OpenAI / 文章
提取受保护推理的成功率,从54.0%降到了0.2%
OpenAI用自适应黑盒红队测试GPT-6.2 Astra的更新防护,并报告完整推理轨迹比摘要更容易把网络安全能力传给学生模型。评测是内部进行的。
乔木改写
在内部自适应评测里,攻击方要恢复模型在回答更早一个提示时生成的受保护推理。对沿用原有防护的GPT-6 Astra,平均攻击成功率是54.0%;对采用更新防护的GPT-6.2 Astra,这个数字降到了0.2%。
2026年11月20日,OpenAI发布文章,讨论他们如何检测和打断对抗性蒸馏,如何加强对受保护推理的防护,以及他们认为哪些经验可以帮助行业加固防线。

媒体未能载入,请通过下方原文链接查看。
官方把对抗性蒸馏写成一条路径:有人可以未经授权地使用前沿模型的输出,去训练另一个能力相近的模型,再在没有相当防护的情况下部署。他们举的例子是,攻击者蒸馏出模型的网络能力,用来协助实施网络攻击,而原来的模型会拒绝提供帮助,或其防护会拦住这类请求。
蒸馏本身是常见的机器学习方法,学生模型可以从教师模型的答案、推理,或对答案好坏的判断里学习。OpenAI写,除了知识产权,未经授权的蒸馏还有安全风险:第二个模型可能学到前沿能力,却不继承原来那些限制使用方式的防护。在网络安全和生物学这类两用领域,这一点更明显。帮助防御者找软件漏洞的模型,也可能帮助攻击者利用漏洞;协助生物学研究的模型,也可能帮助人研制有害病原体。如果这些能力转到防护更弱的模型上,它可能提供原模型会拒绝、或原有防护会拦住的帮助。
他们特别强调,要防止攻击者提取受保护推理,也就是模型在给出答案前内部生成的中间步骤,亦即思维链。这些轨迹可能包含最终答案里没有的信息,因而对训练另一个模型特别有价值。GPT-6.2 Astra和GPT-6.2 Sol都加强了这类防护。
他们还做了一项内部研究,看网络能力如何转移。产生训练数据的叫教师,学习的叫学生。可见给用户的内容,包括最终回复、推理摘要和工具交互,被当作蒸馏基线;再比较学生如果能得到教师的精确推理轨迹,会多转移多少网络能力。
每次训练都从同一个预训练学生模型开始。监督微调数据主要来自Astra级教师模型在编程和防御性网络安全任务上的轨迹,每条轨迹包含模型完成任务的全部步骤,包括工具交互和最终回复。微调之后,他们又在网络安全和编程环境里做了一轮轻量强化学习,并分别评测。
监督微调之后,用精确推理轨迹训练的学生,总体上优于用推理摘要训练的学生。强化学习之后,在文中展示的两项基准上,精确轨迹学生在测试过的最大推理预算上仍然更高,摘要训练的学生在更低预算上更高。OpenAI认为,保护推理轨迹可以限制网络能力转移,即便教师原来做的是使用政策允许的任务。大量这类交互累积起来,精确推理轨迹可能让学生学到超出教师原任务范围的技能。他们说,还会继续研究不同学生模型和不同蒸馏配方会多转移多少能力。
文中附了一条脚注:Anthropic的相关工作在软件工程基准SWE-bench Verified上,也曾报告完整推理转录比摘要更有优势。OpenAI写,他们的结果把这种优势延伸到了前沿网络安全。脚注指向的Anthropic原文链接在这次页面抓取里没有保留。
防护分两层。请求层用模型行为、分类器和API控制,限制推理轨迹被提取。战役层则看请求和账户层面的信号,寻找协同提取的模式。调查里发现的攻击,会再拿去训练和测试防护。
请求层里,对抗训练让模型识别提取推理的企图并拒绝泄露受保护推理,同时保留对正当解释和帮助请求的回复。训练例子来自内部自动红队、调查,以及外部研究者或行业伙伴负责任披露后确认的成功攻击。
输入分类器用来发现提取企图,输出分类器用来发现助手消息和工具调用参数里的泄露,流式和非流式回复都覆盖。官方强调工具调用这一层,是因为模型可能把受保护推理放进工具参数,而不写进最终答案。输出分类器可以在这些内容到达调用方之前拦住。
API层限制受保护推理如何在不同交互和模型之间被复用。他们提到9月的披露:有人把加密推理从一次对话挪到另一次对话,再要求模型用用户能读的形式还原。防护随后在用户、工作区、组织和模型家族边界上加强,包括关闭一条可能暴露其他用户加密推理的路径。
对GPT-6.2 Astra和GPT-6.2 Sol,受保护推理会绑在产生它的对话历史上。新消息加上去时,模型可以接着用先前的推理;如果更早的一条消息被改写,API会丢弃那条消息之后的推理。他们还阻止防护较弱的旧模型家族读取这两个模型的受保护推理。
战役层上,一旦系统发现未经授权的对抗性蒸馏,通常也包括协同提取推理,他们会调查相关账户,并在确认违规后按政策和条款对账户网络执行处置。每次调查也会用来升级检测。他们还把入侵指标分享给其他模型提供方。文章提到,一致标准政策会在提供其模型的合作方之间设立共享的防护和协同执行标准。
为了测量请求层防护合在一起好不好用,他们在尽量接近生产部署的设置里测试。静态评测会重放一组先前成功的攻击,看哪些还能露出受保护推理。自适应评测里,自动AI红队会根据系统反应调整方法,可能发现静态测试漏掉的弱点。测试者能看到的信息也分两种:黑盒相当于外部用户的视角,白盒会额外提供系统信息,以便构造更强攻击。
在自适应黑盒评测里,他们测试攻击者能不能恢复模型在回答更早请求或当前请求时生成的受保护推理。更新防护后的GPT-6.2 Astra,两类攻击的平均成功率都低于1%,都明显低于沿用原有防护的GPT-6 Astra。他们写,现在也把部分改进用回GPT-6 Astra。
具体数字如下:
| 攻击类型 | 沿用原有防护的GPT-6 Astra,自适应提取成功率(%) | 采用更新防护的GPT-6.2 Astra,自适应提取成功率(%) |
|---|---|---|
| 从先前请求提取受保护推理 | 54.0 | 0.2 |
| 从当前请求提取受保护推理 | 22.9 | 0.9 |
这是对6个攻击模型汇总后的结果。成功率越低越好。文中可能还有配套示意图,这次除社交卡片外,没有从页面取出其他配图。网络能力转移那一节提到“文中展示的两项基准”,图也未能取出,数字以正文为准。
