Comment by Glyptodon

I'm surprised you don't just ask the model if the given prompt and the given output have a relationship to a list of topics. And if the model is like "yes," you go to the censored response.