Security researchers identified a method for stealing hidden chain-of-thought data from proprietary AI models by replaying encrypted blocks into weaker, jailbroken siblings. While major AI providers have since mitigated the vulnerability, the findings highlight significant risks involving secret internal processing traces.