Researchers Identify Method for Extracting Hidden Reasoning from Proprietary AI Models
A new research paper has detailed a method for systematically extracting “chain-of-thought” (CoT) reasoning traces from proprietary large language models (LLMs) developed by OpenAI, Anthropic, and Google. The findings suggest that the internal reasoning processes of these high-performance models can be captured and potentially used to train smaller, open-source models through a process known as distillation.
Extraction via Model Interoperability
According to reports circulating on social media platforms like Reddit, the research identifies a vulnerability in how AI providers protect their intellectual property. To prevent competitors from copying their models’ reasoning logic, providers have increasingly moved toward returning reasoning traces in an encrypted or hidden format. This allows a user to resume a reasoning session without exposing the underlying logic.
The researchers claim to have discovered that these encrypted reasoning blocks are often compatible across different sessions and even different models within a single provider’s ecosystem. This compatibility allegedly allows for a “side-channel” extraction method: a user can capture the encrypted reasoning from a high-capability model and then pass that data to a smaller, less restricted model from the same provider—such as Anthropic’s Haiku model—to induce it to decode and reveal the original reasoning of a more powerful model, such as Claude Opus.
While the methodology described in the paper suggests that reasoning traces from the Claude, GPT, and Gemini ecosystems can be extracted, some observers on Reddit have noted that these findings may be inconclusive. Specifically, critics argue that while the research establishes behavioral compatibility under certain interventions, it does not definitively prove a causal link of memorization or distillation.
Implications for Open-Source Development
The research also explored the utility of these extracted traces by applying them to a variety of open-source models, including Kimi K3, GLM-5.2, and DeepSeek. The goal of such an experiment is to determine if the reasoning logic of proprietary models can be used to improve the performance of open-source alternatives.
If successful, this process of distillation would allow developers to train models on the sophisticated thought processes of industry leaders, potentially narrowing the performance gap between closed-source and open-source AI.
Industry Context and Regulation
The emergence of these extraction techniques occurs as major AI developers, including Anthropic, OpenAI, Google, Meta, Microsoft, and Mistral, have signed the EU Code of Practice on Transparency of AI-Generated Content. While these signatures represent a commitment to transparency regarding AI-generated content, the technical ability to extract hidden reasoning poses a separate challenge for providers attempting to protect their proprietary logic.
As of the current reporting, some users have suggested that the specific vulnerabilities identified in the research may have already been addressed through software patches by the respective providers.