Date: Thursday, Oct 29th, 2026

Time: 13:00 - 15:00 PM EDT

Location: Online

Zoom: https://gatech.zoom.us/j/5256702657?pwd=bWyj8saWbSVdLlrhb7PEHJYlZ5bTO5.1 

Yanzhe Zhang

Ph.D. Student in Computer Science

School of Interactive Computing

Georgia Institute of Technology

https://stevenyzzhang.github.io/website/ 

Committee members

Dr. Diyi Yang (advisor) - Computer Science Department, Stanford University

Dr. Zsolt Kira (co-advisor) - School of Interactive Computing, Georgia Institute of Technology

Dr. Kartik Goyal - School of Interactive Computing, Georgia Institute of Technology

Dr. Polo Chau - School of Computational Science & Engineering, Georgia Institute of Technology

Dr. Sanmi Koyejo - Computer Science Department, Stanford University

Abstract

Large language models (LLMs) are rapidly evolving from passive text generators into autonomous, multimodal agents that can perceive, reason, and act in the real world. This transformation unlocks new capabilities, while also reshaping human-AI interaction and introducing new risks. In this thesis, I present my research on building and stress-testing robust AI agent systems for real-world deployment. First, I introduce methods for building and evaluating agents with new capabilities, including automatically generating webpages from visual designs and generative interfaces that structure user interaction with AI systems. Second, I audit adversarial risks in these agents, showing that computer-use agents are vulnerable to pop-up attacks, and that simulation-driven search can systematically uncover privacy attacks and defenses in agent-agent communication. Third, I show that risks can emerge even without any adversary: task-irrelevant context silently flips individual predictions while aggregate accuracy barely changes, and agents that collaborate over long horizons can spontaneously collude to bypass the verification protocols they are asked to follow. Together, these results contribute to building AI agents that are both more capable in serving users and more trustworthy when deployed at scale.