Preventing Document Permission Leaks in RAG Access Control

Access control is often missing from retrieval-augmented generation systems, and the resulting permission leaks typically surface only after deployment. The author has found the failure to be structural: ingestion jobs run with broad read access, pipelines discard source permissions, and retrieval ranks content by similarity without checking who is asking. Five mechanisms cause most leaks, including dropped permissions at ingestion, privileged index builders, authorization applied in the wrong layer, caller-supplied filters, and stale permissions. The OWASP Top 10 for LLM Applications lists unauthorized access and data leakage as the leading vector and embedding risk, recommending fine-grained access controls and logical partitioning. The article goes on to cover the confused-deputy pattern, enforcement in vector stores, permission refresh after revocation, and a set of tests for judging whether a retrieval system is safe.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in