Document Poisoning in RAG Systems: How Attackers Corrupt Your AI's Sources
Summary
A detailed exploration of document poisoning in retrieval-augmented generation (RAG) systems. Amine Raji demonstrates how three crafted documents can mislead an LLM about a company's finances, outlines the PoisonedRAG theory, the practical lab setup, and compares defense layers, highlighting ingestion-time embedding anomaly detection as the most effective defense.