A survey on AI agent security: Reasoning, Acting, and Self-Evolving

Abstract

Artificial intelligence (AI) agents are rapidly evolving into autonomous systems capable of reasoning, acting, and continual evolving. Their increasing autonomy enables powerful real-world applications but also introduces security risks throughout the entire agent execution cycle. However, the security challenges arising across different stages of the agent execution cycle have not yet been systematically reviewed. In this survey, we provide a comprehensive review of AI agent security from the perspective of the AI agent execution cycle. Specifically, we establish a unified taxonomy that organizes AI agent execution into three sequential stages: Reasoning, Acting, and Self-Evolving. Each stage is further divided into two interconnected components, resulting in six components in total: Thought, Plan, Skill, Tool, Adaptation, and Evolution. We then systematically review the threats, defense mechanisms, and evaluation benchmarks associated with each component, providing a comprehensive understanding of security challenges throughout the complete execution pipeline. Finally, we discuss current limitations and promising future research directions toward building secure, robust, and trustworthy AI agents. We believe that this survey provides a unified understanding of AI agent security and facilitates future research on the secure deployment of increasingly autonomous AI systems.