On November 20, 2026, OpenAI described how it fights adversarial distillation, the unauthorized copying of a frontier model’s capabilities into another model. The focus is the protected reasoning, or chain of thought, that models such as GPT-6.2 Astra and GPT-6.2 Sol generate internally. In an internal test, the success rate of automated attackers trying to recover reasoning from an earlier request fell from 54.0 percent against GPT-6 Astra with its original safeguards to 0.2 percent against GPT-6.2 Astra. The figures come from OpenAI itself and have not been independently verified. The report is mainly interesting because it explains why distillation is more than an intellectual property issue for providers.
What adversarial distillation is
Distillation itself is a common machine learning technique. A smaller or newer model, the “student”, learns from the answers, reasoning or judgments of another model, the “teacher”. Many providers use it legitimately to turn large models into cheaper variants.
It becomes adversarial when someone harvests another provider’s model outputs at scale without permission to train their own model. According to the OpenAI report, the safety problem is that the student inherits the capabilities but not the safeguards. A model that helps defenders find software vulnerabilities could, once copied, help attackers exploit them. The same logic applies to biology and other dual-use domains.
Why the chain of thought is so valuable
Reasoning models generate intermediate steps before answering. Users usually see only a summary. The exact reasoning traces often contain information the final answer leaves out, which makes them especially valuable training data.
OpenAI examined this in an internal study. Each training run started from the same pretrained student model. It was fine-tuned on trajectories from Astra-level teacher models working on coding and defensive cybersecurity tasks. One variant only received what users normally see: final responses, reasoning summaries and tool interactions. The other also received the exact reasoning traces.
After supervised fine-tuning, the student trained on exact traces generally performed better. After an additional round of reinforcement learning, it still scored higher at the largest inference budget tested, while the summary-trained student scored higher at lower budgets. OpenAI’s conclusion: protecting reasoning traces limits the transfer of cyber capabilities, even when the original tasks were benign and allowed. Across many interactions, exact traces can teach a student skills that generalize beyond the teacher’s original tasks. According to OpenAI, Anthropic reported a similar effect on the SWE-bench Verified coding benchmark.
Two layers of defense
OpenAI describes defenses on two levels that are meant to reinforce each other.
At the request level, three mechanisms apply:
- Adversarial training: models learn to recognize extraction attempts and refuse to disclose protected reasoning while still answering legitimate requests for explanations. Training examples come from successful attacks found through internal red-teaming, investigations or reports from outside researchers.
- Classifiers: input classifiers detect extraction attempts, output classifiers check responses and tool-call arguments for leaks. The latter matters because a model can expose reasoning in a tool argument without showing it in the answer.
- API protections: with GPT-6.2 Astra and Sol, protected reasoning is tied to the conversation history that produced it. If an earlier message is rewritten, the API drops the reasoning that followed. Older model families with weaker safeguards can no longer read the new models’ protected reasoning.
The last point responds directly to a campaign OpenAI disclosed on September 30. Attackers had copied encrypted reasoning from one conversation into another and asked the model to reproduce it in readable form. OpenAI now says it strengthened protections across user, workspace, organization and model-family boundaries, including closing a pathway that could have exposed another user’s encrypted reasoning.
At the campaign level, OpenAI looks for patterns in request and account signals that point to coordinated extraction. Weekly enforcement actions against such campaigns this year ranged from thousands to tens of thousands of accounts, with the rate rising significantly over the summer. Indicators are shared with other model providers.
The numbers behind the adversarial distillation defenses
OpenAI measured the effect with automated AI red-teamers that adapt their tactics to the system’s responses, in a black-box setting that mirrors an external user’s view. Results are averaged across six attacker models:
- Extracting reasoning from an earlier request: 54.0 percent success against GPT-6 Astra (original safeguards), 0.2 percent against GPT-6.2 Astra
- Extracting reasoning from the current request: 22.9 percent against GPT-6 Astra, 0.9 percent against GPT-6.2 Astra
OpenAI says it is now applying selected improvements to GPT-6 Astra as well. We covered the models themselves at launch in our article on GPT-6.2 Astra and Sol.
The political dimension
OpenAI explicitly frames the issue as a national security matter. The company notes that it briefed the US Congress in February on activity indicating DeepSeek was attempting to distill US frontier models. It attributed a core part of the September campaign to individuals associated with Moonshot AI, the developer of Kimi, and says it will share more about recent campaigns soon.
OpenAI also backs legislation giving the US government more tools against foreign actors and names the Chip Security Act, the AI OVERWATCH Act and the MATCH Act. The timing matches a report from two House committees released the same day, which cites distillation as one reason for the rise of Chinese open-weight models. It is worth remembering that OpenAI is also arguing its own case here: making distillation harder protects its business model as well as safety.
What other providers can take from it
OpenAI closes with recommendations for other developers and platforms:
- Protect reasoning across the full interaction, including tool arguments, and restrict reuse of encrypted reasoning across conversations and models.
- Keep successful extraction attacks as regression tests and test new safeguards against adaptive attackers.
- Combine request-level mitigations with campaign-level investigation and enforcement.
- Coordinate safeguard requirements with partners that serve your models.
For developers using the OpenAI API, binding reasoning to the conversation history has a practical consequence: if you rewrite earlier messages in a conversation, you lose the reasoning that followed. That is intentional, but it may affect agent frameworks that rewrite or compact histories. Teams running such workflows should test the behavior after moving to GPT-6.2.
Overall, the report is one of the most detailed descriptions so far of how a major provider secures its models’ reasoning. The impressive percentages come from internal tests under conditions OpenAI set itself. How robust the defenses are against determined, well-resourced attackers will only become clear once independent researchers examine them.




