“Contextual Security for the Next Generation of AI Agents“
Location: 177 Huntington Ave, conference room 503
Abstract: As language model capabilities grow, enabling highly autonomous AI agents, and adversaries find new attacks, our community should reinvent security paradigms for upcoming agentic systems. In this talk, I propose dynamic policy generation based on the theory of Contextual Integrity that introduces a concept of appropriate information flows. Our approach dynamically restricts an agent to select actions and parts of the user information that are only necessary in the current context and adjusts them as the context changes. In conclusion, I will discuss the importance of addressing ambiguous scenarios, and outline how contextual policies could be developed for multi-agent systems.
Bio: Eugene Bagdasarian is an Assistant Professor at University of Massachusetts Amherst and a Researcher at Google. His work focuses on studying attack vectors in AI systems deployed in real life and proposing new designs that mitigate these attacks. Previously, he received the Distinguished Paper Award at USENIX Security and Apple AI/ML PhD Fellowship.