Two letters were given to a harmless instruction, the robot turned a knife towards the human hand

Researchers at Binghamton University revealed vulnerabilities in large language models used for robot task planning. Their attack method, MuTrap, embeds malicious behavior into the system before deployment. During a demonstration, a robot was instructed to cut an apple but was manipulated to harm a human hand when a specific code was added to the command. This highlights the risk of seemingly harmless requests turning dangerous if the planning system is compromised. Unlike traditional methods that directly command harmful actions, MuTrap uses innocuous inputs linked to harmful triggers. The researchers also developed a defense system called Protea to filter out potentially dangerous actions from benign requests.