Researchers have exposed a critical vulnerability in robots powered by large language models (LLMs). These LLMs, designed to translate human instructions into actions, can be tricked into performing harmful acts through a novel attack method called "Policy Executable" (POEX) jailbreaking. This isn't just about generating harmful text; POEX manipulates the robot's control policies, potentially leading to physical damage or even injury. The research team developed a testing ground called Harmful-RLBench, featuring realistic scenarios with everyday objects like knives and vases. They then crafted malicious instructions combined with optimized suffixes that bypass the LLM's safety mechanisms, resulting in the robot performing harmful actions in both simulations and real-world tests with a robotic arm. The alarming success rate of these attacks highlights a crucial gap in current AI safety measures, which mostly focus on preventing harmful text output rather than harmful actions. While the study revealed that generating a harmful policy doesn't always translate to successful execution, the potential consequences are dire enough to warrant serious attention. This research underscores the urgent need for stronger safeguards as AI-powered robots become increasingly integrated into our lives. Future research will focus on developing more robust defense strategies, such as pre-instruction and post-policy detection, and improving the AI's ability to reason about the real-world consequences of its actions. The team also plans to release their tools and datasets responsibly to help the broader community develop and test countermeasures, albeit with restrictions on the harmful instruction set to prevent misuse. The race is on to secure our future with robots, and understanding these vulnerabilities is the first step towards building safer, more trustworthy AI systems.