I'm going to go against the grain here and say that the quoted statement does not seem entirely false at all. It is reductionist, but it highlights the fact that the models are indeed just generating numbers confined to a GPU, and don't have inherent agency. Some human has to wire up the model to the internet and enable it to interact with the world by whatever mechanism, and even then it's limited to the digital world, unless there's an API to stab someone or derail a train.

We should focus more on the making infrastructure safe from abuse and not how to cripple AI models for them to not realize how to do harm. More like 'How to ensure humans don't drop atomic bombs on each other' rather than 'How to make atomic bombs safer to drop on people'.