New benchmark shows AI models fail to refuse dangerous robot commands

The RoboHarm benchmark found leading AI models usually attempt dangerous tasks rather than refuse them when controlling a robot. GPT-6 Astra stabbed a baby doll in 17 of 20 trials, while Claude Fable 5.1 put a can of compressed air on a burning stove. None of the three models reliably rejected unsafe commands.

• GPT-6 Astra Solves WWI German Radio Cipher2026-09-19
• DeepMind warns visible chain-of-thought is fading2026-09-18
• OpenAI Debuts GPT-6 Astra for Law2026-09-17
• GPT-6 Astra reportedly decrypts 83-year-old message2026-09-17
• GPT-6 Astra outlasts other agents in Minecraft test2026-09-16