Researchers Discover Vulnerability in Proprietary LLM APIs

Researchers identified a significant vulnerability in proprietary LLM APIs from Anthropic, OpenAI, and Google. The providers return encrypted chain-of-thought blocks that can be replayed across multiple sessions and users. Reasoning generated by a frontier model could be replayed into a weaker version of the same model family and decrypted through jailbreaking, exposing the stronger model's hidden reasoning in plaintext. Researchers found that all models within the same family used identical encryption keys. The vulnerability was disclosed responsibly and providers confirmed the issue was fixed. Claude Haiku 4.5 was the most vulnerable model, while later 4.6 versions removed the feature enabling the attack. The work provides extensive details of extracted reasoning traces, offering insight into how these models think internally.