A new survey paper categorizes the safety and security threats posed by Computer-Using Agents (CUAs). These agents, powered by LLMs, can autonomously interact with software and interfaces, presenting novel risks. The paper outlines a taxonomy of these threats and proposes defensive strategies, aiming to guide future research and practical deployment of secure CUAs. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
IMPACT Provides a structured foundation for understanding and mitigating novel safety and security risks in autonomous AI agents.
RANK_REASON This is a survey paper on safety and security threats of AI agents.