Security Intelligence RSA recap, the LiteLLM breach, and the quest to fix AI agent security
Apr 1, 2026
Jeff Crume, IBM Distinguished Engineer and AI security expert; Dave McGinnis, cyber threat leader; Suja Viswesan, IBM security products VP; and Jake Lundberg, HashiCorp Field CTO. They unpack agentic AI identity and lifecycle security, debate why human IAM fails for agents, examine the LiteLLM supply-chain compromise, and recap RSAC trends like narrow-domain agents and autonomous defense.
AI Snips
Chapters
Transcript
Episode notes
Agentic Identities Are A Different Class
- Agentic AI identities are fundamentally different from traditional human and non-human identities and require new patterns for scope and timing.
- Jake Lundberg warns agents can inherit wide-scoped access from long-lived unmanaged identities, creating unexpected privilege exposure across roles.
Attestation And Audit Are As Important As Authentication
- The problem is twofold: unclear authorization scope and weak attestation/audit for agent-initiated actions.
- Jake explains you must attest the human kickoff and then restrict what the spawned agent can do and audit differences between request and action.
Remediate Unmanaged Secrets And Move To Just In Time
- Start by finding unmanaged credentials, rotate long-lived secrets, and move toward just-in-time session-based credentials.
- Jake recommends vaulting secrets (Vault, Verify) and shifting to session-scoped credentials that are removed when the session ends.