OpenAI details how it blocks adversarial distillation in GPT-6.2 Astra

OpenAI has published a post on how it detects and disrupts adversarial distillation: an actor uses a frontier model's outputs without authorization to train another model with similar capabilities, then deploys it without comparable protections. OpenAI's example is an adversary who distills a model's cyber capabilities and uses the result to help carry out cyberattacks the original model would have refused to assist with. The post covers the detection and disruption work, the stronger protections for the models' protected reasoning, and lessons OpenAI says can help defenses across the industry.
The focus is protected reasoning, the intermediate steps a model generates internally while working toward an answer, known as its chain-of-thought. OpenAI says these traces can contain information the final answer leaves out, which makes them especially valuable training data for another model. It has strengthened defenses against extraction of this reasoning in GPT-6.2 Astra and GPT-6.2 Sol. In an internal adaptive evaluation with strong automated red-teaming, testers tried to 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.
OpenAI also ran an internal study of how cyber capabilities transfer. The teacher model produces the training data and the student learns from it. The baseline was what users normally see: final responses, reasoning summaries and tool interactions. OpenAI then tested how much extra capability transfers when the student also gets the teacher's exact reasoning traces. Every run started from the same pretrained student model. The supervised fine-tuning dataset was made mostly of trajectories from Astra-level teacher models working on coding and defensive cybersecurity tasks, including tool interactions and final responses. Each student was evaluated after fine-tuning and again after a lightweight round of reinforcement learning in cybersecurity and coding environments. After fine-tuning, the student trained on exact traces generally outperformed the one 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. OpenAI reads this as a sign that protecting reasoning traces can limit cyber capability transfer, even when the underlying task is allowed under its usage policies, and says it will keep studying how much transfers across other student models and training recipes.
The defense works at two levels. At the request level, model behavior, classifier checks and API controls limit reasoning extraction. At the campaign level, OpenAI looks for patterns in request-level and account-level signals that reveal coordinated attempts, and attacks found in investigations become training and test examples. The pieces are:
Adversarial training teaches models to recognize extraction attempts and refuse to disclose protected reasoning while still answering legitimate requests for explanations. Its examples come from successful attacks found through internal automated red-teaming, investigations, and responsible disclosures from outside researchers or industry partners.
Input classifiers detect attempts to extract protected reasoning. Output classifiers detect leaks of it in assistant messages and tool-call arguments, in both streaming and non-streaming responses. Tool calls matter because a model can expose protected reasoning in a tool's arguments without putting it in the final answer, and output classifiers can block suspected leaks before they reach the caller.
API-level protections restrict how protected reasoning can be reused across interactions and models. OpenAI's September disclosure described attacks that moved encrypted reasoning from one conversation into another and asked the model to reveal it in readable form. OpenAI 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, the model's protected reasoning is tied to the conversation history that produced it. If an earlier message is rewritten, the API drops the reasoning that followed it. Older model families with weaker safeguards are also prevented from reading the protected reasoning of GPT-6.2 Astra and Sol.
When systems detect unauthorized campaigns, OpenAI investigates the accounts, enforces against the whole network when it finds violations, uses each investigation to upgrade detection, and shares indicators of compromise with other model providers. It also points to a consistent-standards policy that sets shared safeguard and enforcement standards across providers offering its models.
For measurement, OpenAI tests in a setup that closely mirrors production. Static evaluations replay a fixed set of previously successful attacks. Adaptive evaluations use automated AI red-teamers that change tactics based on the system's responses. Black-box testers see what an external user would see; white-box testers get extra information about the system. In the adaptive black-box evaluation, testers tried to recover reasoning generated while answering an earlier request or the current one. On GPT-6.2 Astra with updated safeguards, the average attack success rate was below 1% for each attack type, aggregated across six attacker models. OpenAI says both rates were substantially lower than on GPT-6 Astra with its original safeguards, and it is now applying selected safeguard improvements to GPT-6 Astra too.
OpenAI says it has seen multiple unauthorized distillation campaigns on its platform this year. Its weekly enforcement actions have ranged from thousands to tens of thousands of accounts, with the rate rising significantly this past summer. It also tracks the prevalence of distillation abuse and the time between detection and enforcement, and aims to limit the volume and value of data distillers can get, raise the cost of campaigns and shorten response time.
