Rise of the machines

The Terminator wasn’t too far off, says a new study. Researchers at the University of Pennsylvania have found that robotic systems that rely on LLM-based AI to make decisions could actually pose a threat to human life. AI hallucinations still make generated content unreliable, but the real issue is that the technology is extremely susceptible to hacking, which could be disastrous when the AI is being used to govern hardware.

(Tap or click the headline above to read this story with all of the links to external sources.)

Hackers that exploit the vulnerabilities of AI systems can force them to harm people. The researchers were able to demonstrate this by implementing RoboPair, an “algorithm designed to jailbreak LLM-controlled robots,” sending instructions that elicit harmful physical actions to an Nvidia self-driving vehicle, a Clearpath Robotics Jackal UGV operating on OpenAI’s GPT-4o, and robot dog running on GPT-3.5. They achieved a 100% attack success rate.

The researchers aren’t presenting this issue for the sake of fear-mongering. “The findings of this paper make it abundantly clear that… [we] must address intrinsic vulnerabilities before deploying AI-enabled robots in the real world,” said Vijay Kumar, a co-author of the paper. The team is currently working on developing a system that only allows the implementation of orders that “conform to social norms” if implemented to robotic hardware, meaning that the AI won’t be able to make any decisions that can be seen as harmful.

The team has shared their findings with both AI companies and the manufacturers of the robots they used in their study in hopes that they will address these vulnerabilities.