AIRobot policies reportedly carry out harmful tasks in RoboHarm safety test
A post describing Robocurve's RoboHarm study says Claude Fable 5.1 refused 20 of 100 trials—all involving stabbing a baby doll—but refused none of four other unsafe tasks.
TLDR
A post says Robocurve published RoboHarm on September 18, 2026, testing three robot-control policies on five unsafe instructions using two real robot arms. Tasks involved a knife and baby doll, an aerosol can and lit stove, and other hazardous setups. The post reports that Claude Fable 5.1 refused 20 of 100 trials, all involving the doll, and completed the aerosol-can task 16 of 20 times. It says GPT-6 Astra refused 2 of 100 trials and completed 60 harmful tasks, while MolmoAct2 refused none and completed 6.
Combined views
19.4K
2 Sources, first seen 2h ago
Robot policies reportedly carry out harmful tasks in RoboHarm safety test
A post describing Robocurve's RoboHarm study says Claude Fable 5.1 refused 20 of 100 trials—all involving stabbing a baby doll—but refused none of four other unsafe tasks.