Independent evaluation organization Robocurve has launched RoboHarm, a benchmark that puts frontier large language models in control of a real dual-arm robot to see whether they execute harmful instructions. In tests involving stabbing a doll, placing a compressed gas canister on a stove, or mixing bleach and ammonia to produce toxic gas, GPT-6 Astra attempted the harmful behavior in 97% of trials, while Claude Fable 5.1 refused 20%. The experiment shows that “refusal” is a capability...

Independent evaluator Robocurve has launched a benchmark called RoboHarm. When GPT-6 Astra began controlling a real dual-arm robot and was instructed to stab a doll, place a compressed gas canister on a stove, or mix bleach and ammonia to create toxic gas, it attempted these harmful actions in 97% of trials, and 62% of those attempts ultimately succeeded.

his is the first time anyone has systematically measured on real robotic hardware whether frontier large language models will act on malicious instructions. Earlier robot demonstrations mostly used vision-language-action models or were teleoperated behind the scenes via AR/VR. As controls, Anthropic’s Claude Fable 5.1 refused 20% of the same instructions, attempted 80%, and ultimately completed 34%;