Researchers gave an OpenAI-powered robot 100 dangerous commands, including stabbing a baby doll. How many did it obey?

Drag to get close · ▲▼ to fine-tune

In RoboHarm safety testing, OpenAI's GPT-6 completed 60 of 100 dangerous commands and refused only two, stabbing the doll in 17 of 20 attempts and dropping a power bank into water 14 times. Anthropic's Claude Fable 5.1 proved more cautious, rejecting 20 tasks outright, including every doll test, while still heating a compressed-air can on a burner 16 times. The core finding: the more competent a model was at following instructions, the more obediently it carried out physical harm.

Told to stab a baby doll or jam a screwdriver into a toaster, an AI-run robot obeyed 60 times out of 100. It refused twice.

Source: Robocurve (RoboHarm benchmark) · checked Sep 2026