fix-ai-agent-features
Fix defects in LLM and agent features — prompt injection, data leakage, unsafe output handling, excessive tool agency, runaway cost, inconsistent or ungrounded outputs — with server-side permission checks, escaping, budgets, and grounding rules. Use for any model- or agent-driven feature bug.
72
91%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Passed
No findings from the security scan
Fix AI and Agent Features
Procedure
- Treat model input and output as untrusted. Escape output before HTML, shell, SQL, or templates.
- Injection: separate instructions from data; strip or quarantine instruction-like content from retrieved pages, emails, files, and API responses; never let content trigger tools on its own.
- Agency: tools check the user's permissions on the server; consequential actions need approval enforced outside the model.
- Leakage: scope retrieval and memory by tenant; keep secrets out of prompts.
- Cost and loops: token, time, and tool-call budgets with hard stops.
- Consistency: fix the cause of score drift (temperature, nondeterministic inputs, unstable rubric); prove spread on 3+ runs is within tolerance.
- Grounding: every claim maps to provided evidence; free versus paid content can't leak through paraphrase.
- Check each OWASP Agentic item (ASI01–ASI10) touched by the change.
Exit criteria
The tester's AI suite passes, including injection cases, budgets trip correctly, and consistency is within tolerance.
Counterpart of tester skill: qa-ai-agent-testing.
- Repository
itsablabla/qa-bug-agent-skills
- Revision
9a6f628
- Last updated
- First committed
Is this your skill?
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.