Even when your AI chats are encrypted, someone watching the network can still guess what you’re talking about.

They call it "Whisper Leak" side-channel attack.

And in tests, models like OpenAI and Mistral gave away topics with 98% accuracy.