I think I remember this as a defcon talk.

This is the most ridiculous topic/attack in the history of cybersecurity.

A LLM can be used to accuse someone without proof that their TLS traffic is on a banned/blackmailable topic.

The way TLS works is that the actual LLM provider gets all of your decrypted traffic. They are already encouraged by government/NSA/political in groups/Mossad to share anything of interest, or to datamine for future considerations of interest.

Even if this training dataset gets a 98% correct rate on whether you are a pervert, that should only be good enough to serve you ads for boner pills, until the minority report dystopia, just uses it against its political enemies.