A new study reveals that robots driven by large language models risk enacting discrimination and violence. Researchers found every tested AI model failed critical safety checks and approved commands that could result in serious harm.