A human artist is obviously capable of generating CSAM, even if they have never seen that before.
Filtering of training data is countered by increasing capabilities to generalize:
Two years ago, that was a viable strategy: models could barely produce what was in the training data again.
Today models can generalize much better and compose concepts they have been trained on into new concepts that they haven’t.
Two years from now, filtering will be irrelevant.
Keep in mind, there already are a lot of illustrated/anime style pictures of CSAM on that site of years though (something that is legal in many countries), so it's sort of becoming a blurred area as these AI art generators are still somewhat like that but now are getting to be more photorealistic.
As far as the models not being trained on NSFW content, there was already leaked models that were, and there are unofficial models trained by outsiders using SD that are specifically trained on for example adult image websites.