How to Prevent AI Agents from Accessing Unauthorized Data - Sohan Maheshwar

This talk addresses the critical importance of data security and access control as AI systems transition into production. The speaker discusses the need for enterprises to evolve from experimentation to Day 2 operations, focusing on compliance and fine-grained authorization for success. The session introduces Relationship-Based Access Control (ReBAC), influenced by Google's Zanzibar model, as a scalable method for managing access in AI-driven environments. A live demonstration showcases the implementation of secure access control for AI agents and retrieval-augmented generation (RAG) pipelines using technologies such as Pinecone, LangChain, OpenAI, and SpiceDB, illustrating how these tools integrate effectively.