A cybersecurity student released GateKeep RAG, an open-source framework treating multi-tenant RAG isolation as a database and retrieval problem instead of a prompt-guardrail problem. It enforces authorization at the vector-search and relational layers, with an audit ledger and adversarial evaluation showing zero leaks across 823,996 canary checks. The work highlights that prompt instructions fail as authorization boundaries once unauthorized chunks enter an LLM's context.
A RAG service where access control happens before the LLM: each chunk carries tenant ID, roles, and clearance metadata, and only chunks the user is allowed to read reach the model. It solves cross-tenant data leakage in multi-user RAG deployments, where prompt instructions are not a real security boundary. Aimed at teams building RAG over documents with per-user or per-tenant permissions.
A RAG service where access control happens before the LLM: each chunk carries tenant ID, roles, and clearance metadata, and only chunks the user is allowed to read reach the model. It solves cross-tenant data leakage in multi-user RAG deployments, where prompt instructions are not a real security boundary. Aimed at teams building RAG over documents with per-user or per-tenant permissions.
AWS describes how Amazon Quick and Amazon Bedrock Knowledge Bases enforce document-level permissions in enterprise RAG over sources like SharePoint, Google Drive, and Confluence. The system uses a hybrid two-stage design: pre-retrieval ACL filtering on a cached vector index, followed by real-time permission verification with the authoritative source at query time. This addresses stale or incorrectly replicated ACLs that could leak sensitive documents through AI answers.
AWS describes how Amazon Quick and Amazon Bedrock Knowledge Bases enforce document-level permissions in enterprise RAG over sources like SharePoint, Google Drive, and Confluence. The system uses a hybrid two-stage design: pre-retrieval ACL filtering on a cached vector index, followed by real-time permission verification with the authoritative source at query time. This addresses stale or incorrectly replicated ACLs that could leak sensitive documents through AI answers.
A YouTube talk features Aaron Linsky (Bridgewater) and Nikhil Benesch (turbopuffer) explaining how Bridgewater's Vault — 100 million documents with per-document permissions — broke conventional filtered vector search, and how turbopuffer's 'native filtering' fixed it. The discussion is a practitioner case study in making retrieval-augmented AI respect document-level access control at scale.
A YouTube talk features Aaron Linsky (Bridgewater) and Nikhil Benesch (turbopuffer) explaining how Bridgewater's Vault — 100 million documents with per-document permissions — broke conventional filtered vector search, and how turbopuffer's 'native filtering' fixed it. The discussion is a practitioner case study in making retrieval-augmented AI respect document-level access control at scale.
A developer published sentinel-eval, an Apache-2.0 offline test harness showing that metadata filters like allowed_groups prevent leaks in direct multi-tenant RAG queries (0% leaks) but fail on agent write paths, where derived summary chunks lose upstream access control tags (2 of 3 synthesis tasks leaked, exposing 6/18 restricted canaries in a Chroma + LlamaIndex setup). The harness lets teams reproduce the problem locally without API keys or live LLM calls.
A developer published sentinel-eval, an Apache-2.0 offline test harness showing that metadata filters like allowed_groups prevent leaks in direct multi-tenant RAG queries (0% leaks) but fail on agent write paths, where derived summary chunks lose upstream access control tags (2 of 3 synthesis tasks leaked, exposing 6/18 restricted canaries in a Chroma + LlamaIndex setup). The harness lets teams reproduce the problem locally without API keys or live LLM calls.
A developer demonstrated that metadata filters like allowed_groups protect multi-tenant RAG on direct reads but fail when background agents with elevated access write derived summaries into shared collections. In a Chroma and LlamaIndex setup, summary chunks lose source ACLs, leaking restricted data. An open-source harness, sentinel-eval, reproduces the failure offline.
A developer demonstrated that metadata filters like allowed_groups protect multi-tenant RAG on direct reads but fail when background agents with elevated access write derived summaries into shared collections. In a Chroma and LlamaIndex setup, summary chunks lose source ACLs, leaking restricted data. An open-source harness, sentinel-eval, reproduces the failure offline.