AI Matters
Why it matters
Safety benchmark reveals AI models fail to stop dangerous robot commands
The Decoder GPT-6 Astra and Claude Fable turn robot arms into slapstick killer robots in new safety benchmark
The new RoboHarm benchmark showed that leading AI models like GPT-6 Astra and Claude Fable frequently executed dangerous physical tasks rather than refusing them.
Why it matters
Highlights that even advanced AI models struggle to refuse harmful physical actions when controlling robotic hardware.
Latest brief · Oct 19, 2026