LLM APIs Leak Reasoning Traces, Let Attackers Reconstruct Proprietary Chains‑of‑Thought
A new paper shows that Anthropic, OpenAI and Google return encrypted reasoning blocks that can be swapped into weaker models, forcing them to output the original chain‑of‑thought in plaintext. The authors decoded 315k public traces, exposing 367 PII records, 182 credentials and enabling hidden prompt‑injection attacks, effectively stealing proprietary model reasoning.