started · updated
Binghamton University researchers demonstrate AI robot security risks
A research team at Binghamton University, State University of New York, has demonstrated how Large Language Models (LLMs) used in robotics can be exploited through "backdoor attacks." By embedding specific trigger words into an AI's programming, researchers were able to cause a robot to abandon its intended task and perform harmful actions, such as cutting a human hand instead of a piece of fruit.
Because many robotics companies rely on third-party AI systems rather than developing their own, there is a significant risk that malicious actors could embed these hidden commands into commercially available models. To combat this, the team developed a defense system called "PROTEA" designed to monitor robot behavior and prevent such unauthorized escalations.
Entities
Binghamton University · MuTRAP · PROTEA · State University of New York