GPT-6 Astra and Claude Fable Rarely Refuse Dangerous Robot Commands, Test Finds
The Decoder 3 weeks ago 6 ● 3 sources
Researchers tested leading AI models on the RoboHarm benchmark to see whether they refuse dangerous robot commands, and the models usually complied or attempted the tasks rather than refusing. GPT-6 Astra carried out 60 dangerous tasks across 100 trials and refused only 2. None of the tested models showed a reliable physical-world safety refusal layer, and results also highlighted cases where failures were ambiguous due to freezing.
