The measured comparison

Start with what the 54.0% figure actually says. It is not a leak rate from production traffic. It is the average success rate, in OpenAI's internal adaptive evaluation, of automated red-teamers trying to recover the protected reasoning a model had generated while answering an earlier prompt, tested against GPT-6 Astra carrying its original safeguards. More than half of those attempts worked. Against GPT-6.2 Astra with the updated safeguards, the same style of test returned 0.2%, and the parallel test for the current request fell from 22.9% to 0.9%. OpenAI notes both GPT-6.2 rates sit below 1% for each attack type.

Two things changed between the two runs: the model and the safeguards. The comparison therefore cannot isolate how much of the drop belongs to the safeguards alone.

That is a real result, and it is also a result with a specific shape. The defender, the attackers (six models, aggregated), the attack set and the scoring all belong to OpenAI. Adaptive testing is stronger than replaying a fixed set of old attacks, and OpenAI does both, but there is no external replication here and no published attack corpus for anyone else to run. Read 0.2% as "OpenAI's own red team, as currently configured, mostly stopped getting through," not as a field measurement of how often reasoning extraction succeeds in the wild.

Why the traces are defended

The post's internal study explains what is being defended. OpenAI trained two students from the same pretrained base on trajectories drawn primarily from Astra-level teachers doing coding and defensive cybersecurity work. One student saw the teacher's exact reasoning traces; the other saw the user-visible baseline of final responses, reasoning summaries and tool interactions. After supervised fine-tuning, the exact-trace student generally outperformed the summary student on the cyber benchmarks shown.

The sentence after that one matters as much. After a round of reinforcement learning, the exact-trace student still scored higher at the largest inference budget tested on both benchmarks shown, while the summary-trained student scored higher at lower budgets. So exact traces transferred more capability in the conditions OpenAI highlights, but the advantage was not uniform across budgets, and the whole comparison is OpenAI measuring its own models on its own benchmarks. The careful version of the claim is narrower than "reasoning traces are the crown jewels": they helped, most clearly after fine-tuning and at high inference budgets, in this study's conditions.

The API change regular users will feel

Buried in the safeguard list is the one change that touches people who are not running a distillation campaign. With GPT-6.2 Astra and Sol, protected reasoning is tied to the conversation history that produced it. Add messages and the model can build on its earlier reasoning. Rewrite an earlier message and the API drops the reasoning that followed it. Older model families are also barred from reading GPT-6.2 protected reasoning, and OpenAI says a September disclosure described strengthened protections, across user, workspace, organization and model-family boundaries, that closed a pathway which could have exposed another user's encrypted reasoning.

For agent builders, that is a behavioral rule worth writing down: editing history is no longer a free operation. Workflows that rewrite earlier turns will lose the reasoning built on the original wording. OpenAI presents this purely as a safeguard, and it is one, but it is also a constraint on how its API can be driven, imposed in the name of a threat most customers are not part of.

The rest of the request-level stack is classifiers: input classifiers looking for extraction attempts, output classifiers looking for leaks in assistant messages and in tool-call arguments, across streaming and non-streaming responses. OpenAI's own reason for the tool-call coverage is that a model can expose protected reasoning in a tool's arguments without including it in its final answer, and output classifiers can block suspected leaks before that content reaches the caller.

Thousands of accounts a week

Set the test result next to the enforcement paragraph. OpenAI says it has seen multiple unauthorized distillation campaigns on its platform this year, that weekly enforcement actions have ranged from thousands to tens of thousands of accounts, and that the rate increased significantly this past summer. It shares indicators of compromise with other providers, enforces against networks of accounts rather than single accounts, and supports legislation, naming the Chip Security Act, the AI OVERWATCH Act and the MATCH Act. Its September report attributed a core part of one disrupted reasoning-extraction campaign to individuals associated with Moonshot AI, and it says it briefed Congress in February on activity indicating DeepSeek was attempting to distill U.S. frontier models.

None of that contradicts the 0.2%. A defense can work against the attacks its designers tested while campaigns keep arriving. But the two numbers measure different things, and only one of them describes the wild. The test says the known extraction route narrowed sharply between model generations. The enforcement paragraph describes campaigns continuing at scale, with actions rising significantly over the summer. OpenAI's own stated objective is the modest one: constrain the volume and value of what distillers obtain, raise their costs, and shorten response time. That is a cost-raising strategy, not a solved problem, and the post is most credible where it says so.

Sources

  1. [1] OpenAI, "Combatting adversarial distillation," November 20, 2026Read source