A hands-on session dissecting how AI agents actually fail from a security standpoint — not abstract AI-safety talk, but the specific, well-documented failure modes teams are running into today: prompt injection through tool outputs, excessive agency from over-permissioned tool access, memory/context poisoning, and insecure handling of agent outputs. The first half is a live autopsy of a deliberately compromised demo agent, where attendees trace exactly how it was exploited and what let the attack in. The second half turns adversarial: attendees pair up into red teams and blue teams inside an isolated sandbox — no real systems or data involved — with red teams trying to manipulate the agent into leaking data or calling tools outside its scope, and blue teams patching the vulnerability live. Attendees leave with a working checklist for auditing their own agent stack.