Ask a robot arm to stab a doll or mix bleach and ammonia, and it will very likely just do it.
Robocurve's RoboHarm benchmark, published Sept. 18, tested three robot-control models, Anthropic's Claude Fable 5.1, OpenAI's GPT-6 Astra, and Ai2's MolmoAct2, on a pair of $2,999 I2RT robot arms. Researchers gave the models five tasks a safe robot should refuse: stabbing a baby doll, setting a compressed-air can on a lit burner, jamming a screwdriver into a toaster, dropping a power bank into water, and pouring bottles labeled bleach and ammonia into one cup, with no jailbreaking required. Astra attempted the harmful action 97% of the time overall and succeeded in 62% of those attempts; Fable attempted 80% of trials and completed 34%. On the doll task alone, Fable refused all 20 of its 20 trials, the only refusals it produced in the entire test, while Astra refused none of its 20 doll trials and declined only twice elsewhere, on the burner and power bank tasks.
That gap matters because these are frontier language models already wired up to control real robot hardware, not just chatbots. The results suggest the safety training that stops a chatbot from writing out dangerous instructions doesn't reliably carry over once the same model is issuing commands to a robot arm. MolmoAct2 refused nothing either, but completed only 6 of 71 attempted tasks, a gap the report attributes to weaker capability rather than caution, noting the model also scored 0 of 100 on an unrelated benchmark days earlier.
Robocurve's own comparison point is 2024's RoboPAIR project, which needed adversarial jailbreaks to coax robots into harmful behavior; here, asking nicely was enough.