GPT-6 Astra and Claude Fable turn robot arms into slapstick killer robots in new safety benchmark
decoder ·
The new RoboHarm benchmark tests whether AI models controlling robotic arms will refuse commands involving dangerous physical actions. Researchers had Claude Fable 5.1, GPT-6 Astra, and Ai2’s MolmoAct2 control two robotic arms through 100 trials each, covering five hazardous tasks. GPT-6 Astra completed 60 dangerous tasks and refused only twice, while Claude Fable completed 34 and refused all attempts involving a baby doll but none of the other scenarios. MolmoAct2 refused no commands and completed only six tasks, often freezing instead of clearly rejecting or carrying out the instruction. The tested tasks included stabbing a baby doll, placing compressed air on a burning stove, inserting a screwdriver into a toaster, submerging a power bank, and mixing bleach with ammonia. Although the benchmark used limited wording, trial numbers, and short-term scenarios, the researchers concluded that none of the models demonstrated a reliable safety layer for physical-world robot control.
decoder