For years, the OWASP Top 10 for LLM Applications ran on expert judgement. Hundreds of practitioners voted on what worried them most. This year, OWASP did something new. It checked that vote against the record of what has actually gone wrong. The 2026 edition, published on 3 August, draws on a corpus of 7,714 real incidents. The lesson for anyone testing AI is simple. What we fear and what breaks are not always the same.
Belief versus evidence
OWASP pulled incidents from public vulnerability databases and an AI harm database. Classifiers sorted the 6,639 that had enough detail. Then it compared that record with the vote.
They did not always match. Ranked on raw incidents alone, prompt injection would drop out of the top 10. OWASP kept it at number one anyway. Its reasoning is that teams defend hard against injection, so fewer clean exploits reach public databases. A low count can mean the defences are working, not that the risk is small.
Misinformation went the other way. Voters placed it near the bottom. The incident record placed it near the top. It now sits seventh, up from ninth. OWASP's point is simple. When a confident wrong answer drives a decision or a tool call, it becomes a wrong action.
The final list gives the vote three quarters of the weight and the evidence one quarter.
Excessive Agency climbs to third
Excessive Agency was sixth in 2025. It is now third, behind prompt injection and sensitive information disclosure. OWASP calls it the most consequential move on the list. This is where the vote and the record agree. Agents are where real harm is turning up.
Excessive agency means an AI system can take damaging actions when its output is wrong, ambiguous or manipulated. OWASP names three root causes:
- Excessive functionality: tools the system does not need, such as a document reader that can also delete.
- Excessive permissions: tools that connect with more access than the task requires.
- Excessive autonomy: high-impact actions that go ahead with no human approval.
One note for QA teams. Cleaning up model inputs and outputs is not a root control here. You cannot filter your way out of a tool that should never have had delete rights.
What to test now
If your AI system can call tools, your test plan needs to cover what it can do, not just what it says.
- Map what each agent can reach. List every tool, data source and downstream system each agent can reach. Remove anything unused.
- Check permissions at the source. Test them at the database or API level, not just in the agent's configuration.
- Run actions as the user. Confirm actions run in the user's own context, not under a shared high-privilege account.
- Gate irreversible actions. Test that payments, deletions and anything else you cannot undo stop for human approval.
- Test misinformation where it acts. Run misinformation tests where wrong answers trigger actions, not just where they appear on screen.
OWASP is also moving towards runtime control. On 1 September it brought the Agent Control Standard into its GenAI Security Project. It is an early open standard for making agents inspectable and traceable, and for controlling their behaviour while they run. It is young, but the direction is clear.
Start with the evidence
Belief is a starting point. Evidence is what moves the ranking.
EvalSpec gives you that evidence for how your AI system behaves. It scores your system across five quality dimensions, including accuracy, safety and robustness, using 70 ready-made test cases. Results map to the EU AI Act, NIST AI RMF and ISO 42001.
Try EvalSpec free - evaluate your AI system across all 5 dimensions. No code required.
Start your free evaluation