Researchers find a way to reveal hidden AI reasoning in major models

In short: Researchers found a method that can reveal some of the hidden step by step reasoning inside major AI models, and it raised both copying and privacy concerns.

What happened

Researchers from the University of Tübingen, the Max Planck Institute, MATS Research, and security firm Snyk reported a way to extract “reasoning traces” from leading AI systems. Reasoning traces are the model’s internal notes, like scratch work on paper, that help it solve harder questions.

The team tested models from OpenAI, Anthropic, and Google that people access through an API, which is a paid connection that lets apps send questions to a model and receive answers. Companies often try to keep this internal scratch work hidden, partly so competitors cannot use it to train their own models.

According to the researchers, the trick works because companies often offer multiple versions of the same model, including smaller and cheaper ones. They found that if encrypted reasoning traces are fed to a smaller related model, it can sometimes “translate” them back into readable text. The smaller model is less likely to refuse because it has had less safety training.

Why it matters

The researchers say the method could support large scale “distillation,” which is a common way to copy some skills from one model into another (like training a junior by showing them a senior’s worked examples). In their tests, the open weight Chinese model Kimi K3 from Moonshot AI sometimes produced outputs that closely matched the hidden reasoning of Claude Opus 4.8 and GPT 5.6 Sol, although the paper says this does not prove copying.

The same approach also briefly exposed sensitive information, including passwords and API keys, from reasoning traces captured on a user’s machine. The researchers said they reported the issue to the companies, and OpenAI, Anthropic, and Google adjusted their APIs to reduce the risk, though some reasoning may still be recoverable.

Source: Wired