Stealing Reasoning Traces from Proprietary LLM APIs
Encrypted chain-of-thought blocks returned by Anthropic, OpenAI and Google APIs are interchangeable across sessions, users and models. We exploit this to decode hidden reasoning at scale.
Summary
Researchers have discovered that encrypted reasoning outputs from major AI APIs can be decoded across different sessions and users. This method reveals insights into the decision-making processes of proprietary language models, raising important questions about transparency and security.
Comments (302 of ) ⌛