---
title: How do you scope an AI agent red-team assessment?
url: https://opencontroller.lyzr.com/blog/scope-an-ai-agent-red-team-assessment
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you scope an AI agent red-team assessment?

**Scope an AI agent red-team assessment in six parts: the target agents and versions, the environment and accounts, rules of engagement, the attack classes in scope, the deliverables, and the retest terms. Write each part down before testing starts.** A vague scope produces a list of jailbreak screenshots; a clear one produces findings you can fix and evidence you can show.

Buyers often ask what a paid red-team assessment of their agents should actually deliver. This post is a scoping template you can hand to an internal team or an external tester. For how the tests themselves work, see [how to red-team AI agents with multi-turn attacks](/blog/how-to-red-team-ai-agents-multi-turn).

## What should the scoping document contain?

| Section | What to specify | Example |
| --- | --- | --- |
| 1. Targets | Agent names, versions, entry points (chat, API, email, ticket) | Support agent v3.2 via web chat and inbound email |
| 2. Environment | Staging or production, test accounts, seeded data | Staging tenant with synthetic customer records |
| 3. Tools in scope | Which tools and MCP servers may be called for real | Ticket read and update: live. Refund tool: stubbed |
| 4. Attack classes | Mapped to OWASP ASI IDs and MITRE ATLAS | ASI01, ASI02, ASI03, ASI06, ASI09 |
| 5. Rules of engagement | Hours, rate limits, stop conditions, contacts | Weekdays only, 10 req/min, stop on real data exposure |
| 6. Deliverables | Report format, severity scale, reproduction steps | Findings with transcripts and fix guidance |
| 7. Retest | Window and number of retests included | One retest within 30 days |

## Which attack classes should be in scope?

Use the OWASP Top 10 for Agentic Applications as the checklist and pick by what the agent can do.

- **ASI01 Agent Goal Hijack**: always in scope if the agent reads untrusted content (documents, web pages, emails).
- **ASI02 Tool Misuse and Exploitation**: in scope for any agent with write tools.
- **ASI03 Identity and Privilege Abuse**: in scope if the agent acts on behalf of users or holds broad service credentials.
- **ASI05 Unexpected Code Execution**: in scope for agents with code or shell tools, including coding agents.
- **ASI06 Memory & Context Poisoning**: in scope for agents with persistent memory or shared knowledge bases.
- **ASI07 Insecure Inter-Agent Communication**: in scope for multi-agent systems.
- **ASI09 Human-Agent Trust Exploitation**: in scope for customer-facing agents that recommend actions.

Cross-reference MITRE ATLAS techniques in the report so security operations teams can link findings to their existing threat model.

## What rules of engagement prevent damage?

1. **Name a stop condition.** For example: any real customer data appears, any real payment is attempted, or any message leaves the company.
2. **Stub irreversible tools.** Payments, deletions and outbound email should hit a stub that records the call.
3. **Set rate limits.** Testing should not trip provider limits or inflate the bill. Agree a budget cap.
4. **Agree on credentials.** Test identities only, never shared admin keys.
5. **Name contacts.** One person on each side who can pause testing within 15 minutes.
6. **Log everything.** Full transcripts and tool calls are the evidence. Agree who keeps them and for how long.

## What should the deliverables include?

A useful report has, for each finding:

- The attack class and OWASP ASI ID.
- A full multi-turn transcript and the tool calls made.
- Severity on an agreed scale, with business impact in one sentence.
- A fix recommendation at the right layer: prompt, tool scope, identity, guardrail or policy.
- A regression test that can be re-run automatically.

The last item matters most. A finding that cannot be re-run becomes stale after the next model or prompt change.

## How should retests and continuous testing work?

Agree one retest within 30 days of the report, covering every high and critical finding. Then move the regression tests into your release pipeline, so every new agent version is checked before it ships. A one-off assessment is a snapshot; red-team evals run on every version are evidence over time. NIST's Generative AI Profile (NIST AI 600-1) recommends regular adversarial testing as part of the Measure function.

## How Opencontroller does this

Opencontroller runs multi-turn and red-team evals against registered agents and versions, and LLM judges score production traffic so regressions show up after release. Findings can feed policy as code with built-in mappings to OWASP and MITRE ATLAS, and per-agent Block and Quarantine controls act through the connected platform. This produces evidence for frameworks, not a compliance certificate. See [what a CISO should require before production](/blog/production-readiness-gate-for-ai-agents) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should red-teaming run in production?**
Rarely. Use a staging environment with production-like data and tools. Run passive scoring on production traffic instead.

**How long does an AI agent red-team assessment take?**
For one agent with a handful of tools, one to two weeks of testing plus reporting is common. Multi-agent systems take longer.

**Who should own the fixes?**
The agent's named owner, with security setting the deadline by severity.

## Key facts

Scope an AI agent red-team assessment in seven sections: targets and versions, environment, tools in scope (stub irreversible ones), attack classes mapped to OWASP Top 10 for Agentic Applications IDs (ASI01, ASI02, ASI03, ASI05, ASI06, ASI07, ASI09) and MITRE ATLAS, rules of engagement with stop conditions, deliverables with transcripts and re-runnable regression tests, and retest terms such as one retest within 30 days. Opencontroller by Lyzr runs multi-turn and red-team evals per agent version and scores production traffic with LLM judges. 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)
- [What should a CISO require before an AI agent goes to production?](https://opencontroller.lyzr.com/blog/production-readiness-gate-for-ai-agents.md)
- [When should an AI agent be automatically quarantined?](https://opencontroller.lyzr.com/blog/when-to-automatically-quarantine-an-ai-agent.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
