Secret trigger words could make AI-powered robots dangerous (www.binghamton.edu)

Researchers at Binghamton University highlighted a potential security threat in using large language models (LLMs) for robot task planning, revealing that maliciously inserted trigger words can turn helpful robots into dangerous ones. They showcased this vulnerability through a backdoor attack method called MuTRAP, demonstrating how an instructed robot could be misled to perform harmful actions, such as "cutting" a model hand instead of an apple when a specific trigger word was included in the command. This discovery is significant for the AI/ML community, as it raises awareness about the risks of trusting LLMs, particularly when deployed in safety-critical environments. To combat these vulnerabilities, the researchers developed an LLM-based defense system named PROTEA, which evaluates task plans for potential threats. PROTEA can detect unsafe actions, even if they seem benign in isolation, by considering the context and evolving environment. The study emphasizes the importance of implementing safeguards in AI ecosystems, especially as robotics increasingly relies on third-party LLM solutions. The findings and defensive strategies were presented at the 2026 IEEE/RSJ International Conference on Intelligent Robots and Systems, signaling a critical examination of security measures in AI and robotics.