AI agents are no longer experimental. They browse the web, execute code, and take actions across enterprise systems with minimal human oversight. That creates threats traditional security models don't cover. This session walks through the ones that matter: prompt injection, goal manipulation, tool misuse, and what happens when an agent starts acting on its own. We use the OWASP Agentic AI Threats and Mitigations as the framework and show where the EU AI Act and the Cyber Resilience Act apply. You'll know what goes wrong when AI acts independently and which controls actually matter. The workshop that follows lets you attack these systems hands-on.