Securing AI RAG Pipelines with Fine Grained Authorization
summary
Explains a concrete ReBAC setup based on Google Zanzibar that uses SpiceDB to look up authorized document IDs and then pre-filter a Pinecone vector search before calling an LLM, claiming Zanzibar-scale throughput of 10 million queries per second. Demo code uses Pinecone+LangChain+OpenAI, creates two articles, upserts embeddings with document metadata, shows a lookup returning authorized IDs, then deletes the relationship and repeats the query so the LLM no longer returns the removed document. This proves fine-grained authorization works and is practical at small scale.
AI-written summary · about
chapters (16)
- 0:00Securing AI RAG pipelines
- 1:56Google Plus to Zanzibar auth
- 4:42Evolution of auth systems
- 8:33Zanzibar relationship-based access
- 12:03Open source Zanzibar impls
- 17:03RAG pipelines intro
- 21:00LLM app security risks
- 22:29RAG pipeline components
- 25:02Fine-grained auth for RAG
- 29:58Live demo setup
- 31:23SpiceDB relationships
- 33:07Pre-filtering authorized resources
- 34:31Auth integration with RAG
- 35:35Permission revocation demo
- 37:20Production architecture benefits
- 39:50Summary and next steps
related talks
33mSecuring RAG Pipelines with Fine Grained Authorization
44mSecuring RAG Pipelines with Fine Grained Authorization
39mSecuring RAG Pipelines with Fine Grained Authorization
33mHow to Prevent AI Agents from Accessing Unauthorized Data
29mHow to Prevent AI Agents from Accessing Unauthorized Data
30mFine-Grained Authorization: The Missing Piece in Agentic AI Security
29mHow Google Built a Consistent, Global Authorization System with Zanzibar
55mYour AI Agents Are Overprivileged: The Case for Fine-Grained Authorization