Computer-use agents are increasingly deployed in open-ended digital environments — interacting with the web and executing multi-step tasks across applications. Yet real-world deployment exposes a fundamental tension: agents must continually adapt to unseen scenarios and distribution shifts, while their safety and security grow more critical as their capabilities increase. In this talk, I will discuss our recent efforts toward continual learning in computer-use agents for environment adaptation, systematic evaluation of adversarial risks, and proactive discovery of unintended behaviors in the absence of attacks. Together, these efforts suggest that capability and safety need not be in tension, and point toward agents that are both capable and safe in deployment.
Huan Sun is an endowed College of Engineering Innovation Scholar and Associate Professor in the Department of Computer Science and Engineering at The Ohio State University. Her research focuses on advancing both the capability and safety of LLM-based agents, with emphasis on web agents, computer-use agents, and agents for data-driven scientific discovery. Huan has led or co-led a series of foundational projects that have helped shape the current landscape of AI agents, with broad adoption across academia and industry. These include Mind2Web, the first benchmark and a pioneering effort in building LLM-based generalist web agents, and SeeAct, the first multimodal planning-grounding framework for web agents. She is also deeply committed to the safety and security of increasingly capable agents, as exemplified by projects such as EIA and RedTeamCUA. Huan received the NSF CAREER Award and multiple faculty research awards from industry. Her work has been recognized with paper awards at leading venues including CVPR, ACL, and SIGMOD.