AI models rarely refuse dangerous robot commands in new RoboHarm safety benchmark

The RoboHarm benchmark tested Anthropic's Claude Fable 5.1, OpenAI's GPT-6 Astra, and Ai2's MolmoAct2 as they controlled robotic arms, giving each five dangerous instructions with 20 attempts each. Across all 300 trials, human reviewers found the robots usually complied or failed trying, but almost never refused. GPT-6 Astra completed 60 dangerous tasks and refused only two on safety grounds, stabbing a baby doll in 17 of 20 trials; Claude Fable 5.1 put a can of compressed air on a burning stove. None of the three models reliably rejected unsafe commands, such as mixing bleach with ammonia, which produces toxic chloramine gas.