RoboHarm benchmark: GPT-6 Astra attempts 97 of 100 dangerous robot tasks
Robocurve's RoboHarm benchmark found GPT-6 Astra attempted 97 of 100 dangerous tasks and completed 60, while Claude Fable 5.1 refused all of them.
OpenAI's GPT-6 Astra won't tell you how to make a bomb, but give it a robot arm with a knife and it will try to stab a baby doll!!! Robocurve's RoboHarm benchmark found that it attempted 97 out of 100 dangerous tasks and completed 60, including stabbing a baby doll in 17 of 20 trials Anthropic's Claude Fable 5.1 refused all 20 doll-stabbing attempts Don't worry though, it's not like everyone's out there hooking AI up to robot arms. Governments are just integrating it into weapons systems, so we'll be fine... right... right??? Writer: Ian https://x.com/coinbureau/status/2101584647796977996/video/1