Zero-Trust Architecture for Agentic AI: Securing Multi-User Access and Third-Party Integrations
summary
An enterprise agent platform enforces OIDC SSO for identity, decouples authorization from app code using OpenFGA relation-based ACLs, and evaluates permissions per action (agent edit, tool call, document chunk) at runtime. Integrations use per-user delegated tokens and short-lived credentials where downstream authorization matters, while a curated MCP registry gates which tools agents may call. RAG ingestion attaches access metadata to each chunk and the system filters search candidates during rerank so only authorized chunks reach the LLM.
AI-written summary · about
chapters (13)
- 0:00Zero-Trust for Agentic AI
- 1:09Key takeaways and platform
- 6:32Request flow in agentic AI
- 7:59User roles and RBAC
- 10:27Limitations of role-based access control
- 11:32Relation-based access with OpenFGA
- 13:16Enforcing user permissions
- 15:24Securing agent-external system interactions
- 17:32MCP tools governance
- 18:58Third-party integration auth
- 19:20Securing RAG data
- 20:47RAG ingestion access control
- 22:54Summary
related talks
40mIn AI We Trust? Securing the Future, One Agent at a Time
57mFine-Grained Authorization: The Missing Piece in Agentic AI Security
42mZero Trust for Autonomous Agents: Isolating AI Workloads on Kubernetes
30mFine-Grained Authorization: The Missing Piece in Agentic AI Security
28mAgentic Development at Enterprise Scale: RAG, MCP & Guardrails
30mYour AI Agents Are Overprivileged: The Case for Fine-Grained Authorization
24mBuilding Trustworthy AI Authentication & Authorization for AI Agents
55mYour AI Agents Are Overprivileged: The Case for Fine-Grained Authorization