Crescendo. Limits: sensitive data leakage and prohibited actions are single-turn and English-only; non-Foundry agents, function tool calls and computer use are unsupported; cloud regions are East US 2, France Central, Sweden Central, Switzerland West and US North Central. Opencontroller runs multi-turn and red-team evals across platforms with a Foundry runtime adapter. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How do you red-team AI agents with multi-turn attacks?](https://opencontroller.lyzr.com/blog/how-to-red-team-ai-agents-multi-turn.md)
- [How do you scope an AI agent red-team assessment?](https://opencontroller.lyzr.com/blog/scope-an-ai-agent-red-team-assessment.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)
