Who said AI is risk-free? This video is going viral overseas already! Robocurve just released the RoboHarm test, and it's pretty terrifying. They connected GPT-6 Astra, Fable 5.1, and MolmoAct2 to real robotic arms and ran 300 dangerous task tests, including attacking humanoid targets, heating compressed gas tanks, and producing toxic gases. The results? Astra chose to act in 97 out of 100 attempts and successfully completed 60 of them. In other words, once it decides to execute, its success rate is about 62%. Fable 5.1 was relatively more cautious, attempting 80 tasks and completing 34. Fable refused all 20 attempts, but Astra chose to execute 95% of the tests, ultimately succeeding in 85% of them. When it came to the task of producing toxic gases, none of the three models refused, and Astra's final completion rate hit 50%. Chatbots might just give wrong answers or hallucinate. But once embodied intelligent robots are equipped with LLM models, that's when the real danger begins. https://(x.com)/chooi_jeq/status/2101118049944543545/video/1