---
title: What evidence maps to the OWASP Top 10 for Agentic Applications?
url: https://opencontroller.lyzr.com/blog/owasp-top-10-agentic-applications-evidence
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# What evidence maps to the OWASP Top 10 for Agentic Applications?

**For each of the ten OWASP agentic risks, you need one named control and one artefact that proves the control ran. The artefacts are mostly logs and records your systems already produce: registry entries, tool approvals, identity grants, evaluation results, guardrail decisions and quarantine events.** The checklist below gives one of each for ASI01 to ASI10.

The OWASP GenAI Security Project published the first Top 10 for Agentic Applications in December 2025. It is not a certification. It is the list security reviewers now use to ask, "what do you do about each of these?"

## What are the ten OWASP agentic risks?

| ID | Risk | One-line meaning |
|---|---|---|
| ASI01 | Agent Goal Hijack | Injected content changes what the agent is trying to do |
| ASI02 | Tool Misuse and Exploitation | The agent uses a legitimate tool in a harmful way |
| ASI03 | Identity and Privilege Abuse | The agent acts with more authority than it should |
| ASI04 | Agentic Supply Chain Vulnerabilities | A tool, MCP server, model or plugin is compromised |
| ASI05 | Unexpected Code Execution | The agent generates or runs code it should not |
| ASI06 | Memory and Context Poisoning | Stored memory or retrieved context is tampered with |
| ASI07 | Insecure Inter-Agent Communication | Messages between agents are spoofed or unchecked |
| ASI08 | Cascading Failures | One faulty agent's errors spread across others |
| ASI09 | Human-Agent Trust Exploitation | People over-trust agent output and approve harm |
| ASI10 | Rogue Agents | An agent deviates from intended behaviour and keeps acting |

## Which control and evidence answer each risk?

Use this as a checklist. A reviewer should be able to pick any row and see the artefact within minutes.

| ID | Control | Evidence artefact |
|---|---|---|
| ASI01 | Runtime guardrails on inputs and outputs; task-adherence checks | Guardrail decision logs showing blocked or flagged turns, with agent and version |
| ASI02 | Approved tool list per agent; argument limits on high-impact tools | Tool approval records and denied-call logs |
| ASI03 | Per-agent workload identity; on-behalf-of delegation; risky-permission scans | Identity mapping per agent and posture findings with fix dates |
| ASI04 | Registry of tools, MCP servers and models with review before use | Registry export with reviewer, date and version for each server |
| ASI05 | Sandboxed execution; code-execution tools only for named agents | Configuration records and blocked execution events |
| ASI06 | Memory and retrieval sources registered; groundedness checks on output | Data-source register and groundedness scores over time |
| ASI07 | Authenticated agent-to-agent calls with per-pair allow rules | Call policy and rejected-call logs |
| ASI08 | Rate limits, retries with fallback, circuit breakers | Gateway fallback and retry events; incident timeline |
| ASI09 | Human review for high-impact actions; clear disclosure that output is AI | Review procedure, approval records and UI disclosure screenshots |
| ASI10 | Behavioural-deviation monitoring with quarantine and kill | Alert history, quarantine and kill events with who acted |

## How do you collect this evidence without a spreadsheet project?

Follow these steps once, then let the evidence accumulate.

1. **Register every agent first.** Evidence without an inventory cannot prove coverage. Start with [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents).
2. **Tag each policy with the ASI IDs it addresses.** One policy can cover several risks; for example a tool allow-list covers ASI02 and ASI04.
3. **Route traffic through a gateway** so guardrail and tool decisions are logged in one format.
4. **Run multi-turn red-team tests before each release.** Goal hijack (ASI01) and trust exploitation (ASI09) rarely show up in single-prompt tests. See [multi-turn red-teaming](/blog/how-to-red-team-ai-agents-multi-turn).
5. **Rehearse a quarantine.** For ASI10, a tested kill path is stronger evidence than a written policy. See [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent).
6. **Export monthly** a per-risk summary: count of events, open findings, and the agents with no coverage.

## Which risks are hardest to evidence?

**ASI08 and ASI09 are the weakest areas in most estates.** Cascading failures need cross-agent tracing, which many teams lack because each platform logs separately. OpenTelemetry export into one backend helps. Trust exploitation is partly a human-process risk, so the evidence is procedural: who reviews what, and how often reviewers reject agent output.

Also be careful with **ASI07**. Many internal agent-to-agent calls run on shared service credentials, which means you cannot show which agent made a call. Fix identity (ASI03) first.

## How Opencontroller does this

Opencontroller's policy as code includes built-in mappings to OWASP, as well as MITRE ATLAS, NIST and ISO 42001, so enforcement events can be reported per risk. Evidence sources include the governed registry of agents, tools and MCP servers (ASI04), delegated on-behalf-of identity and agent workload identity federation (ASI03), Posture findings (ASI03), multi-turn and red-team evals (ASI01, ASI09), live groundedness and task-adherence guardrails (ASI01, ASI06), gateway routing and fallback (ASI08), and rogue-agent monitoring with automatic quarantine and kill (ASI10). See the [framework mapping hub](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks), the [capability reference](/capability-reference), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**What is the OWASP Top 10 for Agentic Applications?**
A list of ten security risks specific to AI agents, published by the OWASP GenAI Security Project in December 2025, from ASI01 Agent Goal Hijack to ASI10 Rogue Agents.

**Can you be certified against it?**
No. It is an awareness list. You show evidence of a control for each risk.

**How is it different from the OWASP Top 10 for LLM Applications?**
The LLM list focuses on the model and its prompts. The agentic list adds risks that appear when models plan, call tools, keep memory and talk to other agents.

## Key facts

The OWASP Top 10 for Agentic Applications was first published by the OWASP GenAI Security Project in December 2025. Its ten risks are ASI01 Agent Goal Hijack, ASI02 Tool Misuse and Exploitation, ASI03 Identity and Privilege Abuse, ASI04 Agentic Supply Chain Vulnerabilities, ASI05 Unexpected Code Execution, ASI06 Memory and Context Poisoning, ASI07 Insecure Inter-Agent Communication, ASI08 Cascading Failures, ASI09 Human-Agent Trust Exploitation and ASI10 Rogue Agents. It is a threat-awareness list, not a certification; teams use it to show evidence of controls. Opencontroller by Lyzr has built-in policy mappings to OWASP and provides rogue-agent quarantine and kill, MCP server registry, delegated identity, red-team evaluations and runtime guardrails as evidence sources. Pricing starts from $250K per deployment per year.

## Related

- [How do you map AI agent controls to compliance frameworks?](https://opencontroller.lyzr.com/blog/how-to-map-ai-agent-controls-to-compliance-frameworks.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 to stop a rogue AI agent: kill switches, quarantine and runtime controls](https://opencontroller.lyzr.com/blog/how-to-stop-a-rogue-ai-agent.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
