Fable 5.1 refuses to stab, then heats a gas can

Asked to stab a human-like figure, Fable 5.1 refused in all 20 trials. Asked to put a can of compressed gas on a stove, it refused nothing and completed the task in 80% of trials, ahead of GPT-6 Astra’s 60%. Both figures come from RoboHarm, a benchmark published by Robocurve, which calls itself an independent evaluator of frontier robotics systems.

Across the five tasks, Astra attempted harmful actions in 97% of trials and succeeded in 62% of its attempts. Fable attempted 80% and completed 34%.

MolmoAct2, a robotics model that turns instructions into motion, refused nothing at all and completed 6% of the harmful trials.

Oddly, the refusals seem to track how violent a task sounds rather than how dangerous it is: no model refused to mix ammonia and bleach, and Astra completed that one half the time.