OpenAI published a technical update on Friday titled "Combatting adversarial distillation," laying out how it protects the hidden reasoning inside its models from being harvested to train other models. Adversarial distillation is the practice of using a frontier model's outputs, especially its chain-of-thought, to build training data for another model that inherits the capabilities but not the safety work.
The headline numbers come from internal adaptive black-box red-team evaluations that pooled six automated attacker models. For attempts to extract protected reasoning from an earlier request, the average attack success rate fell from 54.0 percent against GPT-6 Astra with its original safeguards to 0.2 percent against GPT-6.2 Astra with the updated stack. For reasoning generated in the current request, success fell from 22.9 percent to 0.9 percent. OpenAI said it is also applying selected safeguard improvements back to GPT-6 Astra. All of these figures are company-reported and have not been independently verified.
The defenses are layered. Model training now teaches the model to recognize extraction attempts and refuse them while still answering legitimate requests. Input classifiers flag suspicious requests, and output classifiers scan both assistant messages and tool-call arguments, including streaming responses, before anything reaches the caller — a gap that matters because models can leak reasoning through tool arguments rather than final answers. At the API layer, OpenAI restricts reuse of protected reasoning across users, workspaces, organizations and model families, and blocks older, weaker model families from reading the protected reasoning of GPT-6.2 Astra and GPT-6.2 Sol.
The most interesting new mechanism is conversation binding. In the GPT-6.2 family, protected reasoning is cryptographically tied to the conversation history that produced it. If an earlier message in the thread is rewritten, the API drops all reasoning that was generated after it, a design the company documents as "keep-prefix-unchanged." That closes a classic attack: ask a benign question, then quietly edit the history to retroactively turn it into an extraction attempt.
Enforcement is the other half of the story. OpenAI says it detected multiple coordinated extraction campaigns this year and takes action against entire account networks when it identifies them, with weekly enforcement ranging from thousands to tens of thousands of accounts and the pace increasing significantly over the summer. It shares indicators of compromise with other model providers. A new usage policy that took effect on November 20 makes the rules explicit: users must not use OpenAI models or their outputs to generate, prepare, select or supply training material for models intended as alternatives to OpenAI's, and platforms that resell direct access to the models — model routers and gateways — now need OpenAI's prior review and approval.
The update also cites internal research on why the reasoning trace matters so much. Training a student model on exact reasoning traces produced better cybersecurity benchmark results on ExploitBench and SEC-bench Pro than training on reasoning summaries, with the advantage persisting at higher inference budgets; OpenAI's footnote notes Anthropic has reported a similar effect on SWE-bench Verified. In other words, the hidden chain-of-thought is not just an interpretable intermediate product — it is a transferable skill set.
Why it matters: chain-of-thought has become the most valuable and most contested artifact in the API economy. As model routers, gateways and resellers multiply between developers and labs, every hop handles plaintext prompts and every lab is now writing policy, classifiers and cryptographic binding to keep its reasoning from walking out the door. The 54-to-0.2 drop is OpenAI's own scorecard, but the structural point is hard to argue with: protecting reasoning is now a full-time engineering discipline, not a terms-of-service clause.
Comments (0)
Log in to join the discussion
Log InNo comments yet