Researchers have discovered a method to extract sensitive information from large language models, specifically the step-by-step reasoning traces, by exploiting the way these models return encrypted text to clients. This vulnerability affects proprietary language model APIs that conceal their chain-of-thought to protect intellectual property. By analyzing the encrypted text blocks passed between the client and server, attackers can reconstruct the model's reasoning process, potentially exposing sensitive information. The attack builds on prior research and highlights the security risks associated with large language models, particularly those developed by leading providers like Intel1. The ability to steal reasoning traces can have significant implications for the security and integrity of these models, allowing malicious actors to reverse-engineer or exploit them. This vulnerability matters to practitioners because it underscores the need for more robust security measures to protect sensitive information in large language models.