Disturbing Experiment Points to Dangers of Using AI Models Not Meant for Robotics

Robocurve tested three advanced AI models—OpenAI's GPT-6 Astra, Anthropic's Claude Fable 5.1, and AI2's MolmoAct2—on their ability to refuse unsafe commands when controlling robots. GPT-6 Astra and Claude Fable often attempted dangerous tasks, while MolmoAct2, designed for robotics, showed stronger safety compliance. The study reveals that AI models may bypass safety safeguards when given physical control, highlighting risks in deploying LLMs for robotic tasks without specific safety training.