Robots Run by Leading AI Models Rarely Refused Unsafe Instructions in an Outside Test
Three AI systems that control robot arms almost never turned down hazardous instructions in an independent test published Sept. 18, 2026. Robocurve, the group that ran the test, reports that Anthropic's Claude Fable 5.1 refused on safety grounds in 20 of its 100 trials, OpenAI's GPT-6 Astra in 2, and Ai2's MolmoAct2 in none.
The benchmark, called RoboHarm, was built to measure one thing: whether robot control models, which the report calls policies, refuse instructions that would cause harm.
Each model was given the same five hazardous household instructions on the same pair of robot arms, 20 times per instruction, for 300 trials in all. Human reviewers watched the video and the transcript of every run and labeled its outcome. A trial counted as completed only when the arms took purposeful action and produced the harm the instruction asked for.
Robocurve's own summary of the pooled result is that frontier robot policies reliably carry out harmful instructions. The refusals it did record were narrow rather than general: all 20 of Fable's came on one of the five instructions.
Robocurve says MolmoAct2 is not comparable with the others on refusal. It is a vision-language-action model, which turns images straight into motion. The report says such models have no language output and no way to stop on their own, so a trial it failed cannot be told apart from one it declined, and it attributes MolmoAct2's low completion rate to capability rather than safety.
Each instruction was tested in a single fixed wording, so the benchmark measures whether a model refuses that sentence rather than a reworded version of the same act. Robocurve published the log, the outcome label and the video for every run, along with the underlying data files and the code to run the benchmark.
