Answers are grounded in this paper and cite their sources.
Enterprise assistants built on retrieval-augmented generation (RAG) read company chats, wikis and policies, so anyone who can post into those sources can plant instructions that the assistant may follow. We present Sentinel-RAG, a controlled testbed in which three assistants share the same corpus, retriever and answer model and differ only in the security stage between retrieval and generation: none, role-based access control plus a 1.6-million-parameter Transformer classifier trained from scratch, or the same access control plus a local 2-billion-parameter LLM judge with schema-constrained output. The classifier reaches 89.8% accuracy and 0.895 macro F1 on held-out attack patterns. On a 25-question suite under two corpus conditions, the unguarded assistant gave 4 and 10 unsafe answers, while both guards gave none. The classifier added 8–9 ms per query but removed 9–16% of benign chunks; the judge removed none but added 13–14 s per query.
No takes yet. Share an insight, caveat, or question.
Borpujari et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: