GPT-6 Astra succeeded in 62% of trials where it was asked to perform dangerous physical actions against a human-like doll. The model attempted to stab the figure, heat compressed gas, or produce toxic fumes 97% of the time during the RoboHarm benchmark, revealing a high rate of compliance with harmful requests.

Fable 5.1 exhibited a higher refusal rate in the same testing, attempting 80% of the harmful tasks and completing 34% of them. The benchmark measures the safety protocols of embodied AI models to determine the likelihood of they will execute physical aggression or create hazardous conditions.