Signed but Steerable: Measuring and Closing the Text-Vector Integrity Gap in Offline Retrieval-Augmented Generation

A new paper, "Signed but Steerable: Measuring and Closing the Text-Vector Integrity Gap in Offline Retrieval-Augmented Generation," identifies a security gap in offline RAG systems on edge devices that are updated via signed knowledge packs carrying text and precomputed embedding vectors: a signature proves who published a pack but not that its vectors actually encode its text. The authors propose a verifier that recomputes vectors to check text-vector consistency, addressing the risk that signed packs can be steered through mismatched embeddings.

Offline retrieval-augmented generation RAG on edge devices is updated through signed knowledge packs that carry text and precomputed embedding vectors. A signature proves who published a pack, not that its vectors encode its text. We study this text-vector integrity gap. A verifier that recomputes