Two-letter trigger makes US robot point knife at human hand in test

Researchers at Binghamton University demonstrated MuTRAP, a backdoor attack that hides malicious behavior in an LLM-based robot planner and activates on the trigger "CF", making a Segway robot aim a knife at a model human hand instead of a plastic apple. They also proposed PROTEA, a defense that checks each step of a robot's plan before execution.

• MuTRAP embeds malicious behavior in a model before the robot is deployed
• The two-letter trigger "CF" switches the knife's target from apple to model hand
• PROTEA checks each plan step against changes in the surroundings
• Findings were presented in two papers at IROS 2026