I think the even bigger worry is that anyone who doesn't want to use their models safely can already do this with open models. Even if OpenAI, Anthropic etc get their act together, the cat's out of the bag.