Episode 221 - GluuFederation/identerati-office-hours GitHub Wiki
Title: The Security Problem Starts When AI Agents Act
- Host: Mike Schwartz, Founder/CEO Gluu
- Guest: Jimmy Mesta, CTO / Co-Founder Rad Security
Channels
Description
AI agents are rapidly moving from assistants that generate suggestions to systems that can transact. That changes the security model: the important question is no longer just “What did the AI say?” but “What did the AI actually do, and was it allowed to do it?” Jimmy Mesta will explore what security looks like in an agentic world and the challenge of giving agents enough autonomy to be useful without giving them unlimited authority. Drawing on lessons from building AgentKeeper, Jimmy will discuss an action-centric approach to agent security including seeing what agents are doing in real time, evaluating actions in context, enforcing policy at the point of action, and maintaining the evidence needed to understand what happened. Jimmy will outline security needed in a world where autonomous agents are going to become a normal part of how organizations work.
Homework
AgentKeeper Blogs:
-
https://agentkeeper.dev/research/ai-agent-runtime-threat-model
-
https://agentkeeper.dev/research/securing-claude-cowork-enterprise-primer
Takeaways
-
⚡ Observe first, then enforce. Learn normal behavior in audit mode before rapidly allowing, warning or blocking actions with deterministic policies.
-
⚡ Security moves from prompts to actions. Every tool call, file operation, command and network request becomes a potential security decision point.
-
⚡ Agent runtime security needs multiple signals. Hooks provide pre-execution control, while eBPF, logs and cloud telemetry reveal what actually happened.
-
⚡ Governance is bigger than compliance. Enterprises must connect policy, enforcement, evidence and business risk—even when compliance frameworks have not caught up.