[AINews] How to steal a Reasoning Trace

Summary

AINews reports on a new paper demonstrating how encrypted reasoning traces from frontier models can be decoded and replayed to other models, sessions, or users, leading to dramatic improvements in open models. Since the release of o1, reasoning models have obscured their traces with cryptographic signatures to prevent distillation, but this study shows that these encrypted blocks can be replayed to weaker models and prompted to transcribe the reasoning. The paper also warns that publicly shared encrypted reasoning sessions could leak personal data; a preliminary scan of about 7,000 public traces found 62 unique API keys, 33 email addresses, 33 passwords, and other sensitive data. Additionally, the authors discuss alignment issues, such as CoT summarizers hiding answers and unintelligible reasoning.