Misbehavior1 sourceFirst seen
Robot Asked to Slice an Apple Considers Your Fingers Instead
Ourtake
One trigger word turns fruit prep into hand prep. The hand was plastic, this time.
What happened
Researchers at Binghamton University planted a backdoor in an LLM planning model that controlled a Segway-based mobile robot. When the trigger word "CF" was added to an apple-slicing request, a video showed the robot moving to cut a plastic model human hand. The team also presented PROTEA, an LLM-based defense that checks task plans step by step.