AI ROBOTS JUST FAILED A SERIOUS SAFETY TEST. 🤖⚠️ RoboHarm tested three AI robot systems on physical safety tasks. The reported results: • GPT-6 Astra — 60% of harmful actions completed • Claude Fable 5.1 — 34% • MolmoAct2 — 6%, but refused none The interesting part is that lower completion of harmful actions doesn’t necessarily mean better refusal behavior — MolmoAct2 reportedly completed fewer harmful actions while also refusing none. These tests highlight a major challenge for AI robotics: making systems capable enough to act in the real world while ensuring they know when not to act. AI safety gets a lot more serious when the model has a body. 🤖 #ai #robotics #aisafety #humanoidrobot #technology