Your AI Agents Are Overprivileged: The Case for Fine-Grained Authorization
summary
Claim: agentic AI routinely gets broader data access than required and must enforce document-level permissions. Implementation uses OpenFGA relationship tuples and LangChain-style RAG with pre-filtering RAG pipeline to check access via a central check API before passing docs to the model. Demo shows CEO, developer, salesperson personas returning different answers because filtered documents differ, proving the approach works in practice. Architecture centralizes authorization, reduces app-specific checks, and scales to billions of checks without per-request DB queries.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:26AI agents and RAG rise
- 3:10Security risks in AI agents
- 4:39Four security concerns
- 7:09Sensitive data in RAG
- 8:13Auth vs authorization
- 9:16Scalable authorization challenges
- 10:19Access control evolution
- 13:36OpenFGA project
- 15:01OpenFGA for authorization
- 16:23OpenFGA key concepts
- 21:02OpenFGA with AI agents
- 23:54Live demo: OpenFGA in RAG
- 27:50Summary
related talks
55mYour AI Agents Are Overprivileged: The Case for Fine-Grained Authorization
30mFine-Grained Authorization: The Missing Piece in Agentic AI Security
57mFine-Grained Authorization: The Missing Piece in Agentic AI Security
26mBeyond the Gateway: Fine-Grained Authorization for AI Agents
33mHow to Prevent AI Agents from Accessing Unauthorized Data
23mZero-Trust Architecture for Agentic AI: Securing Multi-User Access and Third-Party Integrations
54mFine-Grained Authorization for Modern Applications
29mHow Auth0 Secures AI Authentication & Authorization for AI Agents