The anti-distillation defenses are getting stronger: on November 20, OpenAI published details of the adversarial-distillation defenses in GPT-6.2. Adversarial distillation means using a frontier model's outputs to train another model, moving the capability over while leaving the safeguards behind. In OpenAI's internal red-team tests, the success rate for extracting protected reasoning from an earlier request fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra.
The defenses work on two layers:
- Request layer: adversarial training teaches the model to refuse to disclose protected reasoning; input and output classifiers check both answers and tool-call arguments, blocking extraction attempts; API controls stop older models from reading protected reasoning; and reasoning is bound to the conversation history that produced it, so if an earlier message is rewritten, the API drops the reasoning that followed.
- Campaign layer: OpenAI bans thousands to tens of thousands of accounts each week suspected of taking part in extraction campaigns, with enforcement rising sharply since this summer, and shares indicators of compromise with other model providers.
How to Read the 0.2% Figure
The number needs two caveats. First, this is OpenAI's internal adaptive red-team evaluation, in which attackers only see what an ordinary user would see, and no independent third party has replicated it yet. Second, 0.2% measures whether an extraction attempt succeeds, not how much stronger a student model ultimately becomes from stolen reasoning. OpenAI's own conclusion leaves room: protecting reasoning traces can limit the transfer of cyber capability, even when the underlying task is one allowed under its usage policies — the defense targets capability outflow, not the task itself.
The context is familiar. Earlier, DeepSeek and Moonshot AI were accused of training new models by distilling GPT-6 Astra outputs, and Anthropic published its own distillation defenses around the same time — see Anthropic's distillation defenses take effect: extracted reasoning from Claude falls to a fraction of before. OpenAI says selected improvements are also being applied to GPT-6 Astra. What matters next is how long the defenses hold against real attacks, and whether other providers follow with the same kind of disclosure.