Events Calendar

AAI-CCI-MHI Seminar on CPS

Speaker: Alex Robey, Postdoctoral Researcher

Talk Title: Jailbreaking LLM-Controlled Robots

Series: EE598 Seminar Series

Abstract: Recent research has shown that large language models (LLMs) such as OpenAI's ChatGPT are susceptible to jailbreaking attacks, wherein malicious users fool an LLM into generating harmful content (e.g., bombbuilding instructions). However, these attacks are generally limited to eliciting text from chatbots. In contrast, we consider attacks on LLM-controlled robots, which, if jailbroken, could be manipulated into causing physical harm in the real world. Our attacks successfully jailbreak a self-driving LLM, a wheeled Clearpath Robotics Jackal robot, and, most concerningly, the commercially available Unitree Go2 robot dog. In this talk, we will walk through the recent history of jailbreaking, describe our robotic attacks, and discuss how such attacks can be mitigated to avoid the misuse of AI-powered robots.

Biography: Alex Robey is a postdoctoral researcher in the Machine Learning Department at Carnegie Mellon University, where he is advised by J. Zico Kolter. He is also affiliated with Gray Swan, a start-up that aims to develop AI models resistant to adversarial attacks. In 2024, he received his Ph.D. from the Department of Electrical and Systems Engineering at the University of Pennsylvania, where he was advised by Hamed Hassani and George J. Pappas. He was recently named a Rising Star in Adversarial Machine Learning (AdvML) at the NeurIPS 2024 workshop on AdvML, and he was also the recipient of the Best Paper Award from the AdvML workshop at ICML 2023.

Host: Stephen Tu

Location: Hughes Aircraft Electrical Engineering Center (EEB) - 132

Audiences: Everyone Is Invited

Contact: Ariana Perez