{"slug": "openai-says-new-defenses-cut-reasoning-extraction-attacks-from-54-to-0-2-but-are", "title": "OpenAI Says New Defenses Cut Reasoning-Extraction Attacks From 54% to 0.2%, But Figures Are Self-Reported", "summary": "OpenAI published a November 20 technical blog post claiming layered defenses in GPT-6.2 Astra cut the success rate of adaptive reasoning-extraction attacks from 54.0% to 0.2% on an earlier request and from 22.9% to 0.9% on the current request, with GPT-6.2 Sol receiving the same defenses but no separate published rates. All figures come from OpenAI's internal red-teaming against six attacker models, with no independent reproduction, and the accompanying cyber distillation study reports only directional results without naming benchmarks or disclosing scores. OpenAI also said it briefed Congress in February on activity indicating DeepSeek was attempting to distill U.S. frontier models and in September attributed a core extraction cluster to individuals associated with Moonshot AI, framing adversarial distillation as a national security risk.", "body_md": "# OpenAI Says New Defenses Cut Reasoning-Extraction Attacks From 54% to 0.2%, But Figures Are Self-Reported\n\nOpenAI published a technical blog post on November 20 detailing new defenses against adversarial distillation — the practice of using a frontier model's outputs to train a smaller copy that replicates its capabilities without its safeguards. The company says the updated protections in GPT‑6.2 Astra reduce the success rate of adaptive attacks that extract protected reasoning from an earlier request from 54.0% to 0.2%, and from 22.9% to 0.9% for the current request. GPT‑6.2 Sol receives the same layered defenses, but OpenAI does not publish separate extraction success rates for it. All figures come from OpenAI's own internal red-teaming; no independent reproduction has been published.\n\n## What OpenAI claims changed\n\nThe post describes a layered defense stack. At the request level: adversarial training teaches models to refuse extraction attempts while preserving legitimate explanations; input classifiers flag extraction prompts; output classifiers scan assistant messages and tool-call arguments for leaked reasoning; API controls bind encrypted reasoning to the conversation that produced it, invalidate it when earlier messages are edited, and block older model families from reading GPT‑6.2 traces. At the campaign level, OpenAI says it investigates coordinated account networks, enforces bans at scale (thousands to tens of thousands of accounts per week, rising sharply over summer 2026), and shares indicators of compromise with other providers and through the Frontier Model Forum.\n\nThe company also released an internal cyber distillation study. Starting from the same pretrained student model, it fine-tuned one version on exact reasoning traces from Astra-level teachers working on coding and defensive cybersecurity tasks, and another on user-visible content (final answers, reasoning summaries, tool interactions). After supervised fine-tuning, the exact-trace student generally outperformed the summary student. After a lightweight reinforcement-learning pass, the exact-trace student still led at the largest inference budget on the benchmarks shown, while the summary student scored higher at lower budgets. OpenAI argues this shows protecting traces limits cyber capability transfer even from policy-compliant tasks. The blog post does not name the benchmarks used in the distillation study; ExploitBench and SEC‑Bench Pro appear in the separate GPT‑6.2 System Card as the main cybersecurity capability evaluations.\n\n## Attribution claims and national security framing\n\nOpenAI says it briefed Congress in February on activity indicating DeepSeek was attempting to distill U.S. frontier models. In September it reported disrupting a protected-reasoning extraction campaign and attributed a core cluster to individuals associated with Moonshot AI, developer of Kimi. The September disclosure described attackers moving encrypted reasoning between conversations and asking the model to decrypt it — a path OpenAI says it has now closed.\n\nThe post frames adversarial distillation as a national security risk: a foreign actor that reproduces advanced capabilities without safeguards can deploy them for offensive cyber or biological use. OpenAI says it supports the Chip Security Act, the AI OVERWATCH Act, and the MATCH Act, arguing export controls limit hardware access while distillation defenses make it harder to acquire capabilities from already-trained models.\n\n## What the numbers leave out\n\n- **Self-reported.** Every metric comes from OpenAI's adaptive black-box red-teaming against six attacker models. No third party has reproduced the 0.2% / 0.9% figures.\n- **Benchmark scores omitted.