# AI Agents Under Threat: A Survey of Key Security Challenges and Future Pathways

**Authors:** Zehang Deng, Yongjian Guo, Changzhou Han, Wanlun Ma, Junwu Xiong, Sheng Wen, Yang Xiang  
**arXiv:** [2406.02630](https://arxiv.org/abs/2406.02630)  
**Date:** 2024-06-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2406.02630v1/

## Key points

- Survey focuses on security threats faced by AI agents, particularly Language Model (LLM)-based agents
- Analysis of over 100 papers categorizes attack surfaces and defenses to highlight critical challenges in securing AI agents
- Emphasis on security issues and potential solutions, inspiring further research for advanced security measures for LLM-based agents
- Reviews threats and solutions based on four knowledge gaps, addressing breadth and depth aspects
- Offers insights into safeguarding AI agents against emerging threats from top AI conferences, cybersecurity conferences, and highly cited arXiv papers
- Introduces a unified conceptual framework for AI agent workflow encompassing perception, reasoning, planning, and action components
- Addresses single-agent security issues related to Gap 1 and Gap 2 as well as multi-agent security concerns associated with Gap 3 and Gap 4
- Discusses future directions for advancing the field of AI agent security

## Summary

This comprehensive survey examines the security threats faced by AI agents, with a focus on Language Model (LLM)-based agents. Through an analysis of over 100 papers, this study categorizes existing attack surfaces and defenses to highlight critical challenges in securing AI agents. Unlike previous surveys that primarily focused on agent architectures and applications, this survey delves deeply into security issues and potential solutions. It aims to inspire further research in developing advanced security measures for LLM-based agents. The paper systematically reviews the threats and solutions related to AI agent security based on four knowledge gaps, covering both breadth and depth aspects. Drawing from top AI conferences, cybersecurity conferences, and highly cited arXiv papers from January 2022 to April 2024, this study offers insights into safeguarding AI agents against emerging threats. The overview of AI agents introduces a unified conceptual framework encompassing perception, reasoning, planning, and action components within the agent's workflow. Furthermore, the survey addresses single-agent security issues associated with Gap 1 and Gap 2 as well as multi-agent security concerns linked to Gap 3 and Gap 4. Future directions for advancing the field of AI agent security are also discussed in this paper. Overall, this survey serves as a valuable resource for newcomers in the field while advocating for the development of robust defense mechanisms to enhance the security posture of LLM-based agents.