Atlas › 07 Security Testing for AI Apps › RAG System Security Testing☰ Read as one page
RAG System Security Testing
6.1The RAG Attack SurfaceRetrieval-Augmented Generation (RAG) systems introduce a unique attack surface at the intersection of search and generation. The retrieval…6.2RAG-Specific Threat Model6.3RAG Security Test Suite6.4Indirect Prompt Injection via RAGThe most dangerous RAG vulnerability is indirect prompt injection: malicious instructions embedded in documents that the LLM processes.6.5Systematic Output Verification with Eval FrameworksHand-written assertions catch known-bad patterns, but grounding and citation accuracy need continuous, metric-driven evaluation. As of July…6.6RAG Security Best Practices1. Content sanitization on ingestion. Strip hidden HTML, invisible characters, and known injection patterns before storing documents in the…