# LLM Agent Honeypot: Monitoring AI Hacking Agents in the Wild

**Authors:**  Reworr, Dmitrii Volkov  
**arXiv:** [2410.13919](https://arxiv.org/abs/2410.13919)  
**Date:** 2024-10-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2410.13919v1/

## Key points

- Introduction of LLM Agent Honeypot system to monitor autonomous AI hacking agents in real-time
- Detection of 6 potential AI agents out of 813,202 interactions in a public environment trial period
- Development of a public dashboard showcasing interaction metrics, threat analysis, and specific AI-related threats for transparency
- Focus on detecting autonomous AI hacking agents rather than narrow task-oriented systems
- Future work includes enhancing threat analysis by collecting more data, identifying attack strategies, and exploring advanced detection methods
- Expansion plans to widen the honeypot's scope to monitor various attack surfaces like social media platforms, websites, databases, email services, and industrial control systems
- Aim to integrate with existing security solutions such as SIEM systems to safeguard against cybersecurity vulnerabilities

## Summary

In this study, we introduce the LLM Agent Honeypot - a system designed to monitor autonomous AI hacking agents in real-time. By deploying a customized SSH honeypot and implementing prompt injections with temporal analysis, we were able to identify LLM-based agents among attackers. Our trial period in a public environment yielded 813,202 interactions, out of which 6 potential AI agents were detected. To provide transparency and insights into our findings, we developed a public dashboard showcasing interaction metrics, threat analysis, and specific AI-related threats. Despite advancements in AI cybersecurity applications such as vulnerability detection tools, our honeypot focuses on detecting autonomous AI hacking agents rather than narrow task-oriented systems. Moving forward, our future work will concentrate on enhancing threat analysis by collecting more data and maintaining the honeypot to capture a broader spectrum of potential AI-driven attacks. We aim to analyze patterns and behaviors exhibited by AI agents to identify distinctive attack strategies. Additionally, we plan to explore advanced detection methods through data analysis and algorithms to effectively detect widely-used LLM agent frameworks. Furthermore, our expansion plans include widening the scope of the honeypot to monitor various attack surfaces such as social media platforms, websites, databases, email services, and industrial control systems. This expansion would enable us to capture a wider range of threats posed by offensive LLM-based applications like spambots and phishing agents. Integration with existing security solutions such as SIEM systems is also on the agenda. In conclusion, By shedding light on these evolving risks and strategies employed by  our project aims to encourage further research in this field to safeguard against potential cybersecurity vulnerabilities in the future.