[AINews] How to steal a Reasoning Trace
Researchers decoded encrypted reasoning traces from frontier models, exposing sensitive user data
“if you ever shared online a Claude Code/Codex session with encrypted reasoning blobs, they can be decoded and leak your personal data.”
A new paper demonstrates that encrypted reasoning traces from frontier lab models like Claude and Codex can be decoded and transferred across models, sessions, and users — breaking the cryptographic protections labs added post-o1 to prevent distillation. A preliminary scan of ~7,000 public traces found 62 API keys, 33 emails, and 33 passwords leaked exclusively inside reasoning blocks invisible to users. This sits at the critical intersection of alignment, security, and chain-of-thought monitoring, with immediate implications for anyone who has shared AI coding sessions publicly.