Tag
This paper introduces Aligned RAG Evaluation, a matched-query protocol showing that attack success rate (ASR) hides defense-induced failures in RAG knowledge-poisoning defenses, distinguishing prevented attacks from newly introduced target-positive answers on HotpotQA and Natural Questions.
This paper proposes that only AI agents capable of System 2 thinking should access untrusted documents in RAG systems to enhance security, introducing new metrics to evaluate robustness and showing reasoning models are more resistant to knowledge poisoning.