RAG over-retrieval / entitlement-failure register
Public register of incidents in which an AI assistant, agent, or RAG system surfaced content a principal was not entitled to see, or amplified access.
6 posts
Public register of incidents in which an AI assistant, agent, or RAG system surfaced content a principal was not entitled to see, or amplified access.
Purview classifies, labels, and governs your data estate. It does not decide, at query time, whether a user may retrieve a document from AI you build.
A RAG pipeline can enforce access inside the vector query or in the app after results return. Each has a distinct failure mode. Here is what breaks.
Embeddings throw away the permissions your source systems already track. Here is the recipe to carry document-level permissions into a RAG pipeline.
Gateco is not a RAG framework. It is the authorization layer you insert at the retrieval step of LangChain or LlamaIndex. Here is where it goes, and why.
Vector databases retrieve by embedding similarity. They don't know who's asking or check permissions. That is the RAG security gap, and it is wide.