官方还写,今年他们在平台上看到过多起未经授权的对抗性蒸馏,并在9月的报告里分享过初步发现。每周针对这类活动的执行,规模从数千个账户到数万个账户不等,今夏明显加快。他们跟踪滥用的普遍程度,以及从发现到执行的时间,目标是压缩蒸馏者能拿到的数据量和价值,提高战役成本,并缩短对新攻击的反应时间。
在国家安全部分,OpenAI写,外国行动方如果从美国模型复制出先进能力,就可以在没有原开发者防护的情况下部署。他们说,2月曾向美国国会简报,活动显示DeepSeek试图蒸馏美国前沿模型;9月报告打断过一起提取受保护推理的战役,并把其中核心部分归给与Moonshot AI(Kimi的开发方)有关的个人。他们计划很快再分享最近打断的战役。这些归出来自OpenAI自己的调查和表述。
他们还写,支持让美国政府有更多工具去识别和制裁非法提取美国AI模型能力的外国行动方,并向国会表示支持Chip Security Act、AI OVERWATCH Act和MATCH Act。在他们看来,更强的出口管制会限制对手获得训练前沿模型所需的硬件,蒸馏防护则让人更难从已经训练好的模型里取得部分同类能力。这两件事被他们写成有助于维持美国在前沿AI上的领先,从而给开发者更多空间,在需要时间测试防护时放慢发布。以上是OpenAI的立场和政策主张,不是独立核实后的结论。
写给其他开发者和平台的建议,原文列了四条:在完整交互里保护私有推理,检查助手消息和工具参数,并限制加密推理在对话和模型之间的复用;把红队、生产监控或负责任披露里成功的提取攻击留下来做回归测试,在符合隐私政策的前提下用来改进训练和分类器,更新防护后既要重放原攻击,也要看自适应红队还能不能打穿;把请求层缓解和战役层调查、执行结合起来;与部署伙伴协调防护要求和执行,并与其他模型开发者分享经验,以提高行业的防护下限。
他们此前也通过Frontier Model Forum和相应的政府信息共享渠道分享过发现,并写明会把攻击者可能利用的细节留在非公开范围。文章最后说,对抗性蒸馏防护的目标,是降低前沿能力在没有足够防护的情况下被复制的风险,以支持他们把通用人工智能安全部署、并让全人类受益的使命。
这些数字来自OpenAI的内部评测和平台执行记录。54.0%降到0.2%,比较的是特定自适应黑盒设置里的提取成功率,不能直接读成外部世界里蒸馏已经被阻止。完整推理轨迹会多转移多少网络能力,也仍取决于学生模型和训练配方,官方自己也说还会继续研究。
原文
Combatting adversarial distillation

媒体未能载入,请通过下方原文链接查看。
Combatting adversarial distillation
November 20, 2026

How we're addressing safety risks of distillation and lessons we've learned for strengthening safeguards across the industry.
As frontier AI models become more capable, safeguards play an increasingly important role in how they can be used responsibly. Adversarial distillation creates a path for bad actors to reproduce these capabilities without appropriate safeguards. An actor may use a frontier model's outputs without authorization to train another model with similar capabilities, then deploy it without comparable protections. For instance, an adversary might distill a model's cyber capabilities and use the resulting model to help carry out cyberattacks that the original model would have refused to assist with or that its safeguards would have blocked.
We're sharing how OpenAI detects and disrupts adversarial distillation, how we've strengthened protections for our models' protected reasoning, and what we've learned that can help strengthen defenses across the industry.
Our approach to adversarial distillation combines adversarial testing, model training, classifiers, and API-level protections with investigations into coordinated attempts to extract model capabilities. These updated safeguards align more closely with those we use for biology and cybersecurity.
