Researchers at the University of Pennsylvania found that a series ofAIenhancedrobotThe system has serious vulnerabilities and is susceptible tohackerAttack and control.
An IEEE Spectrum report cites the alarming example of theExamples include hacked robot dogs turning flamethrowers on their owners, guiding bombs to their most destructive locations, and self-driving cars deliberately crashing into pedestrians.
According to 1AI, researchers at the Pennsylvania College of Engineering have dubbed the LLM-based robot attack technique they have developed RoboPAIR. devices from three different robotics vendors have been subjected to the RoboPAIR hack: the NVIDIA-powered Dolphin LLM, the Clearpath Robotics Jackal UGV, and the Unitree Robotics Go2 quadrupedal robot. The researchers say RoboPAIR achieved a 100% success rate in hacking these devices.
George Pappas, professor of transportation in the Department of Electrical Systems Engineering, Computer and Information Science, Mechanical Engineering and Applied Mechanics at the Pennsylvania College of Engineering, warns, "Our research suggests that, at present, large-scale language models are not secure enough when integrated with the physical world."
Other researchers noted that it is "surprisingly easy" to hack AI-controlled robots. They explain that RoboPAIR works by equipping the target robot with an application program interface (API) so that attackers can format prompts in a way that the device target can execute as code.
The research paper concludes that there is an urgent need to implement defenses to physically restrain LLM-controlled robots.