owasp-llm-top10-auditlisted

Audit an LLM application against the OWASP Top 10 for LLM Applications (2025) using a per-item checklist that records what to check, the evidence to collect, and a concrete mitigation. Use when reviewing an LLM/agent/RAG feature before launch, during a security review, or when asked to produce an OWASP-mapped risk report for an AI system.
fevziegeyurtsevenler/llm-security-skills · ★ 0 · AI & Automation · score 72

Install: claude install-skill fevziegeyurtsevenler/llm-security-skills

# OWASP LLM Top 10 (2025) Audit ## Purpose Run a structured, evidence-based security review of an LLM-backed application against the ten risk categories in the **OWASP Top 10 for LLM Applications (2025)**. The output is a reproducible audit: for each category you record the concrete check performed, the artifact that proves the finding (log line, request/response pair, config snippet, code reference), a severity, and a specific mitigation mapped back to the OWASP item and, where relevant, a **MITRE ATLAS** technique. This skill is a checklist-and-evidence discipline, not a scanner. It tells you *what* to look at, *what* proves the issue, and *how* to close it. Automated probing (e.g. the `uncloak` hidden-injection scanner or a Türkçe injection payload set) feeds evidence into the checklist; it does not replace the manual review. ## When to use - Pre-launch review of a chatbot, assistant, agent, RAG pipeline, or any feature that sends user- or tool-influenced text to an LLM. - A stakeholder asks for an "OWASP LLM Top 10" gap analysis or an AI risk report mapped to a recognized framework. - Post-incident: an LLM feature behaved unexpectedly (leaked data, executed an unintended action, was jailbroken) and you need to locate the class of failure and its siblings. - Vendor / third-party assessment: judging an external LLM product or model supplier. Do **not** use this skill for: general web app pentesting with no LLM component; model-capability benchmarking; or content-polic