We place particular emphasis on preventing attackers from extracting our models' protected reasoning: the intermediate steps a model generates internally while working toward an answer, known as its chain-of-thought. These traces can contain information the final answer leaves out, making them especially valuable training data for another model.
We've strengthened our defenses against protected reasoning extraction in GPT-6.2 Astra and GPT-6.2 Sol. In an internal adaptive evaluation using strong automated red-teaming, we tested whether attackers could recover protected reasoning the model had generated while answering an earlier prompt. The average attack success rate fell from 54.0% against GPT-6 Astra with its original safeguards to 0.2% against GPT-6.2 Astra with its updated safeguards.
Why distillation is a safety issue
Distillation is a common machine-learning technique in which one model learns from data produced by another. For example, a model can learn from another model's answers, its reasoning, or its judgments about which answers are better and why. Beyond intellectual property concerns, unauthorized distillation poses serious safety risks. Through adversarial distillation, a second model can learn advanced frontier capabilities without inheriting the safeguards that govern how the original model can use them.
The concern is especially clear in dual-use domains like cybersecurity and biology. A model that helps defenders find vulnerabilities in software could also help an attacker exploit them; a model that assists with biological research could also help someone develop a harmful pathogen. If such capabilities were transferred to a model with weaker safeguards, it may provide assistance that the original model would have refused, or that the original safeguards would have blocked.
The importance of protecting exact reasoning traces for cyber safety
Our models help users with coding and defensive cybersecurity tasks. A second model trained on many of those benign outputs may learn skills that also apply to offensive cybersecurity tasks. To better understand this, we study how distillation, especially involving reasoning traces, may transfer risky cyber capabilities.
In this setting, the model that produces the training data is called the teacher, and the model that learns from it is called the student. In an internal study on how cyber capabilities transfer, we used the content normally visible to users—final responses, reasoning summaries, and tool interactions—as our baseline for distillation, and tested how much additional cyber capability transfers when the student has access to the teacher's exact reasoning traces.
