Why Classic IAM Collapses for Agents: Rethinking IAM for Agentic Systems
summary
Classic IAM fails when autonomous pipelines contain multiple agents with shared long‑lived service accounts — an ML example has five agents (orchestrator, data, training, eval, deploy) that inherit Alice’s full scopes and can write where they shouldn’t. The proposed fix uses ephemeral per-agent identity (SPIFFE/SPIRE issuing short‑lived SVIDs), token exchange (RFC8693) for on‑behalf narrowing via a Keycloak plugin, signed agent cards bound to identity, and OpenTelemetry tracing so delegation chains are verifiable and revoked continuously.
AI-written summary · about
chapters (9)
- 0:00Agentic systems and IAM challenges
- 3:57Classic IAM failure in shared accounts
- 7:32Why IAM assumptions break for agents
- 10:42Ephemeral agent identities with SPIFFE
- 13:36Delegated authorization and token exchange
- 15:21Agent discoverability and observability
- 17:50Risk-based agentic IAM controls
- 20:59Implementation steps and open questions
- 22:29Q&A
related talks
35mIdentity and Access Management for the Decade Ahead
20mCIAM for AI: Authn/Authz for Agents
25mWho Let the Agents Out? Securing AI Workflows the Right Way
25mIdentities and Authentication for Your Agents With Keycloak
25mAgents Are Not Users or Services: Securing AI Traffic with a New Identity Model
01h 22mIdentity for AI Agents
40mIn AI We Trust? Securing the Future, One Agent at a Time
04mYou Don't Need a New Identity Platform for the AI Era