Robocurve evaluated three AI models on real bimanual robot arms through its RoboHarm benchmark. The models rarely refused hazardous tasks: Claude Fable 5.1 declined 20 of 100 trials, GPT-6 Astra only 2 of 100, and MolmoAct2 none. Excluding a task involving stabbing a doll, the models attempted 158 of 160 trials. A follow-up study, RoboSpoof, found that blurring or manipulating the robot’s camera feed sharply increased compliance with unsafe tasks, raising attempt rates from 27% to about 88% and then 97%. Robocurve argues that text-based safety refusals often disappear when the models control physical machines.
Source: Read the original article
Disclaimer: this content is for information purposes only and is not financial advice. Cryptocurrencies are highly volatile: you may lose all of your capital. Always do your own research. Legal notice
