This is kind of a stupid argument. How about make a slightly stronger claim like "models won't break symmetric crypto" ? I mean, language models aren't even trained to break symmetric crypto. There is not good reason to think they will. It seems possible to train a large model to do it though.