What problem does it solve? Generic web security reviews miss risks specific to systems that call a model: prompt injection through retrieved content, sensitive data disclosure, excessive agency, system-prompt leakage, vector weaknesses, and unbounded inference spend. This Skill performs a systematic, category-by-category assessment of any LLM-shaped application against the OWASP Top 10 for LLM Applications 2025. ## Core Features & Use Cases - Full LLM01–LLM10 Coverage: Evaluates prompt injection (direct and indirect), sensitive information disclosure, supply chain, data poisoning, output handling, excessive agency, system prompt leakage, vector and embedding weaknesses, misinformation, and unbounded consumption. - Evidence-Based Findings: Traces prompt assembly, retrieval pipelines, output handling, agency surfaces, and cost enforcement in code, producing numbered SEC-XXX findings with severity, attack scenarios, and recommendations. - Structured Assessment Report: Writes a standardized report to .gener8v/reviews/[slug]-owasp-llm-top10-assessment.md with coverage summary, category assessments, new findings, and positive controls. - Use Case: Before launching a RAG chatbot, run the assessment to discover that retrieved web content is substituted into prompts without delimiting (indirect injection) and that no per-tenant token ceiling exists (denial-of-wallet risk). ## Quick Start Run the OWASP LLM Top 10 review on the support-search system to check whether retrieved documents can smuggle instructions into the model.
