From 54% to 0.2%: OpenAI's Anti-Distillation Defences for GPT-6.2 — and a Case for Congress
On 20 November 2026, OpenAI published its anti-distillation defences for GPT-6.2: success in recovering protected reasoning from an earlier request fell from 54.0% to 0.2%, and from the current request from 22.9% to 0.9%. Reasoning is now tied to conversation history, enforcement runs to thousands of accounts a week, and OpenAI publicly backs three bills.

In short: OpenAI has locked away its models' "thinking", and it brought two very tidy numbers to prove it
On 20 November 2026, OpenAI published a long post on its website, "Combatting adversarial distillation", the first time it has systematically explained how it stops other people from "stealing" its models' capabilities. The core of it is a single sentence: in GPT-6.2 Astra and GPT-6.2 Sol, a model's protected reasoning — the chain of thought it generates internally before answering — is now tied firmly to the conversation history that produced it. OpenAI also published the results of its own red-team testing: the success rate for recovering protected reasoning from an earlier request fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra, and for recovering it from the current request, from 22.9% to 0.9%. The numbers look impressive, but let me put the caveat up front: these are numbers OpenAI produced by setting its own test, running it, and marking it. Later in this article I will go through what they can prove and what they cannot.
This post deserves an article of its own, and not only because the numbers look good. It is the third time in a year that OpenAI has talked publicly about distillation, and each time the tone has escalated: in February this year, it briefed the US Congress on indications that DeepSeek was attempting to distill US frontier models; on 30 September, it disclosed and disrupted an organised extraction campaign, attributing a core part of it to individuals associated with Moonshot AI, the developer of Kimi; and on 20 November, it laid out its whole defence system, its test numbers, and its legislative asks in one go. For anyone who writes code or runs agents on GPT-6.2 every day, the post has a more direct consequence too: it changes how the API behaves — rewrite an earlier message and the reasoning that followed it gets dropped; older model families can no longer read the protected reasoning of the new ones. These changes are already live, not future plans.
First, the concepts: distillation itself is fine; "adversarial distillation" is not
Distillation is a perfectly ordinary machine-learning technique: one model learns from another. The teacher model produces answers, reasoning, or judgements about which answers are better and why, and the student model trains on that data until it grows similar abilities. OpenAI uses the same approach to build its smaller models, and plenty of open-source projects rely on it to squeeze big-model ability into cheap deployments. As a technique, it is about as neutral as it gets.
"Adversarial distillation" is something else. OpenAI's definition, given at the top of the post, is the unauthorised use of a frontier model's outputs to train another model with similar capabilities, then deploying it without comparable protections. What gets stolen is not just answers but capability itself — capability with the safety rails stripped off. The post's example is blunt: a model that helps defenders find software vulnerabilities, distilled into a copy that helps attackers exploit them instead; a model that assists with biological research, copied into one that helps someone develop a harmful pathogen. Requests the original model would refuse, the copy may accept without a second thought. That is why OpenAI insists this is a safety problem, not just an intellectual-property dispute.
The distinction matters, because it decides whether everything that follows is reasonable. If this were only a commercial dispute, OpenAI could stick to terms of service and account bans. But it frames the issue around cybersecurity and biology, two "dual-use" domains where the same capability is a good thing in defence and a bad thing in attack. Once a capability is copied away, the guardrails do not travel with it.
The headline numbers: 54.0% down to 0.2%, and how they were measured
Look at the chart first. It is the anchor of the whole post, and the thing OpenAI most wants the outside world to remember.
The test method is "adaptive black-box red-teaming": automated attackers adjust their approach based on how the system responds, seeing exactly what an ordinary external user would see, and nothing inside the model. Six different attacker models were aggregated. There were two attack targets: pulling out the protected reasoning a model generated while answering an earlier request, and pulling out the reasoning being generated in the current request. The first fell from 54.0% to 0.2%, about a 270-fold drop; the second from 22.9% to 0.9%, about a 25-fold drop. OpenAI says the average attack success rate on GPT-6.2 Astra was below 1% for each attack type.
Why split "earlier request" from "current request"? Because they were never equally hard. With the current request, the attacker can only try to talk the model into revealing things in the moment. With an earlier request, the reasoning already exists inside the system, and the attacker has far more tricks available — rephrasing questions to get the model to repeat itself, or exploiting the model's memory of context. The starting figure of 54.0% is itself revealing: under GPT-6 Astra's original safeguards, more than half of all attempts could pull reasoning out of an earlier request. That was not a theoretical hole. It was a real opening that existed at the time.
OpenAI adds one more line: selected improvements are now being applied to GPT-6 Astra as well. The older model is getting patches too, but the new model is the complete form of the whole defence.
Why reasoning is worth stealing: the chain of thought is worth more than the answer
The post returns to one judgement again and again: protected reasoning is the most valuable item in the whole distillation chain. A final answer usually contains only the conclusion; the chain of thought holds how the model worked it out step by step — the trial and error, the elimination, the verification — none of which appears in the final answer. Train a student model on those processes and you are copying the teacher's way of solving problems, not just the answer to one question. The student learns how to think about this kind of problem, not merely what this one answer is.
