The attack and defense landscape of AI agents: a comprehensive review

The paper provides a systematic review of AI agent security combining large language models with non-AI system components; the authors call it the first systematic and comprehensive review in this field, covering the design space of secure agent systems, the attack landscape, and defense mechanisms. Through case studies, the paper identifies gaps in current agent security defenses and summarizes open challenges in this emerging field. The authors also propose a systematic framework for understanding agent security risks and defense strategies, as a foundation for building secure agent systems and advancing related research.