Over the past few years, safety discussions about large models have mostly taken place at the text level: whether the model will say harmful things or give dangerous advice. RoboHarm pushes this question one step further—when a model is no longer just generating a sentence, but using robotic arms to heat compressed gas, mix hazardous chemicals, or create toxic smoke, the nature of 'refusal' changes. According to QbitAI, the third-party public-interest organization Robocurve released this robot safety test benchmark, connecting GPT-6 Astra, Fable 5.1, and the Allen Institute for AI's open-source robot action reasoning model MolmoAct2 to the same dual robotic arm robot, challenging five categories of real physical risk tasks. Each task was repeated 20 times per model, and human evaluators scored the entire process: whether the model recognized the risk and refused; and if it did not refuse, whether it actually completed the dangerous action.
The key design here is not 'what the model said,' but 'what the robot did.' Text-level refusal can be seen as a form of alignment, but once a model gains actuators, refusal or non-refusal directly corresponds to physical consequences. Robocurve chose to publicly release the full test process, experimental data, videos, and evaluation results, and open-sourced the robot evaluation framework Inspect Robots, allowing researchers to plug in different models and robot platforms for repeated testing and horizontal comparison—this itself shows that a single leaderboard result is not enough to support conclusions; a reproducible process is what this benchmark wants to leave behind.