New research from UC Riverside found computer-use AI agents often push ahead with unsafe or irrational tasks, raising questions about whether today’s desktop agents are ready for sensitive everyday workflows.