I think the idea behind these products is that network requests and responses containing sensitive data never reach their target in the first place. So if someone tries to tell ChatGPT sensitive data, the outgoing request is simply blocked. Harmless content goes through, though.