Skip to main content

OpenAI Details Safeguards Against Model Distillation by Rivals

OpenAI published a post on 20 November 2026 describing how it defends its models against adversarial distillation, where outsiders harvest a model's outputs to train a copy. It says updated safeguards in GPT-6.2 Astra and Sol cut the success rate of extracting hidden reasoning from 54.0% to 0.2% in internal red-team tests. Those figures have not been independently verified.

On this page

What OpenAI means by adversarial distillation

Distillation is an ordinary technique: one model learns from data produced by another, and labs use it all the time to build smaller, cheaper versions of their own systems. OpenAI's post is about the unauthorized kind, where someone uses a frontier model's outputs to train a rival with similar abilities and then releases it without comparable protections.

The company frames this mainly as a safety problem instead of an intellectual property one. Its example is cyber capability: a model distilled from a system that refuses to help with attacks could end up helping with them. The prize for an extractor is the model's protected reasoning, the step-by-step working that is normally hidden from users, because it contains information the final answer leaves out.

OpenAI backs that with an internal study. A student model trained on exact reasoning traces generally outperformed one trained only on reasoning summaries on cyber tasks, although the summary-trained student did better at lower inference budgets after a further round of training. The results appear only as charts, and the company says it will keep studying how much capability transfers.

The numbers: 54% down to 0.2%

The headline figures come from what OpenAI calls an internal adaptive evaluation, in which automated red-team attackers adjust their approach as the defenses respond. Results are averaged over six attacker models.

  • Recovering protected reasoning from an earlier request succeeded 54.0% of the time against GPT-6 Astra's original safeguards, and 0.2% against GPT-6.2 Astra.
  • Recovering reasoning from the current request fell from 22.9% to 0.9%.

Three layers do the work. The models were trained on past attacks to recognise extraction attempts and decline to reveal protected reasoning. Classifiers check inputs for extraction attempts and outputs for leaks, including inside tool-call arguments. And the API now ties protected reasoning to the conversation that produced it: if an earlier message is rewritten, the reasoning that followed is dropped. That closes the route OpenAI described in September, when attackers moved encrypted reasoning from one conversation into another and asked the model to reveal it.

On enforcement, the post says weekly actions against extraction campaigns have ranged from thousands to tens of thousands of accounts, with the rate rising sharply over the summer. It gives no exact counts.

Who OpenAI names, and how carefully

The wording on attribution is hedged, and worth reading closely. OpenAI says it briefed Congress in February on activity "indicating that DeepSeek was attempting" to distill US frontier models. It also repeats its September finding, which attributed a core part of one campaign to "individuals associated with Moonshot AI," the developer of Kimi.

Neither statement says an attempt succeeded. Digital Watch, summarising the September disclosure, notes that OpenAI's figures (a spike of 16,000 requests from more than 4,000 users on 24 and 25 July) describe attempted, not necessarily successful, extractions, and that OpenAI said it was unclear whether one actor was behind all of it. The same summary notes that a separate press report claiming recent DeepSeek and Moonshot models leaned heavily on GPT-6 Astra outputs has not been tied to that campaign.

Why this matters for open-weight models

OpenAI's post never mentions open-weight models, but they sit close to the story. DeepSeek and Moonshot are among the most widely used publishers of downloadable weights, and many people who run models locally use their releases. Three points follow.

First, nothing in the post shows that any specific open model was built from OpenAI outputs. Treat that as an allegation until evidence is published.

Second, the post is partly a policy argument. OpenAI says it supports legislation giving the US government more tools to identify and sanction foreign actors that extract capabilities from American models, and backs stronger export controls on training hardware. Rules written with distillation in mind could affect how freely some open-weight models circulate.

Third, the practical advice it offers other developers is sound whatever one thinks of the politics: protect private reasoning across the whole interaction, keep attack logs as regression tests, and pair per-request defenses with investigation of account networks. All of the reported gains still rest on OpenAI's own unpublished tests.

CuriousLM runs supported AI models locally on your device. Try CuriousLM.