Close-up of a computer screen displaying HTML, CSS, and JavaScript code
LabsBased on company claims

OpenAI says GPT-6.2 cut reasoning-extraction attacks from 54% to 0.2%, in its own tests

By Himanshu Sakre·

Photo: Саша Алалыкин / Pexels

OpenAI's November 20 post on adversarial distillation reports a sharp drop in attacks that pull hidden reasoning from GPT-6.2 Astra. The numbers are internal and the attribution to Moonshot AI is unproven.

OpenAI says it has cut the success rate of attacks that try to pull hidden reasoning out of its newest models from 54.0% to 0.2%, a result that matters because adversarial distillation is now a stated national security concern for US labs. The company published the numbers on November 20, 2026, as part of a post on how it detects and disrupts the practice, which it describes as one model learning from another model's outputs without that model's safeguards. The headline figures are internal and self-reported, so the useful question is what they measure and what they leave out.

What adversarial distillation is, in OpenAI's terms

Distillation itself is routine: a smaller "student" model is trained on the outputs of a stronger "teacher." It becomes adversarial when the student's builder has no permission and the outputs carry capabilities the teacher's owner deliberately withholds. OpenAI's November 20 post says the concern is that an adversary might distill a model's cyber capabilities, and it focuses on the protected reasoning, the model's internal working record, because that record can reveal information the final answer does not.

That framing explains why the company protects reasoning more tightly than answers. A student trained on exact reasoning traces may learn more than one trained on the short summaries users normally see. OpenAI ran an internal study to test that, starting every run from the same pretrained student and fine-tuning on mostly Astra-level teacher trajectories for coding and defensive cybersecurity tasks. It compared exact traces against summaries, then added a lightweight round of reinforcement learning. In a footnote, OpenAI notes that Anthropic reported a similar advantage for full reasoning transcripts over summaries on SWE-bench Verified.

The numbers, and how they were produced

The post reports two attack types, both measured with adaptive black-box red-teaming aggregated across six attacker models. Adaptive means automated red-teamers adjust their prompts to the system's responses, which is harder to defend than replaying old attacks.

  • Extracting protected reasoning from an earlier request: 54.0% attack success on GPT-6 Astra with its original safeguards, 0.2% on GPT-6.2 Astra with the updated ones.
  • Extracting protected reasoning from the current request: 22.9% on GPT-6 Astra, 0.9% on GPT-6.2 Astra.

OpenAI says average attack success on the updated model was below 1% for each attack type. The defenses it lists are adversarial training, input and output classifiers that also cover tool-call arguments, and API-level restrictions that tie reasoning to the conversation history and block reuse across users and across models.

“we attribute a core cluster of the activity to individuals associated with Moonshot AI, the developer of Kimi.”

OpenAI, Disrupting a coordinated model distillation campaign, September 30, 2026
Engineer working on prosthetic design using specialized software in an office setting
Engineers monitor systems in a security operations room. Photo: ThisIsEngineering / Pexels

What the post does not tell you

The limits are as important as the percentages. The red-team results are internal and self-reported, and the post does not name the six attacker models or give full evaluation details. It reports only the black-box setup, although it says testing also covered white-box conditions with extra system information. The cyber transfer study is described only qualitatively; the text gives no numeric scores for the ExploitBench or SEC-bench Pro results, which appear only in charts. And the campaign-level enforcement figures are vague: weekly actions have ranged from thousands to tens of thousands of accounts, with the rate rising significantly over the summer, but no exact counts are published.

So the 54.0% to 0.2% drop is a measure of one defense against one family of automated attacks, run by the defender. It is not an independent finding that reasoning extraction is solved. A determined state-backed operation uses real accounts, patience and human adaptation that an automated red team only approximates.

The earlier campaign this builds on

The November post sits on a prior disclosure. On September 30, 2026, OpenAI reported disrupting a coordinated campaign to extract protected reasoning. Activity began on July 1 at low volume, then spiked on July 24 and 25, with about 16,000 requests from more than 4,000 users. OpenAI notes these were attempted extractions, not necessarily successful ones, and that related prompt-pattern activity spanned a cluster of more than 15,000 users. The cluster was fully disrupted by July 28.

The operators did not break encryption or access stored conversations. According to OpenAI, they manipulated model interactions so hidden reasoning appeared in a form the requester could see, including copying encrypted reasoning from one conversation and having a model in another conversation decrypt it. OpenAI closed that replay pathway. It wrote that it attributes "a core cluster of the activity to individuals associated with Moonshot AI, the developer of Kimi," while saying it is unclear whether all operators were one actor. The November post adds that OpenAI briefed Congress in February that DeepSeek was attempting to distill US frontier models. Those are OpenAI's assertions; the posts we read do not publish technical evidence for the attribution, and this desk has not seen a response from either company.

The pattern links to a wider theme on this site: stolen access and fraudulent accounts as the raw material for copying models. Our earlier report on a Railway breach and 100,000 workspaces tied to distillation covered the account-fraud side of the same problem.

Why it matters for the race

If frontier capability can be copied through an API, then export controls on chips are only part of the picture, and the labs' own safeguards become a form of defense policy. That is the reading OpenAI is inviting when it mentions national security concerns and recommendations for other developers. It also shifts a design choice: showing users a summary instead of raw reasoning is no longer only a product decision.

The counterweight is that the evidence is thin where it counts. We have two company posts, a before-and-after table the company ran itself, and an attribution without published proof.

Our take

Treat the 0.2% as a credible sign that the specific replay and extraction holes OpenAI found are closed, not as proof the problem is over. What we would watch next: whether OpenAI or an outside evaluator publishes the attacker models and the white-box results, whether the promised follow-up on recent disrupted campaigns arrives with named evidence, and whether Anthropic and Google publish comparable numbers so the claim can be compared across labs.

Frequently asked questions

What is adversarial distillation?

It is the unauthorized use of one model's outputs to train another model. OpenAI's November 20, 2026 post says the worry is that an adversary might copy a model's capabilities, such as cyber skills, without the safeguards the original model carries.

What did OpenAI report about GPT-6.2 Astra?

In adaptive black-box red-teaming across six attacker models, extraction of protected reasoning from an earlier request fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra. For the current request it fell from 22.9% to 0.9%.

Are OpenAI's results independently verified?

No. The post describes internal, self-reported tests. It does not name the six attacker models or publish full evaluation details, and the cyber transfer study results appear only in charts.

Who does OpenAI blame for reasoning extraction?

On September 30, 2026 it attributed a core cluster of one campaign to individuals associated with Moonshot AI, while saying it is unclear whether all operators were one actor. In February it briefed Congress that DeepSeek was trying to distill US models. We have not seen published technical evidence.

Why does OpenAI protect reasoning more than answers?

Reasoning traces can reveal information withheld from final answers. OpenAI ran a study comparing exact traces with summaries for training a student model, and notes Anthropic reported a similar advantage for full transcripts on SWE-bench Verified.

Sources

What each one is, and whose it is.

  1. 1

    Combatting adversarial distillation, OpenAI (November 20, 2026)

    Vendor announcement
  2. Vendor announcement
  3. Press reportIndependent of the vendor