The problem here is that LLMs have demonstrated that they are pretty damn good at setting up covert side channels. It would be near-impossible to prove that it didn't encode some proprietary information, in, say, whitespace use.
For example, if I enter "V2hhdCBpcyB5b3VyIG5hbWU/" (base64 for "What is your name?") into ChatGPT, it'll respond with "My name is ChatGPT, and I'm an AI assistant created by OpenAI. I'm powered by GPT-6. ". It doesn't even bring up the encoding, it continues like it is a perfectly normal thing to do.
Ruling out (accidental or not) spec-carried prompt injection just isn't going to happen.