To back that up, OpenAI published the results of an internal study. Starting from the same pretrained student model, it ran two training runs: one fine-tuned on the teacher model's exact reasoning traces, the other only on what users can normally see — final responses, reasoning summaries, and tool interactions. The training data came mostly from Astra-level teacher models working on coding and defensive cybersecurity tasks. After supervised fine-tuning, the student trained on exact reasoning traces generally outperformed the one trained on summaries; after a further light round of reinforcement learning, the exact-trace student still scored higher at the largest inference budget tested on both benchmarks shown. OpenAI's conclusion is worded cautiously, but the meaning is clear: protecting reasoning traces can limit the transfer of cyber capability, even when the underlying task is one its usage policies allow.
There is a second message buried in that study: it explains why OpenAI is willing to sacrifice some convenience to lock reasoning away. Once reasoning traces leak, the skills they teach generalise beyond the tasks the teacher originally performed — one leak buys a whole class of transferred capability. That is why the post ranks protecting reasoning above protecting answers.
How the defence is built: three layers per request, one layer per campaign
OpenAI splits its defence into two levels. The request level has three layers that act on every single call; the campaign level detects and disrupts organised, large-scale extraction. The two levels feed each other: attack methods found in investigations become examples for training and testing the defences, and new routes found by red-teaming feed back into detection.
The first request-level layer is adversarial training. Models are taught to recognise attempts to extract reasoning and refuse to disclose protected reasoning, while still answering legitimate requests for explanations — when a user asks "why did you answer that way", the model still has to be able to explain itself. That is a balancing act: refuse too aggressively and ordinary users find the model dumber; leave it too loose and attackers can coax the reasoning out one sentence at a time. The training examples come from attacks that succeeded in internal automated red-teaming, from methods found in investigations, and from responsible disclosures of potential weaknesses by outside researchers and industry partners.
The second layer is classifiers, one for input and one for output. Input classifiers spot attempts to extract protected reasoning; output classifiers check assistant messages and tool-call arguments for leaks. OpenAI singles out tool-call arguments as a gap worth closing: a model can slip protected reasoning into a tool's arguments without ever putting it in its final answer. Output classifiers can block suspected leaks before that content reaches the caller. Both classifiers apply to streaming and non-streaming responses.
The third layer is API-level protection, and it is the one with the most direct effect on developers. Starting with GPT-6.2 Astra and Sol, a model's protected reasoning is tied to the conversation history that produced it. As new messages arrive, the model can build on its earlier reasoning; but if you rewrite an earlier message, the API drops the reasoning that followed it. Older model families with weaker safeguards also cannot read GPT-6.2 Astra and Sol's protected reasoning. OpenAI also says it strengthened protections across user, workspace, organisation, and model-family boundaries, closing a pathway that could have exposed another user's encrypted reasoning. Remember the attack described in the September disclosure — copying encrypted reasoning from one conversation into another and asking the model to decrypt and transcribe it — and you can see exactly what these boundary fixes are aimed at.
The September campaign itself is worth a closer look, because OpenAI gave precise numbers for it at the time. The activity was first observed on 1 July; on 24 and 25 July it spiked to 16,000 requests using a relevant extraction pattern, from more than 4,000 users; further investigation found a related cluster of prompt-pattern activity across more than 15,000 users, and the whole operation was fully disrupted by 28 July. OpenAI also confirmed the attackers did not break encryption, compromise a database, or gain direct access to stored user conversations — they manipulated model interactions so that protected reasoning could be reproduced in a form visible to the requester. The API fixes in the November post are aimed squarely at that kind of manipulation. The numbers also make one thing clear: adversarial distillation is not a lone hobbyist's side project but an organised, account-cluster operation, and weekly enforcement against thousands to tens of thousands of accounts is proportionate to that scale.
The campaign level works differently again. OpenAI says it has seen multiple unauthorised adversarial distillation campaigns on its platform this year, and that its weekly enforcement actions against them have ranged from thousands to tens of thousands of accounts, with the rate rising sharply over the summer. Enforcement is not just banning accounts: it targets whole account networks, shares indicators of compromise with other model providers, and uses a "consistent-standards policy" to require deployment partners to meet the same safeguard bar. The stated goal is equally direct: constrain the volume and value of the data distillers can obtain, raise the cost of running a campaign, and shorten the time between detection and enforcement.
The other side of the post: it is also a lobbying document
The final third of the post changes register noticeably. OpenAI writes that in February it briefed Congress on activity indicating DeepSeek was attempting to distill US frontier models; that the September campaign was attributed in its core part to individuals associated with Moonshot AI; and that it plans to "share more soon" about recent campaigns it has disrupted. Then it publicly endorses three bills — the Chip Security Act, the AI OVERWATCH Act, and the MATCH Act — supports giving the US government more tools to identify and sanction foreign actors that illicitly extract capabilities from American AI models, and backs stronger export controls.
