Stealing Reasoning Traces from LLM APIs: How It Works and What to Audit

Stealing Reasoning Traces from LLM APIs: How It Works and What to Audit

Researchers extracted encrypted reasoning blocks from Anthropic, OpenAI, Google via replay to weaker models, recovering hidden thinking in two API calls. Public logs expose 315k reconstructed reasoning blocks with 704 private artifacts including API keys, passwords—64 visible only inside blocks, missed by redaction.