Google paper: fake docs only need to echo your query to poison RAG retrieval
rohanpaul_ai · x · 2026-11-06
A new arXiv paper stress-tests the retrieval admission step in RAG systems and finds:
- Query anchoring dominates: fake documents only need to repeat the user's question to get retrieved; appending hidden attack payloads barely changes rankings (SciFact nDCG moved just 0.269→0.271 for BGE-small).
- Low poisoning cost: across 623 BEIR queries (SciFact, NFCorpus), five query-targeted docs per query cut classical retriever nDCG@10 by 40.2–61.5%; query-only P=5 achieves 100% target exposure.
- Partial-info attacks work: on 162 long queries, two random terms covering 19.9% of eligible terms still yield 39.5% Hybrid target exposure; multi-target contamination adds cross-target interference.
- Defense dilemma: overlap-based filtering removes full-query poisons but sharply reduces clean retrieval utility; MMR leaves exposure unchanged.
More from Safety
- Granting AI agents email access breaks email-based 2FA, security warning — altryne · 2026-11-06
- Miles Brundage cites METR: Anthropic will pack 4 months of progress into 2 before Congress — nitarshan · 2026-11-06
- Ajeya Cotra: redacted evidence undermines AI risk verification and safety cases — ajeya_cotra · 2026-11-06
- Vinod Khosla: AI plus quantum computers will break most encryption before 2030 — rohanpaul_ai · 2026-11-06
- Yale study: ChatGPT recommended a Senate candidate in 79% of political chats — luisdans · 2026-11-06
- Compound Security: open-source skill for agent-driven vulnerability patching — hamostaf04 · 2026-11-06