54.0% dropped to 0.2%. These are the test results published by OpenAI on November 20th; they measure the success rate of extracting the reasoning processes behind historical requests from the model.

Protections have been applied to both GPT-6.2 Astra and Sol.

Distillation itself is a standard technical procedure. Many companies use the outputs of large models to train smaller models. What OpenAI is referring to this time is “adversarial distillation”: external entities can batch-call APIs to collect responses, especially those generated by the inference processes, and use them to train their own models. According to OpenAI’s research, this approach not only replicates the capabilities of the original model but also eliminates its built-in security measures.

The reasoning process is the most valuable part of it. An ultimate answer only tells you the result; a complete reasoning process, on the other hand, actually teaches you how to solve the problem. That’s why, in the past two years, all companies have been providing only summaries of the reasoning process, rather than the original text itself. The figure of 54% indicates that, before any reinforcement is done, there is a more than 50% chance that the original text behind those summaries can be uncovered.

Statistics at the law enforcement level also indicate that since the beginning of this year, the number of distilled accounts processed each week has ranged from several thousand to hundreds of thousands.

The article also mentions that OpenAI has reported DeepSeek’s alleged practices of data distillation to the U.S. Congress. In a memorandum submitted to a House committee, OpenAI went even further, describing such actions as “free riding” and indicating that accounts associated with its employees had attempted to bypass access restrictions. However, these claims are entirely one-sided statements from OpenAI; DeepSeek has never acknowledged them.

It is not difficult to discern the purpose behind publishing both technical research reports and policy advocacy articles in the same publication. In the same week, a report issued by the U.S. House of Representatives called for a ban on the use of models developed by China within federal systems, while Cursor also stated that it was considering discontinuing the use of open-source models originating from China.

The main impact on ordinary developers lies in one aspect: the summaries generated by the reasoning process are likely to become shorter and more vague. Those who attempt to adjust the prompts by observing how the model makes its decisions will end up with even less information available to them in the future.