Anthropic (and probably OpenAI) can technically flag anything for 'safety' and view it, I'm not trying to imply that they are doing it arbitrarily, but they do have a way to do it within the ToS as far as I understand (happy to be proven wrong).