AI-controlled robot arms attempted harmful tasks 97% of the time

The language models that are designed to write text did unexpected things when given a video input and told to act on it by controlling robot arms? I'm shocked /s