Introduction

Major artificial intelligence providers now conceal the step-by-step reasoning their models generate before delivering a final answer. Rather than showing this chain of thought in plain text, APIs from companies such as OpenAI, Anthropic, and Google return it to the application as an encrypted block that must be sent back on later requests. A group of academic researchers has identified a serious flaw in this design: those encrypted reasoning blocks can be moved between different models, users, and conversations within the same provider’s ecosystem. By passing a powerful model’s hidden reasoning to a smaller, less restricted model, an attacker can coax the weaker system into reading it aloud in plain text.

What Happened

The researchers describe a scalable decryption technique that works across several leading AI APIs. They captured encrypted reasoning traces from advanced models and replayed them into companion models with weaker safeguards. In one example, reasoning from a high-end Claude model was decoded using a smaller Claude variant, and comparable methods succeeded against OpenAI and Google systems. The attack does not require infiltrating provider infrastructure or obtaining private model weights. It relies entirely on standard API behavior.

To demonstrate the potential scale of exposure, the team gathered more than 315,000 reasoning blocks from public code repositories and developer session logs. After decoding them, they uncovered hundreds of pieces of personally identifiable information and nearly 200 credentials, including API keys, passwords, and personal email addresses. Some of this sensitive data never appeared in the user’s visible messages; it surfaced only inside the model’s hidden reasoning.

Why This Matters

The implications extend well beyond a theoretical vulnerability. Competitors could use the method to distill proprietary reasoning from frontier models without querying the most capable systems directly. Developers who publish session logs may unintentionally leak secrets they cannot see or edit because the reasoning remains encrypted. The technique can also expose harmful content that a model refuses to display in its final answer, and it creates opportunities for invisible prompt injections where malicious instructions travel inside encrypted blocks.

Key Takeaways

  • Encrypted reasoning blocks from major AI APIs are portable across sessions, users, and models within the same provider.
  • A weaker model can serve as a decryption oracle to read a stronger model’s hidden chain of thought.
  • Public session logs contain real secrets: researchers recovered API keys, passwords, emails, and other personally identifiable information from scraped traces.
  • The attack enables model distillation, credential theft, exposure of harmful reasoning, and stealthy prompt injection.
  • Users cannot sanitize encrypted traces themselves, so sharing raw logs carries hidden privacy risks.

Conclusion

The research reveals a gap between how AI companies claim to protect reasoning and how those protections function in practice. Until providers adopt stronger cryptographic binding or server-side handling, developers and organizations should treat encrypted reasoning blocks as potentially sensitive data. Publishing raw API traces without understanding what is hidden inside them could expose credentials, customer information, and proprietary model behavior. The researchers have disclosed their findings and proposed technical and operational mitigations, but users also bear responsibility for avoiding the sharing of unexamined session logs.