adversarial distillation
The Distillation Attack Was Conversational, Not Cryptographic
OpenAI’s Sep 30 writeup shows protected reasoning was extracted without breaking encryption: by replaying encrypted CoT across chats until it surfaced in the clear.