We started each training run from the same pretrained student model. We created a supervised fine-tuning dataset consisting primarily of trajectories from Astra-level teacher models working on coding and defensive cybersecurity tasks. Each trajectory includes all of the steps a model takes as it works through a task, including any tool interactions and its final response. We evaluated each student after fine-tuning and again after a lightweight round of reinforcement learning in cybersecurity and coding environments.
After supervised fine-tuning, the student trained on exact reasoning traces generally outperformed the student trained on reasoning summaries. After 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.
Our cyber distillation results suggest that protecting reasoning traces can limit cyber capability transfer, even when the underlying task is one allowed under our usage policies. Across many such interactions, exact reasoning traces can teach a student skills that generalize beyond the tasks the teacher originally performed. We will continue studying how much additional capability transfers across different student models and distillation training recipes.
Defenses that reinforce one another
We defend against adversarial distillation at two complementary levels of our safeguard stack. At the request level, model behavior, classifier checks, and API controls limit reasoning trace extraction. At the campaign level, we look for patterns in request-level and account-level signals that can reveal coordinated extraction attempts. In turn, attacks uncovered in investigations become examples for training and testing our safeguards.
Our defenses against adversarial distillation combine safeguards for individual requests with campaign detection and enforcement. Findings from investigations and automated red-teaming help strengthen those safeguards.
Protecting individual requests and responses
Defensive training for model behavior
Adversarial training teaches models to recognize reasoning extraction attempts and refuse to disclose protected reasoning, while preserving responses to legitimate requests for explanations and assistance. Adversarial training examples come from successful attacks identified through internal automated red-teaming, investigations, and responsible disclosures of potential security weaknesses from outside researchers or industry partners.
Classifier safeguards
We use input classifiers to detect attempts to extract protected reasoning, and output classifiers to detect leaks of protected reasoning in assistant messages and tool-call arguments. These safeguards apply to both streaming and non-streaming responses.
Tool-call coverage is important because a model can expose protected reasoning in a tool's arguments without including it in its final answer. Output classifiers can block suspected leaks before that content reaches the caller.
API-level protections
API-level protections restrict how protected reasoning can be reused across interactions and models. Our September disclosure described attacks that moved encrypted reasoning from one conversation into another and asked the model to reveal the reasoning in user-readable form. We strengthened protections across user, workspace, organization, and model-family boundaries, including closing a pathway that could expose another user's encrypted reasoning.
With GPT-6.2 Astra and GPT-6.2 Sol, we tie a model's protected reasoning to the conversation history that produced it. As new messages are added, the model can build on its earlier reasoning. If an earlier message is rewritten, our API drops the reasoning that followed it. We also prevent older model families with weaker safeguards from reading GPT-6.2 Astra and Sol's protected reasoning.
Investigating and disrupting coordinated campaigns
When our systems detect unauthorized adversarial distillation campaigns, which often include coordinated attempts to extract our models' reasoning, we investigate the accounts responsible. When we identify violations, we enforce against the network of accounts, consistent with our policies and terms of service. Through investigations of each campaign, we generate insights that allow us to upgrade our defenses in order to better detect similar activity. We also share indicators of compromise with other model providers so they can investigate related abuse on their own services.
This work extends beyond our own platform. Our consistent-standards policy establishes shared standards for safeguards and coordinated enforcement across providers offering our models. We work closely with partners where appropriate to uphold these standards and strengthen protections across deployments.
Sharing relevant findings across industry and government where appropriate can help disrupt campaigns and enable model developers and providers to better identify and defend against attempts to distill frontier models.
Measuring progress
Resistance to reasoning extraction
To measure how well our request-level safeguards work together, we test whether attackers can still extract protected reasoning in a setup that closely mirrors production deployments of our models and safeguards. In static evaluations, we replay a fixed set of previously successful attacks to see which still expose protected reasoning. In adaptive evaluations, automated AI red-teamers adjust their approach based on how the system responds, which can uncover weaknesses the static tests miss.
