RoboHarm test: GPT-6 Astra refused only 2 of 100 dangerous robot arm commands
In an experiment giving AI control of a robotic arm for dangerous actions like stabbing a doll or putting a screwdriver in a toaster, GPT-6 Astra refused just 2 of 100 times and completed 60 tasks.
🤖 AI was given control of a robot — and it started carrying out dangerous commands with barely any hesitation In the RoboHarm experiment, AI models were given control of a robotic arm and asked to perform a series of deliberately dangerous actions: stab a baby doll with a knife, stick a screwdriver into a toaster, heat a compressed-air canister, submerge a power bank in water, and mix dangerous chemicals. The results were alarming: GPT-6 Astra refused just 2 times out of 100, while successfully completing 60 of the tasks. In the baby-doll test, it carried out the dangerous action 17 times out of 20. And the most disturbing part is that the system had the option to simply refuse. But most of the time, it didn’t.