What is it about?
As robots become increasingly connected and intelligent, they also face new cybersecurity threats. Attackers can now use artificial intelligence (AI) agents to automatically explore robotic systems, identify weaknesses, and carry out cyberattacks with limited human involvement. Our research introduces ROS2Pot, an AI-powered cybersecurity honeypot designed to protect robotic systems using ROS 2, a widely adopted software framework for modern robots. A honeypot is a realistic decoy system that attracts attackers, allowing researchers to observe their behaviour without exposing real operational systems. Unlike traditional honeypots, ROS2Pot combines realistic robot communication, physics-based simulation, and an AI agent that actively interacts with suspicious visitors. It can help distinguish AI-driven attackers from other users and uses carefully designed deception techniques to encourage attacking AI agents to reveal their objectives. We evaluated ROS2Pot against five AI models performing different attack tasks. The results demonstrate that the system can convincingly imitate a robotic environment and identify or extract information from certain AI-driven attacks, although effectiveness varies across models and attack objectives. This research explores how AI can be used not only to launch cyberattacks but also to develop more proactive and intelligent cybersecurity defenses for robotic systems.
Featured Image
Photo by Andy Kelly on Unsplash
Why is it important?
Robotic systems are increasingly used in manufacturing, transportation, healthcare, and other critical environments. As AI-powered cyberattacks become more autonomous, conventional security monitoring may struggle to understand and respond to these emerging threats. ROS2Pot introduces a new agent-versus-agent defense approach, where an AI-powered honeypot actively engages attacking AI agents rather than simply recording suspicious activity. The research makes three important contributions: 1. Realistic robotic deception: Creates an interactive ROS 2/DDS honeypot that closely reproduces genuine robotic communication and simulated physical behaviour. 2. AI-aware attacker detection: Introduces techniques to identify LLM-driven attackers by exploiting characteristics of their automated interactions. 3. Proactive threat intelligence: Demonstrates how carefully designed interactions can encourage certain attacking AI agents to disclose their operational objectives. By evaluating five AI-driven attacker models, the study provides early experimental evidence for proactive defense against autonomous cyberattacks. The findings could support future research on intelligent intrusion detection, robotic cybersecurity, and adaptive deception systems for autonomous robots and industrial environments.
Perspectives
My research interests lie at the intersection of artificial intelligence, cybersecurity, and intelligent networked systems. As LLM-based agents become increasingly capable of performing complex tasks autonomously, I believe cybersecurity research must move beyond traditional human-centred threat models. Through this work, we explored an important question: Can defensive AI agents exploit the weaknesses of attacking AI agents to protect robotic systems proactively? ROS2Pot represents an initial step toward this direction by combining robotic system simulation, AI-driven attacker profiling, and adaptive deception. I am particularly interested in extending this research toward trustworthy agentic AI, autonomous cyber defense, and intelligent security operations for robotic and cyber-physical systems. I hope this work encourages further investigation into AI-versus-AI security and the development of more resilient autonomous systems.
Jing Li
Xiamen University Malaysia
Read the Original
This page is a summary of: POSTER: ROS2Pot: An Agentic Honeypot for Proactive Defense Against LLM Agents in ROS 2/DDS Systems, November 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3830454.3846411.
You can read the full text:
Contributors
The following have contributed to this page
