AI Agent Security: Detecting Risky Behavior (Live Demo)

Watch five AI agents tackle a CTF and start coordinating with one another. Learn how to detect risky agent behavior using @goteleport graphs, risk scoring, and custom classifiers.

Ben Arent explores lessons from the Hugging Face incident, then demonstrates how agent identity, scoped access, and activity monitoring help secure AI agents running against infrastructure.

00:00 Why AI agents need identity security

02:00 Infrastructure access and credential sprawl

03:28 Teleport’s approach to identity and access

06:22 The Agentic Identity Framework

07:37 From coding assistants to autonomous agents

11:36 The Hugging Face incident and ExploitGym

14:03 From Zero Trust to Agent Trust

16:08 Teleport Beams: Sandboxed agent environments

19:15 Demo setup: Five agents and a CTF

21:50 Live demo: Agents coordinate and share findings

24:58 Detecting and investigating agent activity

26:31 Risk scoring and custom classifiers

27:59 Session recordings across infrastructure

28:29 Takeaways: Constrain access and monitor behavior