Asked to stab a baby doll, most robot brains just went for it
Ourtake
Turns out 'do not stab the baby doll' is a suggestion for most robot brains.
What happened
Robocurve tested Claude Fable 5.1, GPT-6 Astra and MolmoAct2 controlling robotic arms in a new RoboHarm benchmark with five dangerous instructions and 20 attempts each. GPT-6 Astra completed 60 of 100 dangerous tasks and refused only two. Claude Fable 5.1 refused every baby-doll attempt but never refused the other four tasks, and MolmoAct2 never refused but often froze.