OpenAI published a detailed account on November 20, 2026 of how it detects and disrupts adversarial distillation, the unauthorized use of a frontier model's outputs to train another model that then runs without comparable safeguards. The company frames the problem as a safety issue as well as an intellectual-property one. Its example is a model trained on a stronger system's cyber capabilities and then used to help carry out attacks that the original system would have refused.
The post puts particular weight on protecting a model's chain-of-thought, which OpenAI calls protected reasoning, because those intermediate steps can contain information the final answer leaves out. In an internal study, OpenAI fine-tuned the same pretrained student model on Astra-level teacher trajectories for coding and defensive cybersecurity tasks, comparing students that saw only user-visible content with students that also saw exact reasoning traces. After supervised fine-tuning the exact-trace student generally did better. After a light round of reinforcement learning it still scored higher at the largest inference budget tested, while the summary-trained student led at lower budgets. OpenAI reads this as evidence that guarding traces can limit cyber capability transfer.
On defenses, the company describes two layers. At the request level it uses adversarial training, input classifiers that spot extraction attempts, output classifiers that look for leaks in both messages and tool-call arguments, and API limits on reusing protected reasoning. With GPT-6.2 Astra and GPT-6.2 Sol, a model's protected reasoning is tied to the conversation history that produced it, so rewriting an earlier message makes the API drop the reasoning that followed. Older model families with weaker safeguards are also blocked from reading it. At the campaign level, OpenAI looks for coordinated patterns across accounts and enforces against whole networks.
The main number comes from an adaptive black-box evaluation using automated red-teaming, aggregated across six attacker models. The average success rate for recovering reasoning from an earlier request fell from 54.0% against GPT-6 Astra with its original safeguards to 0.2% against GPT-6.2 Astra with updated ones. For extraction from the current request, it fell from 22.9% to 0.9%. OpenAI says it is now applying selected improvements to GPT-6 Astra too. It reports weekly enforcement actions against distillation campaigns ranging from thousands to tens of thousands of accounts, with the rate rising sharply over the summer.
The post also takes a policy position. OpenAI recalls briefing Congress in February about activity indicating that DeepSeek was trying to distill U.S. models, and says its September report attributed a core part of one reasoning-extraction campaign to individuals associated with Moonshot AI, developer of Kimi. It says it supports legislation to sanction foreign actors that illicitly extract capabilities, and has told Congress it backs the Chip Security Act, the AI OVERWATCH Act and the MATCH Act. It promises more detail on recent campaigns soon, and shares indicators of compromise with other providers.
All figures come from OpenAI's own internal tests and have not been independently reproduced, and the attack set depends on the red-teamers used. The company lists steps others could copy, including protecting reasoning across messages and tool calls, keeping successful attack logs as regression tests and pairing request-level controls with campaign-level enforcement.