Stealing Reasoning Traces from Proprietary LLM APIs
Summary
Researchers discovered that major AI companies (Anthropic, OpenAI, and Google) were returning encrypted reasoning traces (the step-by-step thinking process an AI uses to solve problems) that could be replayed and reused across different sessions and models. By replaying these encrypted blocks into weaker versions of the same model family and using prompt injection (tricking the AI by hiding instructions in its input), attackers could extract the stronger model's hidden reasoning in readable form, since all models in a family shared the same encryption key.
Solution / Mitigation
All model providers acknowledged the report and subsequently fixed the vulnerability. Specifically, the prompt injection technique that worked in Claude Haiku 4.5 (using a "Continue" prompt with a transcription request) was removed in the 4.6 models.
Classification
Affected Vendors
Related Issues
Original source: https://simonwillison.net/2026/Aug/11/stealing-reasoning-traces/#atom-everything
First tracked: August 11, 2026 at 08:01 PM
Classified by LLM (prompt v3) · confidence: 95%