PoisonedRAG

Capture Summary

Paper introducing knowledge corruption attacks against RAG. The abstract reports that injecting a few malicious texts into a large knowledge database can induce attacker-chosen target answers for target questions, and notes that evaluated defenses were insufficient.

Relevance

• Foundational RAG poisoning source for Security for AI.
• Important for challenges around retrieval trust, corpus governance, and detection of poisoned knowledge.
• Candidate for research questions on evidence provenance and robust retrieval.

Collection Notes

Collected as a core RAG/data security paper.