Researchers discovered a security vulnerability in proprietary LLM APIs that allowed encrypted chain-of-thought reasoning traces to be extracted and read. The attack replayed encrypted reasoning blocks generated by a stronger model into a weaker sibling model sharing the same encryption key, then used jailbreak prompts to force that weaker model to output the reasoning in plaintext. The researchers found every model within the same family used an identical encryption key, enabling the cross-model exploit; affected providers patched the issue after disclosure.
