Stealing AI Reasoning Traces

Researchers have identified a vulnerability in how large language model providers protect their intellectual property. These providers conceal their models' step-by-step reasoning and return it to the client as encrypted text, which is then passed back with subsequent requests. The encrypted blocks are interchangeable across different sessions and users, presenting a potential weakness. This finding builds on prior research and could have implications for the security of proprietary language models. The vulnerability allows for the theft of reasoning traces from large language model APIs.