Signed but Steearable: Study Exposes Vector-Text Integrity Gap in Offline RAG
securagenai · hf · 2026-11-07
A Hugging Face paper shows that signed retrieval packs in offline retrieval-augmented generation can contain embedding vectors that don't match their text, leaving many queries vulnerable even with practical vector verification and compressed-index audits.
The authors propose code-level verification with calibrated ambiguity handling that accepts honest vectors while rejecting tampered ones. Accompanying plots on BGE-M3 and all-MiniLM-L6-v2 show the share of hijackable queries climbing from near zero to 1 between FP32 noise levels and common CPU INT8 thresholds across SQuAD, SciFact, NFCorpus, FiQA and ArguAna — meaning slightly loose tolerance thresholds leave almost everything exposed.
More from Safety
- Google team confirms Hacktron AI submitted high-impact vulns, expands partnership — ithilgore · 2026-11-07
- AI governance is plumbing: the last 20% breaks without the right foundations — vasuman · 2026-11-07
- Sandberg dissects the 1940 argument that harnessing atomic energy was impossibly remote — anderssandberg · 2026-11-07
- AI Safety Researcher Vows to Stay, Calls on OpenAI and Anthropic to Stop RSI — trevposts · 2026-11-07
- Researchers Propose "Inter-AI Morality" Agenda: How AI Systems Treat One Another — dioscuri · 2026-11-07
- US Unveils National DNA Synthesis Screening Policy; Expert Backs Independent Verification, Flags Benchtop Concerns — anshulkundaje · 2026-11-07