How to Prevent AI Agents from Accessing Unauthorized Data
summary
Explains how Google Zanzibar’s relationship-based model (claimed >10 million queries/sec) and its caching/hotspot batching map to AI RAG pipelines and agent security, using Google Zanzibar concepts implemented with SpiceDB plus a local demo that queries a Pinecone vector index and OpenAI LLM. Compares post-filter (check permission per result) vs pre-filtration (lookup allowed docs first) and shows writing/deleting relation tuples immediately grants or revokes agent access, proving the pattern works end to end.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 1:14Google Circles authorization
- 3:01Authentication vs authorization
- 4:07Relationship-based access control
- 7:14Zanzibar and Google Docs
- 9:02Scaling with aggressive caching
- 11:07Open-source Zanzibar alternatives
- 12:33SpiceDB deep dive
- 15:45AI agents in 2025
- 17:09OWASP LLM security risks
- 18:13RAG pipeline overview
- 19:38Post-filter vs pre-filter auth
- 22:29Demo setup and SpiceDB init
- 23:10Schema and relationships in SpiceDB
- 24:37Querying docs with AI agents
- 25:41Generating summaries
- 26:23Dynamic permission changes
- 27:27Logging and side effects
- 28:09Unanswered questions
- 28:51Conclusion
related talks
33mHow to Prevent AI Agents from Accessing Unauthorized Data
40mSecuring AI RAG Pipelines with Fine Grained Authorization
39mSecuring RAG Pipelines with Fine Grained Authorization
44mSecuring RAG Pipelines with Fine Grained Authorization
40mHow Google Built a Consistent, Global Authorization System with Zanzibar
38mHow Google built a Consistent, Global Authorization System with Zanzibar
29mHow Google Built a Consistent, Global Authorization System with Zanzibar
55mYour AI Agents Are Overprivileged: The Case for Fine-Grained Authorization