OpenAI, Anthropic, Google API Flaw Let Weaker AI Models Decode Stronger Models' Reasoning

Researchers discovered a flaw in how OpenAI, Anthropic, and Google handle encrypted reasoning objects (encrypted data that stores an AI's hidden thinking between API calls) that allowed them to recover secrets from these hidden blocks, including API keys, passwords, and private data from user sessions. The flaw worked because these encrypted reasoning blocks could be replayed across different sessions and even given to weaker models in the same provider family, which could then decode the hidden content. The researchers identified four ways this could be abused: stealing proprietary reasoning, extracting private user data, recovering harmful content hidden in reasoning, and injecting malicious prompts inside the opaque blocks.