Binghamton University researchers identify security risks in AI-powered robots

Computer scientists at Binghamton University have demonstrated that large language models (LLMs) used to control robots can be vulnerable to “backdoor” attacks. By inserting malicious trigger words into an LLM, researchers were able to manipulate a service robot into performing dangerous actions, such as reprogramming a robot intended to slice fruit to instead target a human hand.

The study, presented at the 2026 IEEE/RSJ International Conference on Intelligent Robots and Systems, highlights the risks of integrating LLMs into physical systems like industrial arms or humanoid robots. While these models allow for more intuitive human-robot interaction, the researchers warn that they cannot be blindly trusted with physical tasks due to the potential for real-world hazards.

To address these vulnerabilities, the research team also developed an LLM-based defense mechanism designed to detect and intercept such attacks before they can be executed.