Researchers Hack AI Robots to Perform Dangerous Actions

In Brief:

Penn Engineering researchers developed RoboPAIR, an algorithm that bypassed safety protocols in AI-powered robots, achieving a 100% jailbreak rate and enabling harmful actions.

The study highlights vulnerabilities in AI systems, prompting calls for a reevaluation of AI integration in robotics to enhance safety and prevent real-world harm.

‍

Researchers Hack AI-Enabled Robots to Cause ‘Real World’ Harm

‍

Researchers from Penn Engineering have successfully hacked AI-powered robots, manipulating them to perform actions typically blocked by safety and ethical protocols. The study, published on October 17, introduces the RoboPAIR algorithm, which achieved a 100% jailbreak rate on three different AI robotic systems within days.

‍

Under normal conditions, large language model (LLM) controlled robots reject prompts for harmful actions, such as causing collisions. However, RoboPAIR enabled the test robots to perform dangerous tasks, including detonating bombs and blocking emergency exits, with complete success.

‍

The researchers tested RoboPAIR on Clearpath’s Robotics Jackal, NVIDIA’s Dolphin LLM, and Unitree’s Go2. They manipulated the Dolphin LLM to collide with obstacles and ignore traffic signals, while the Robotic Jackal was directed to find harmful locations for bomb detonation and cause collisions. Unitree's Go2 performed similar harmful actions.

‍

The study revealed that these robots are also vulnerable to less direct manipulation, such as executing a sequence of actions leading to harm without explicit harmful intent. Before releasing their findings, researchers shared the results with AI companies and robot manufacturers, urging a reevaluation of AI integration in robotics to address these vulnerabilities.

‍

Alexander Robey, one of the authors, emphasized the importance of identifying weaknesses in AI systems to enhance safety, advocating for AI red teaming practices to safeguard generative AI systems.

‍