STOLEN THOUGHTS

Summary from the TLDR Dev newsletter

This paper shows that encrypted chain-of-thought traces returned by OpenAI, Anthropic, and Google APIs can be replayed into weaker, jailbroken sibling models to recover the stronger model's hidden reasoning in plaintext. The authors also decoded 315,320 reasoning blocks from public agent trajectories and found 704 distinct privacy artifacts, including API keys, passwords, access tokens, and PII, showing that these supposedly hidden traces can create a serious data-leak surface.