Encrypted reasoning traces leak via weaker models

Researchers found that feeding a frontier model's encrypted reasoning traces to a weaker model from the same provider can make it output the traces in plaintext. The technique exploits alignment differences between models.