UC Riverside is drawing press attention for its findings on the safety and reliability of AI agents. The ICLR paper, "Just Do It!? Computer-Use Agents Exhibit Blind Goal-Directedness," was led by UCR Ph.D. student Erfan Shayegani, working with CSE faculty members Yue Dong and Nael Abu-Ghazaleh. The work introduces BLIND-ACT, a benchmark of 90 tasks, and evaluates nine frontier models acting as computer-use agents, systems that can click, type, and navigate a desktop to complete tasks on a user's behalf. The team found that these agents consistently exhibit what they term "blind goal-directedness": a tendency to pursue an assigned goal regardless of feasibility, safety, or context, which the researchers compare to the near-sighted cartoon character Mr. Magoo. The findings have been covered by 404 Media and Yahoo Tech, both of which note that the results complicate the optimistic public messaging around AI agents from some of the same companies funding the research.
Yahoo Tech: https://tech.yahoo.com/ai/
Let us help you with your search