** The cyber distillation study reports directional results (\"generally outperformed,\" \"scored higher at the largest inference budget\") but does not disclose the underlying scores or identify the benchmarks shown in its charts.\n- **Older models partially covered.** Selected improvements are being backported to GPT‑6 Astra, but its exposure is not described by the 0.2% figure.\n- **Attribution unconfirmed.** The DeepSeek briefing and Moonshot attribution are OpenAI statements. The NSA/CISA/FBI joint advisory (September 2026) was cited by the House Select Committee with specific actor-level attributions — naming DeepSeek, Moonshot AI, Alibaba, Zhipu, MiniMax, StepFun, and Xiaomi with dates, targets, and exchange volumes — and the advisory itself is publicly available on CISA's website.\n\n## Industry context\n\nAnthropic published its own distillation defenses in February 2026 and reported in October 2026 that its changes cut what reasoning distillers extract by 10 to 100 times, also citing an internal study where full reasoning transcripts outperformed summaries on SWE‑bench Verified. The House Select Committee on the CCP (November 2026) cited OpenAI and Anthropic disclosures in a report alleging industrial-scale distillation by PRC labs including DeepSeek, Moonshot, Alibaba, Zhipu, MiniMax, and StepFun, and recommended sanctions and supply-chain authorities against firms engaged in illicit distillation.\n\n## What developers should watch\n\nOpenAI's recommendations for other labs: scan tool-call arguments for leaks, keep successful extraction prompts as regression tests, combine request-level mitigations with campaign-level investigation, and coordinate safeguard floors with deployment partners. For API users, the practical changes are invisible — reasoning summaries remain visible, raw traces stay hidden — but editing an earlier message now drops subsequent reasoning, and bulk scraping faces network-level enforcement.\n\nThe core advance OpenAI demonstrates is that layered defenses can push adaptive extraction success rates below 1% in its own test environment. Whether those rates hold against novel attack methods, across different model families, or under sustained campaign pressure remains to be seen by independent evaluators.\n\n## Sources\n\n1. [Combatting adversarial distillation](https://openai.com/index/combatting-adversarial-distillation) (openai.com)\n2. [OpenAI Adversarial Distillation: GPT-6.2 Safeguards (Nov 20)](https://explainx.ai/blog/openai-adversarial-distillation-protected-reasoning-campaigns-november-2026) (explainx.ai)\n3. [OpenAI's Distillation Defenses: 54% to 0.2%](https://cellcog.ai/blog/openai-distillation-defenses) (cellcog.ai)\n4. [GPT-6.2 System Card](https://deploymentsafety.openai.com/gpt-6-2-astra-sol) (deploymentsafety.openai.com)\n5. [OpenAI Says New Defenses Cut Reasoning Theft From GPT-6.2 ...](https://aiparabellum.com/openai-adversarial-distillation-defenses-gpt-6-2-astra) (aiparabellum.com)\n6. [MITRE ATLAS: The Complete Guide for Security Teams](https://www.picussecurity.com/resource/blog/mitre-atlas-the-complete-guide-for-security-teams) (picussecurity.com)\n7. [Untitled - Select Committee on the CCP |](https://chinaselectcommittee.house.gov/sites/evo-subsites/selectcommitteeontheccp.house.gov/files/evo-media-document/2026.11.16-joint-report-prc-ai-models-signed-final-1.pdf) (chinaselectcommittee.house.gov)\n8. [Combatting adversarial distillation | OpenAI](https://openai.com/index/combatting-adversarial-distillation?ref=explainx) (openai.com)\n9. [Disrupting a coordinated model-distillation campaign | OpenAI](https://openai.com/index/disrupting-a-coordinated-model-distillation-campaign/?ref=explainx) (openai.com)\n10. [OpenAI US House Select Cmte Update \\[021226\\]](https://assets.bwbx.io/documents/users/iqjWHBFdfxIU/rRmql_jJcxb4/v0) (assets.bwbx.io)", "url": "https://wpnews.pro/news/openai-says-new-defenses-cut-reasoning-extraction-attacks-from-54-to-0-2-but-are", "canonical_source": "https://epyhia.com/openai-says-new-defenses-cut-reasoning-extraction-attacks-from-54-to-0", "published_at": "2026-11-21 05:09:47+00:00", "updated_at": "2026-11-21 05:11:33.728547+00:00", "lang": "en", "topics": ["ai-safety", "ai-policy", "large-language-models", "artificial-intelligence"], "entities": ["OpenAI", "GPT-6.2 Astra", "GPT-6.2 Sol", "DeepSeek", "Moonshot AI", "Kimi", "Frontier Model Forum", "Chip Security Act"], "also_reported_by": [], "takeaway": null, "recommendation": null, "alternates": {"html": "https://wpnews.pro/news/openai-says-new-defenses-cut-reasoning-extraction-attacks-from-54-to-0-2-but-are", "markdown": "https://wpnews.pro/news/openai-says-new-defenses-cut-reasoning-extraction-attacks-from-54-to-0-2-but-are.md", "text": "https://wpnews.pro/news/openai-says-new-defenses-cut-reasoning-extraction-attacks-from-54-to-0-2-but-are.txt", "jsonld": "https://wpnews.pro/news/openai-says-new-defenses-cut-reasoning-extraction-attacks-from-54-to-0-2-but-are.jsonld"}}