Research
OSAI AI Red Teaming — Lesson 17: RAG Poisoning Attacks
Publication date unknown · Discovered youtube.com
Page published
Publication date unknown · First observed: 27 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
SecureTechIn's experiment shows how attacker-controlled documents can dominate retrieved context and illustrates source-provenance controls defenders can evaluate at ingestion.
SecureTechIn demonstrates RAG retrieval poisoning in a synthetic, offline Python lab, where nine planted documents produced 26 poisoned chunks that ranked first for all eight target queries. An ingestion provenance allowlist reduced poisoned retrieval from 100% to 0% while preserving legitimate recall; no live models were tested, so the results demonstrate retrieval manipulation rather than confirmed downstream instruction execution.