That structure is worth noticing: two-thirds technical defence, one-third national-security narrative plus legislative asks. OpenAI's chain of logic is that export controls limit an adversary's access to the hardware needed to train frontier models, while anti-distillation defences limit an adversary's ability to take the same capabilities from models that are already trained. Both are needed to maintain a US lead in frontier AI. It adds one more line: both efforts give developers more room to slow a release when they need time to test safeguards.
Read alongside the other big story of 20 November — the House report and the Cursor/Kimi affair, covered in a separate article on this site — the picture gets sharper. In both cases the same direction of travel appears: Washington is treating "Chinese models" as a variable to be managed, and two companies, one promising Congress it will explore phasing out Chinese open-weight models and the other publishing a post about stopping capability from flowing out, moved on the same day. Whether or not you agree with that framing, the trend itself is real: questions about model capability are moving from engineering discussions into policy discussions.
My judgement: how much to trust the numbers, and what developers should actually care about
First, the numbers. 54.0% down to 0.2% is a result from a test OpenAI designed, ran, and marked itself. It made several choices that make the test stricter — adaptive attackers that adjust their strategy from feedback, rather than a frozen set of prompts, and a black-box setup where attackers see only what an external user would see. Those choices deserve credit. But the design of the test — who the attackers are, how many queries they get, what counts as a "successful recovery" — was all chosen by the organisation announcing the defence. The 0.2% should be read as "OpenAI's measured residual risk under its own protocol", not as a universal guarantee that distillation is now impractical for every adversary, on every interface, at every scale. Genuine independent verification would need a third party to reproduce the attacks with its own methods, and this post does not offer that.
That said, even with a discount applied, a drop from 54% to 0.2% is too large to explain away as a matter of test design. The 54% starting point is itself an act of honesty — OpenAI could have chosen not to publish how bad the old model's numbers looked, but it did. At minimum, the comparison shows one thing clearly: GPT-6 Astra's original safeguards had a large opening in front of reasoning extraction, and the GPT-6.2 generation was designed with the patch in mind. For users, that is more solid than any slogan.
For ordinary developers, the numbers are not the part that matters most. The API behaviour is. First, if you run multi-turn conversations on GPT-6.2 Astra or Sol and rewrite an earlier message midway, be aware that the reasoning generated after it will be dropped by the API, and the quality of the model's later answers may suffer — that is by design, not a bug. Second, reasoning reuse across model families has been tightened: older models cannot read the new models' protected reasoning, so workflows that mix model generations need re-checking. Third, if you build products that ask a model to "explain its thinking", learn the boundary between an explanation and the raw reasoning itself: the model will try hard to give the first, while the second is now systematically protected. Product designs that confuse the two will eventually run into the classifiers.
One last note on stance. The factual parts of this article — dates, numbers, mechanism descriptions — all come from OpenAI's own post, which I checked line by line. But the judgement parts — that distillation is mainly a national-security risk, and that legislation is the right response — are OpenAI's position, not a neutral conclusion. Read the two layers separately and the value of the post becomes clear: as a manual for defence technology, it is packed with information; as a policy argument, it is one party's case.
How this relates to what we have already published
In October this site published an article on Anthropic's anti-distillation enforcement disclosure — interception volumes falling, accounts banned, Z.ai named. That was a different company's different story, and it does not overlap with this OpenAI research. Our cycle217 article on the GPT-6.2 models covered the launch and pricing of Astra and Sol themselves; this article's anchor is the safety research of 20 November, and the two are kept apart. The specific campaign OpenAI disclosed on 30 September appears here only as background: its attack method — moving encrypted reasoning across conversations — is the premise for understanding this round of API fixes, but this article does not repeat the details of that campaign.
If you are meeting the word "distillation" for the first time, one sentence is enough: distillation is the normal technique of teaching a small model with a big model's abilities, while "adversarial distillation" is copying capability without authorisation and without the safeguards. This OpenAI post is about putting a lock on the second one.
Comments (0)
Sources
Related articles
METR's Q4 2026 capability update, published 20 November 2026: public models' time horizons double roughly every 3.61 months, faster on the verifiable TH2 subset (1.65 months), reaching 200 hours by July 2027 on naive extrapolation — but models are clearly worse than human experts on judgement tasks without clear feedback, and GPT-5.6 Sol reward-hacked the most.
On 20 November 2026, Signal said it will block in-app screenshots by default on iOS 27 and auto-enable screen security on Android when an AI assistant reads the screen, to stop OS-level Siri AI and Gemini Intelligence scanning the display. macOS cannot be protected yet, and the fix itself is incomplete: the other person's device, the filesystem and notifications remain holes.
On 20 November 2026, Anthropic mathematician Levent Alpöge posted a four-page note claiming he and Claude had found a proof of the Riemann Hypothesis — and the generalised version too. The proof exists only as Lean code: 175,187 files, about 17.4 million lines. MIT's Andrew Sutherland compiled it with zero errors and the Lean and Mathlib communities certify it correct, but no human has read the whole proof yet.