OpenAI Says New Defenses Cut Reasoning Theft From 54% to 0.2%
OpenAI says GPT-6.2 defenses cut reasoning-extraction success from 54.0% to 0.2%, and attributes campaigns to DeepSeek and Moonshot AI — but every figure is self-reported.
By Daniel Okafor
4 min read

What's News
- OpenAI's November 20 post claims GPT-6.2 Astra defenses cut reasoning-extraction success from 54.0% to 0.2%.
- All figures come from OpenAI's internal red-teaming; no independent reproduction exists.
- OpenAI attributes a September extraction campaign to individuals linked to Moonshot AI and briefed Congress on DeepSeek in February.
- Anthropic reported in October 2026 that its own defenses cut distillation extraction by 10 to 100 times.
- A House Select Committee report in November 2026 recommended sanctions over distillation by PRC labs.
By Daniel Okafor
OpenAI claims its new defense stack cuts the success rate of attacks that extract protected model reasoning from 54.0% to 0.2%. The figure covers GPT-6.2 Astra, comes from the company's own red-teaming, and has not been reproduced by any independent evaluator. OpenAI published the technical details in a November 20 blog post on adversarial distillation — the practice of using a frontier model's outputs to train a smaller copy that replicates its capabilities without its safeguards.
The headline numbers are specific. Adaptive attacks that extract reasoning from an earlier request now succeed 0.2% of the time, down from 54.0%. Attacks on the current request drop from 22.9% to 0.9%. GPT-6.2 Sol receives the same layered defenses, but OpenAI does not publish separate extraction rates for it.
How the defense stack works
OpenAI describes a two-layer architecture. At the request level:
- Adversarial training teaches models to refuse extraction attempts while preserving legitimate explanations.
- Input classifiers flag extraction prompts before they reach the model.
- 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.
What the internal distillation study found
OpenAI also released an internal cyber distillation study. It started from the same pretrained student model and fine-tuned two versions: one on exact reasoning traces from Astra-level teachers working on coding and defensive cybersecurity tasks, and one 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, while the summary student scored higher at lower budgets. OpenAI argues this proves protecting traces limits cyber capability transfer even from policy-compliant tasks.
The blog post does not name the benchmarks used. ExploitBench and SEC-Bench Pro appear in the separate GPT-6.2 System Card as the main cybersecurity capability evaluations.
Who does OpenAI blame?
OpenAI 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 attackers moved encrypted reasoning between conversations and asked the model to decrypt it — a path OpenAI says it has now closed.
The post frames 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 backs 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.
What the numbers leave out
Every metric comes from OpenAI's adaptive black-box red-teaming against six attacker models. No third party has verified the 0.2% and 0.9% figures. The distillation study reports directional results only, without underlying scores or named benchmarks. Selected improvements are being backported to GPT-6 Astra, but its exposure is not covered by the 0.2% figure. The DeepSeek and Moonshot attributions are OpenAI statements. The NSA/CISA/FBI joint advisory of September 2026, cited by the House Select Committee with actor-level detail — naming DeepSeek, Moonshot AI, Alibaba, Zhipu, MiniMax, StepFun, and Xiaomi, with dates, targets, and exchange volumes — is publicly available on CISA's website.
Industry context and what comes next
Anthropic 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. Anthropic cited an internal study where full reasoning transcripts outperformed summaries on SWE-bench Verified. The House Select Committee on the CCP, in a November 2026 report, cited both companies' disclosures and alleged industrial-scale distillation by PRC labs including DeepSeek, Moonshot, Alibaba, Zhipu, MiniMax, and StepFun. It recommended sanctions and supply-chain authorities against firms engaged in illicit distillation.
OpenAI advises other labs to 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.
The core claim is that layered defenses can push adaptive extraction success below 1% in OpenAI's own test environment. Whether those rates hold against novel attacks, across model families, or under sustained campaign pressure is a question only independent evaluators can answer.
Source: epyhia.com
More from Daniel Okafor
Show full bio
Correspondent covering money, funding and market shifts around entrepreneurship, startups and venture funding at Business Bearings.
2,538 articles