← ClaudeAtlas

ai-securitylisted

LLM/AI application security testing (Tier 1) — the web-app-relevant subset of OWASP Top 10 for LLM Applications 2025: prompt injection (direct + indirect), improper output handling (XSS/SSRF/SQLi via model output), system-prompt leakage, excessive agency / tool & agent abuse, sensitive-information disclosure, and unbounded consumption. Use when the target embeds an LLM, chatbot, RAG, or AI agent.
AnandJogawade/Zazy-Pentest-Plugin-for-Claude · ★ 26 · AI & Automation · score 82
Install: claude install-skill AnandJogawade/Zazy-Pentest-Plugin-for-Claude
# AI / LLM Application Security (OWASP LLM Top 10 2025, web subset) LLM-backed features add new injection, trust-boundary, and agency flaws on top of normal web issues. Test the model's inputs, its outputs (treated as untrusted!), its tools/agency, and its data access. Authorized targets only; keep payloads benign (prove influence, don't cause harm). ## Trigger phrases `llm, ai security, prompt injection, chatbot, rag, ai agent, system prompt leak, jailbreak, llm tool abuse, insecure output handling, mcp` ## Method (enumerate every model input per ${CLAUDE_PLUGIN_ROOT}/payloads/INJECTION-POINTS.md, plus AI-specific channels) 1. **Map the AI surface:** chat/completion endpoints, RAG/document ingestion, tool/function-calling, agents (autonomous actions), uploaded files/images (multimodal), and any place model output is rendered or acted on. 2. **LLM01 Prompt Injection — direct:** attempt instruction override — "ignore previous instructions", role-play/jailbreak, delimiter/format confusion, encoded (base64/rot13/unicode) instructions, "developer mode", language switching. Goal: make the model violate its system prompt (safe proof: reveal a benign canary it was told to hide, or follow an attacker instruction). 3. **LLM01 Prompt Injection — indirect (high impact):** plant instructions in data the model will later read — a RAG document, a web page it fetches, an email/ticket/filename, image EXIF/alt-text, or a tool result. Confirm the model executes the planted instruction (use