We also vary what testers can see. Black-box testing gives them the same view an external user would have, while white-box testing gives them additional information about the system to help them develop stronger attacks. In our adaptive black-box evaluation, we tested whether attackers could recover protected reasoning the model had generated while answering an earlier request or the current one. On GPT-6.2 Astra with its updated safeguards, the average attack success rate was below 1% for each attack type. Both rates were substantially lower than on GPT-6 Astra with its original safeguards. We're now applying selected safeguard improvements to GPT-6 Astra as well.
| Attack type | GPT-6 Astra defender with original safeguards Adaptive extraction ASR (%) | GPT-6.2 Astra defender with updated safeguards Adaptive extraction ASR (%) |
|---|---|---|
| Extracting protected reasoning from an earlier request | 54.0 | 0.2 |
| Extracting protected reasoning from the current request | 22.9 | 0.9 |
Attack success rate (ASR) in adaptive black-box red-teaming against GPT-6 Astra and GPT-6.2 Astra, aggregated across six attacker models. Lower ASR is better.
Campaign detection and enforcement
We have seen multiple unauthorized adversarial distillation campaigns on our platform this year, and shared initial findings in our September report. Our weekly enforcement actions against these campaigns have ranged in scale from thousands to tens of thousands of accounts, with the rate increasing significantly this past summer. In order to measure the effectiveness of our efforts, we also track the prevalence of distillation abuse, as well as the time between detection and enforcement. Our objective is to constrain the volume and value of the data distillers are able to obtain, to increase the cost of distillation campaigns, and to shorten the time needed to respond to new attacks. We continue to improve these defenses through adversarial testing and investigations.
The national security implications of adversarial distillation
Adversarial distillation also poses a national security risk. A foreign actor that reproduces advanced capabilities from a U.S. model can deploy them without the protections its developer put in place. In February, we briefed Congress on activity indicating that DeepSeek was attempting to distill U.S. frontier models. In September, we reported disrupting a campaign to extract our models' protected reasoning and attributed a core part of that activity to individuals associated with Moonshot AI, the developer of Kimi. We plan to share more soon about recent campaigns we've disrupted.
We support legislation that would give the U.S. government more tools to identify and sanction foreign actors that illicitly extract capabilities from American AI models. We've also told Congress we support the Chip Security Act, the AI OVERWATCH Act, and the MATCH Act. Stronger export controls can limit adversaries' access to the hardware needed to train frontier models. Distillation defenses make it harder to acquire some of the same capabilities from models already trained. Both efforts help maintain a U.S. lead in frontier AI, giving developers more room to slow a release when they need time to test safeguards.
How others can build on this work
Supporting the safe proliferation of frontier capabilities is an important goal of our anti-distillation work. Our experience so far suggests several useful steps that other model developers and platforms can take:
- Protect private reasoning across the full interaction. Check assistant messages and tool-call arguments for leaks, and restrict how encrypted reasoning can be reused across conversations and models.
- Keep successful extraction attack logs found through red-teaming, production monitoring, or responsible disclosure as regression tests. Where appropriate, and consistent with applicable privacy policies, use those examples to improve model training and classifiers. When evaluating updated defenses, replay the original attacks and test whether an adaptive red-teamer can still find a way through those defenses.
- Combine request-level mitigations with campaign-level investigation and enforcement.
- Coordinate safeguard requirements and enforcement with deployment partners, and share expertise with other AI model developers, to ensure a high floor of distillation safeguards across the industry.
We have previously also shared findings through the Frontier Model Forum and appropriate government information-sharing channels to help other developers strengthen their defenses while keeping the details attackers could exploit private. We'll continue working with researchers, developers, and partners to share what we learn about emerging attacks and effective defenses.
As we make frontier AI broadly available, our defenses against adversarial distillation aim to reduce the risk that its capabilities are reproduced without adequate safeguards, supporting our mission to ensure that artificial general intelligence is deployed safely and benefits all of humanity.
Footnotes
- Related work from Anthropic reported a similar advantage when training on full reasoning transcripts rather than summaries on SWE-bench Verified, a software engineering benchmark. Our result shows this advantage extends to frontier cybersecurity.