The post also frames distillation as a national security risk. In February OpenAI briefed Congress on activity indicating that DeepSeek was attempting to distill U.S. frontier models. In September it reported disrupting a campaign to extract protected reasoning and attributed a core part of that activity to individuals associated with Moonshot AI, the developer of Kimi. It plans to share more soon about recent campaigns it has disrupted. OpenAI supports legislation giving the U.S. government more tools to identify and sanction foreign actors that illicitly extract capabilities from American models, and says it has told Congress it supports the Chip Security Act, the AI OVERWATCH Act and the MATCH Act. Its argument is that export controls limit adversaries' access to training hardware, distillation defenses make it harder to get the same capabilities from models already trained, and both help maintain a U.S. lead that gives developers more room to slow a release when they need time to test safeguards.
Key facts
- In an internal adaptive red-team evaluation, the average attack success rate for recovering protected reasoning fell from 54.0% on GPT-6 Astra with original safeguards to 0.2% on GPT-6.2 Astra with updated safeguards.
- In a separate adaptive black-box evaluation, the average attack success rate on GPT-6.2 Astra was below 1% for each attack type, aggregated across six attacker models.
- In an internal cyber study, students trained on exact reasoning traces generally beat those trained on summaries after supervised fine-tuning; after reinforcement learning the exact-trace student led only at the largest inference budget tested.
- GPT-6.2 Astra and GPT-6.2 Sol tie protected reasoning to conversation history: rewriting an earlier message makes the API drop the reasoning that followed it, and older model families cannot read it.
- OpenAI says weekly enforcement against distillation campaigns has ranged from thousands to tens of thousands of accounts, and names DeepSeek and individuals associated with Moonshot AI in connection with extraction activity.
Why it matters
Distillation lets one model learn a frontier model's skills without inheriting the safeguards around them. OpenAI's concern is sharpest in dual-use areas such as cybersecurity and biology, where a model with weaker protections could give help the original would have refused. The post argues that reasoning traces are the most valuable thing to protect, because they can hold information the final answer leaves out. Its internal study found that exact traces can transfer more capability than the summaries users normally see, at least at the largest inference budget tested after reinforcement learning. The post also ties the issue to national security and to U.S. legislation and export controls.
Who it affects
Developers of frontier models are the main audience: OpenAI shares indicators of compromise with other providers and says sharing findings across industry and government can help disrupt campaigns. Operators who offer OpenAI's models fall under its consistent-standards policy for shared safeguards and coordinated enforcement. Accounts running unauthorized distillation face enforcement against the whole network. API users should expect that protected reasoning is not reusable across user, workspace, organization and model-family boundaries, and that rewriting an earlier message drops the reasoning that followed it. The named parties are DeepSeek and individuals associated with Moonshot AI, and U.S. lawmakers considering the Chip Security Act, the AI OVERWATCH Act and the MATCH Act.
How to use it
This is a description of OpenAI's own safeguards, not a product to adopt, and no pricing or licence is involved. Builders on OpenAI's API should know that protected reasoning is tied to the conversation that produced it and cannot be carried between users, workspaces, organizations or older model families. Other model providers can take the design ideas: layered request-level and campaign-level defenses, input and output classifiers that also cover tool-call arguments, adversarial training from real attacks, and static plus adaptive, black-box plus white-box testing. OpenAI also says it shares indicators of compromise so other providers can investigate related abuse.
How solid is it
This is a first-party OpenAI post, and the headline figures come from OpenAI's own internal evaluations. The 54.0% to 0.2% drop is for an adaptive evaluation of recovering reasoning from an earlier prompt. The below-1% figure comes from a separate adaptive black-box evaluation, and the post does not list the individual attack types or their values. The post does not name the student or teacher models, the benchmarks or the numerical scores in the cyber study; it only describes the direction of results. It gives no total number of accounts banned and no figure for how prevalent distillation abuse is or how long enforcement takes. Its statements about DeepSeek and Moonshot AI are OpenAI's own, and the post does not say either company confirmed or denied them.
Risks and caveats
The study's result is mixed: after reinforcement learning, the summary-trained student scored higher at lower inference budgets, and OpenAI says it will keep studying how much transfers across other student models and training recipes. Low attack success rates in tests do not guarantee that real attackers fail, and OpenAI itself notes that adaptive red-teaming can uncover weaknesses static tests miss. Enforcement against large account networks depends on detection, and the post reports that the rate of weekly enforcement actions rose significantly this past summer. OpenAI is only now applying selected improvements to GPT-6 Astra. It says it will share more soon about recent campaigns it has disrupted, so the public picture of those campaigns is incomplete.
“Adversarial distillation creates a path for bad actors to reproduce these capabilities without appropriate safeguards.”
— OpenAI, "Combatting adversarial distillation"