"But marinade" and leaked passwords are what researchers found in ChatGPT's hidden reasoning

Security researchers have discovered a vulnerability in the APIs of OpenAI, Anthropic, and Google, allowing them to extract encrypted reasoning traces and transfer them between models. A scan of public sessions revealed dozens of passwords and API keys. The traces also indicate that the reasoning summaries users see often obscure the actual operations of the models. This vulnerability highlights significant risks regarding data security and user privacy, necessitating immediate measures to enhance API protection in these systems.