Choose timezone
Your profile timezone:
Description
As AI systems increasingly operate as autonomous agents capable of controlling browsers, operating systems, and other tools, they introduce a fundamentally new security paradigm. This talk examines the emerging vulnerability landscape of computer-use AI agents, including prompt injection across tool boundaries, indirect control-flow manipulation, environment-based attacks, privilege escalation via delegated actions, and failures of intent alignment under adversarial conditions. We will discuss concrete attack classes observed in real systems, why traditional sandboxing and policy approaches are insufficient, and outline research directions for securing agentic AI in practice.
Videoconference via https://us02web.zoom.us/j/82479747034
