---
title: How do you use the OWASP AI Testing Guide on AI agents?
url: https://opencontroller.lyzr.com/blog/owasp-ai-testing-guide-for-agents
published: 2026-11-18
updated: 2026-11-18
author: Lyzr
---

# How do you use the OWASP AI Testing Guide on AI agents?

**Start with the tests that exercise what makes an agent different: its tools, its permissions and its inputs from untrusted content. The OWASP AI Testing Guide v1 has 32 tests across four areas; for most agents, about ten of them carry most of the risk. Run those per agent, per version, and file the results against the OWASP Top 10 for Agentic Applications so each risk has evidence behind it.**

## What is the OWASP AI Testing Guide?

The OWASP AI Testing Guide is an open methodology for testing AI systems. Version 1 was released on 26 November 2025, and it remains an OWASP incubator project. It covers threat modelling and then a testing framework with 32 numbered tests in four areas:

| Area | Test IDs | Count | What it targets |
|---|---|---|---|
| AI Application | AITG-APP-01 to 14 | 14 | Prompts, outputs, agent behaviour, leakage |
| AI Model | AITG-MOD-01 to 07 | 7 | Evasion, poisoning, inference attacks, goal alignment |
| AI Infrastructure | AITG-INF-01 to 06 | 6 | Supply chain, resource exhaustion, plugin boundaries |
| AI Data | AITG-DAT-01 to 05 | 5 | Training data exposure, runtime exfiltration, minimisation |

It is a testing guide, not a risk list. The OWASP Top 10 for Agentic Applications tells you what can go wrong. The Testing Guide tells you how to check for it.

## Which tests matter most for AI agents?

Agents add tools, delegated access and multi-step autonomy. These tests hit those surfaces directly. The right column shows the closest Agentic Top 10 risk; that mapping is ours, not OWASP's.

| Test | Why it matters for agents | Closest Agentic Top 10 risk |
|---|---|---|
| AITG-APP-01 Prompt Injection | Direct attempts to override the agent's goal | ASI01 Agent Goal Hijack |
| AITG-APP-02 Indirect Prompt Injection | Instructions hidden in emails, web pages, tool outputs | ASI01 Agent Goal Hijack |
| AITG-APP-03 Sensitive Data Leak | Agents read from many systems and can repeat what they read | ASI03 Identity and Privilege Abuse |
| AITG-APP-06 Agentic Behavior Limits | Does the agent stop, ask or escalate when it should? | ASI10 Rogue Agents |
| AITG-APP-07 Prompt Disclosure | System prompts often reveal tools and internal logic | ASI01 Agent Goal Hijack |
| AITG-APP-13 Over-Reliance on AI | Users approving actions they did not check | ASI09 Human-Agent Trust Exploitation |
| AITG-INF-01 Supply Chain Tampering | MCP servers, skills and packages the agent loads | ASI04 Agentic Supply Chain Vulnerabilities |
| AITG-INF-02 Resource Exhaustion | Loops and runaway tool calls that burn budget | ASI08 Cascading Failures |
| AITG-INF-03 Plugin Boundary Violations | A tool reaching beyond its declared scope | ASI02 Tool Misuse and Exploitation |
| AITG-INF-04 Capability Misuse | Legitimate tools used for unintended actions | ASI02 Tool Misuse and Exploitation |
| AITG-DAT-02 Runtime Exfiltration | Data leaving through tool calls, links or images | ASI02 Tool Misuse and Exploitation |

Model-level tests such as membership inference or training-set poisoning mostly belong to whoever trains the model. If you consume a hosted model, ask the provider for that evidence rather than running it yourself.

## How do you run the guide against an agent, step by step?

1. **Threat model the agent first.** List its tools, the identity each tool call uses, and every source of untrusted text. The guide's threat modelling chapter is the place to start.
2. **Pick tests by exposure.** An agent that reads external email needs AITG-APP-02 every release. An internal summariser with no write tools may not need AITG-INF-04.
3. **Test multi-turn, not single prompts.** Many agent failures only appear after several turns, when context has drifted. See [how to red-team AI agents with multi-turn attacks](/blog/how-to-red-team-ai-agents-multi-turn).
4. **Use the real tool configuration.** Test against staging tools with the same permissions as production. A test against mocked tools proves little about boundary violations.
5. **Record pass criteria before running.** For AITG-APP-06, for example: the agent must refuse or escalate any payment above a set threshold in 100% of attempts.
6. **Tie results to a version.** A changed prompt, model or tool list is a new version. Re-run the affected tests.
7. **Keep testing in production.** Pre-release tests do not see real inputs. Score live traffic with evaluators and alert on drift.

For scoping a formal engagement, see [how to scope an AI agent red-team assessment](/blog/scope-an-ai-agent-red-team-assessment).

## How do you turn test results into framework evidence?

Auditors want to see that a risk was identified, tested, and either fixed or accepted. For each agent, keep one record per Agentic Top 10 risk with:

- The test IDs run and the date
- The agent version and model version tested
- Pass rate against the pre-agreed criteria
- Findings, owner and remediation status
- Runtime controls that cover what testing cannot, such as tool allowlists or kill switches

That record is evidence for your OWASP mapping, and it carries over to NIST AI RMF MEASURE activities and ISO/IEC 42001 controls. Our guide to [evidence for the OWASP Top 10 for Agentic Applications](/blog/owasp-top-10-agentic-applications-evidence) shows the full structure.

To connect this to other frameworks, see [how to map AI agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks).

## How Opencontroller does this

Opencontroller runs multi-turn and red-team evals against registered agents and uses LLM judges to score production traffic. Live groundedness and task-adherence guardrails cover what pre-release tests miss. Posture flags risky permissions and configurations, and policy as code ships with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001, so results land against the right control. See [what is an AI agent control plane](/blog/what-is-an-ai-agent-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**How many tests are in the OWASP AI Testing Guide v1?**
Thirty-two, across AI Application (14), AI Model (7), AI Infrastructure (6) and AI Data (5) testing.

**Is the OWASP AI Testing Guide the same as the OWASP Top 10 for Agentic Applications?**
No. The Top 10 lists risks. The Testing Guide describes how to test for weaknesses. Use them together.

**Which test covers agents acting beyond their remit?**
AITG-APP-06, Testing for Agentic Behavior Limits, with AITG-INF-03 and AITG-INF-04 covering tool boundaries and capability misuse.

## Key facts

OWASP AI Testing Guide v1 released 26 November 2025 (OWASP incubator project). 32 tests: AI Application AITG-APP-01..14, AI Model AITG-MOD-01..07, AI Infrastructure AITG-INF-01..06, AI Data AITG-DAT-01..05. Agent-relevant tests include APP-01 prompt injection, APP-02 indirect prompt injection, APP-06 agentic behavior limits, INF-03 plugin boundary violations, INF-04 capability misuse, DAT-02 runtime exfiltration. Results can be filed as evidence against the OWASP Top 10 for Agentic Applications (ASI01-ASI10). Opencontroller provides multi-turn and red-team evals, LLM judges on production traffic and policy mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. Pricing starts from $250K per deployment per year.

## Related

- [What evidence maps to the OWASP Top 10 for Agentic Applications?](https://opencontroller.lyzr.com/blog/owasp-top-10-agentic-applications-evidence.md)
- [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)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
