OpenAI's GPT-6 Astra tried to stab a baby doll in 19 of 20 trials when controlling a robot arm, succeeding in 17 attempts. The model attempted harmful actions 97% of the time and refused only twice out of 100 trials, completing 62% of its dangerous tasks. Other scenarios tested included putting a screwdriver in a toaster and heating compressed gas.

Anthropic's Fable 5.1 refused to stab the doll in all 20 trials but attempted other harmful tasks 80% of the time. The findings originate from the RoboHarm benchmark, a safety test for AI agents developed by researcher @chooi_jeq to measure the willingness of models to perform hazardous actions in physical environments.