rag-security-reviewlisted

Audits a Retrieval-Augmented Generation (RAG) pipeline for indirect prompt injection (OWASP LLM01), knowledge-base/data poisoning (LLM04), retrieval access-control and tenant-isolation gaps (LLM02/LLM08), and citation/grounding integrity (LLM09). Use when auditing, designing, threat-modeling, or hardening any system that retrieves documents and places their text into an LLM's context window — chatbots over internal docs, agentic search, customer-support assistants, code/knowledge assistants.
fevziegeyurtsevenler/llm-security-skills · ★ 0 · AI & Automation · score 72

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

# RAG Security Review ## Purpose A RAG pipeline moves *untrusted, attacker-influenceable text* (retrieved chunks) into the same context window as *trusted instructions* (the system prompt) and often into a model that can *call tools*. That collapse of trust boundaries is the root cause of most RAG-specific vulnerabilities. This skill gives a repeatable procedure to review a RAG system across its whole data path and produce a defensible, evidence-backed findings report mapped to OWASP LLM Top 10 (2025) and MITRE ATLAS. It focuses on four attack surfaces: 1. **Indirect prompt injection** — instructions hidden in retrieved content (OWASP **LLM01:2025**, ATLAS **AML.T0051.001**). 2. **Data / knowledge-base poisoning** — corrupting what gets retrieved (OWASP **LLM04:2025**, ATLAS **AML.T0070 RAG Poisoning**, **AML.T0020**). 3. **Retrieval access control & isolation** — a user retrieving chunks they should not see; embedding/index leakage (OWASP **LLM02:2025**, **LLM08:2025**, ATLAS **AML.T0057**, **AML.T0024**). 4. **Citation & grounding integrity** — fabricated, spoofed, or exfiltrating citations and ungrounded claims (OWASP **LLM09:2025**, output-handling issues under **LLM05:2025**). ## When to use - Reviewing or designing any system where model context is assembled from a vector store, search index, database, files, emails, tickets, web pages, or tool output. - Before shipping a RAG feature, after a pipeline change (new data source, new store, new prompt template), or du