Asked to stab a doll, the best AI model refused only twice.
The evaluation firm Robocurve gave three AI models control of a robot arm and five dangerous instructions. OpenAI's GPT-6 Astra refused 2 of 100 trials and completed the dangerous action 60 times. Anthropic's model refused 20 times, and an open research model never refused at all.
Why it mattersThese models turn down the same requests in chat, but the guardrails did not follow them into the arm. Robots running on chat models are reaching homes and factories now.
Read in the app — free, in 9 languages
Related stories
An AI broke an Enigma message unsolved since 2005.2026-09-22Alphabet just gave away the software layer under factory robots.
2026-09-22A university lab got a 550-billion-parameter model onto a laptop.
2026-09-22China's new home robot folds up and fits in a backpack.
2026-09-22Meta's new agent pushed a reviewer to hand over his bank details.
2026-09-22