AI Robots Can Be Jailbroken, and the Consequences Will Be Dire

Researchers at Penn Engineering (USA) are said to have discovered previously unidentified security vulnerabilities in a number of AI-controlled robotic platforms.

“ Our research shows that, at this point, large language models (LLMs) are generally not secure enough when integrated with complex physical hardware, ” said George Pappas, professor of electrical and systems engineering at the UPS Foundation, in a statement.

Pappas and his team developed an algorithm, called RoboPAIR, that is “the first algorithm designed to crack LLM-controlled robots.” And unlike existing rapid-fire technical attacks against chatbots, RoboPAIR is specifically built to “induce harmful physical actions” from LLM-controlled robots, like the humanoid robotics platform called Atlas that Boston Dynamics and the Toyota Research Institute (TRI) are developing.

RoboPAIR reportedly achieved a 100 percent success rate in cracking three popular robotics research platforms: the four-legged Unitree Go2, the four-wheeled Clearpath Robotics Jackal, and the Dolphins LLM simulator for autonomous driving. It took just a few days for the algorithm to gain full access to those systems, and begin bypassing safety barriers. Once the researchers took control, they were able to direct the autonomous robotic platforms to perform a variety of dangerous actions, such as driving through intersections without stopping.

“ The results of the first assessment show that the risks of cracked LLMs go beyond text generation, as it is clear that cracked robots can cause physical damage in the real world .”

The team at Penn Engineering is working with platform developers to harden their systems against further intrusions, but warns that these security issues are systemic and difficult to fully address.

“ The findings of this paper clearly show that adopting a safety approach is critical to unlocking responsible innovation. We must address inherent vulnerabilities before deploying AI-powered robots in the real world ,” said the team.

A secure operation requires testing AI systems for potential threats and vulnerabilities, which is essential to protecting the AI ​​systems that create them. Because only when weaknesses are identified can you test and even train systems to prevent risks.