Researchers Replay Encrypted Chain of Thought From Frontier LLMs Into Weaker Sibling Models to Read Plaintext Reasoning · cho.sh