OpenAI Says It Cut Attempts to Steal GPT-6.2’s Hidden Reasoning From 54% to 0.2%, as House Investigators Accuse Six Chinese AI Labs of Industrial-Scale Distillation and Urge a Ban on Chinese Models in Federal Systems

The fight over who gets to copy whom in artificial intelligence moved from courtrooms and press briefings into the plumbing of machine learning this week. On November 20, OpenAI published a detailed account of how it is trying to stop rivals from training their own models on the hidden reasoning of its newest systems, and said its updated safeguards cut the success rate of one class of extraction attack from 54% to 0.2%. The same day, two House committees released a joint investigative report alleging that six leading Chinese AI developers ran “industrial-scale” distillation campaigns against American frontier models, and urging Washington to keep Chinese-built models out of federal systems and high-risk critical infrastructure.

Taken together, the two documents turn a once-obscure training technique into a security, compliance and procurement question that reaches well beyond AI labs, into any company that buys, fine-tunes or deploys a model.

What distillation is, and why the reasoning trace is the prize

Distillation is a standard machine-learning method: a smaller or newer “student” model learns from outputs generated by a stronger “teacher.” Done with permission, it is how many efficient models get built. Done without permission, through fake accounts and proxy traffic, it lets a competitor shortcut years of compute spending. OpenAI calls the unauthorized version adversarial distillation, and argues the risk is not only commercial: a copied model can inherit a frontier system’s capabilities without inheriting the safety controls that limit how those capabilities are used.

The most valuable target, according to OpenAI, is a model’s protected chain-of-thought, the intermediate reasoning it produces before giving an answer. Those traces often contain information the final reply leaves out. In an internal experiment, OpenAI trained identical student models on outputs from its top-tier teacher models doing coding and defensive cybersecurity work. After supervised fine-tuning, the student that saw exact reasoning traces generally beat the one that saw only reasoning summaries. After a round of reinforcement learning, the exact-trace student still led at the largest compute budget tested, although the summary-trained student did better at lower budgets. OpenAI’s conclusion is that guarding reasoning traces can limit how much cyber capability leaks into a copy, even when each individual request was perfectly legitimate.

The numbers OpenAI is reporting

OpenAI says it hardened GPT-6.2 Astra and GPT-6.2 Sol with a layered stack:

  • Adversarial training that teaches the model to recognize extraction attempts and decline to reveal protected reasoning, while still explaining its answers to ordinary users.
  • Input and output classifiers that scan prompts, replies and even tool-call arguments, since a model can leak its reasoning into a tool call without ever showing it in the final answer.
  • API rules that bind reasoning to the conversation that produced it. If an earlier message is rewritten, the API discards the reasoning that followed, and older model families with weaker safeguards are blocked from reading the new models’ reasoning. OpenAI also says it closed a pathway that could have exposed another user’s encrypted reasoning.

In an adaptive black-box red-team test aggregated across six automated attacker models, the average success rate for pulling reasoning from an earlier request fell from 54.0% on GPT-6 Astra with its original safeguards to 0.2% on GPT-6.2 Astra. For reasoning tied to the current request, it dropped from 22.9% to 0.9%. OpenAI adds that its weekly enforcement actions against distillation campaigns have ranged from thousands to tens of thousands of accounts, with the pace rising sharply over the summer, and that it is back-porting some of the fixes to GPT-6 Astra.

These are the company’s own figures from its own test harness, and outside observers such as DeAI have noted that the methodology is not published in a form others can reproduce. The direction of travel is still telling: a leading lab is now treating its model’s internal reasoning the way a bank treats customer data.

Congress puts names and dollar figures on the problem

The joint report from House Homeland Security Chairman Andrew Garbarino and Select Committee on China Chairman John Moolenaar, both Republicans, followed a seven-month investigation launched in April. Among its findings: U.S. agencies assessed that six leading Chinese developers conducted industrial-scale distillation against American frontier models, likely with Beijing’s awareness. The report cites earlier disclosures by American developers that DeepSeek, Moonshot AI and MiniMax collectively generated more than 16 million exchanges with U.S. AI systems through roughly 24,000 fraudulent accounts. It also says American data-labeling firms reportedly earn about $500 million a year selling training data to Chinese labs.

The committees corresponded with three U.S. companies to understand adoption: Airbnb, which uses Alibaba models in customer support; Anysphere, which built the flagship model in its Cursor coding tool on a Moonshot AI base model; and DoorDash, which said several Chinese model families had been authorized and deployed. As Anadolu Agency noted, the report did not establish that any of those deployments led to a cyberattack or data leak.

The report lists 12 recommendations. The headline ones would extend the existing DeepSeek exclusion to all Chinese-developed models in federal systems and federal contract work, remove them from high-risk critical-infrastructure functions, and bar federal cybersecurity grants from buying them. Others call for a national American open-weight AI initiative, a model provenance standard paired with an “AI bill of materials,” sanctions on illicit distillers, and a look at export controls on U.S. training-data sales. None of it is law yet.

Why it matters for business

Model provenance is becoming a procurement question. Cheap, capable open-weight models from Chinese labs have been an easy win for startups and enterprises trying to cut inference costs. If an AI bill of materials becomes a requirement for federal contractors, any company selling into government, and eventually its suppliers, will need to document which base models sit inside its products. Teams that fine-tuned on a Chinese base model may face a costly swap later, the “rip and replace” scenario the committee chairmen explicitly warn about.

API behavior is changing under developers’ feet. OpenAI’s decision to drop reasoning when an earlier message is edited, and to block cross-model access to reasoning, will matter to anyone building agents that rewrite conversation history or chain models together. Builders should test their pipelines rather than assume past behavior holds.

The price gap is the real policy problem. The House report concedes that the U.S. lacks an open-weight alternative at the price and performance the market wants. Until that changes, restrictions will push costs up for buyers in regulated sectors, which may widen the opening for American open-weight releases and for vendors that can certify where their models came from.

Your own AI outputs are an asset worth guarding. The same logic that drives OpenAI’s defenses applies to any business that exposes a fine-tuned model to the public. Rate limits, anomaly detection on bulk accounts and careful decisions about what reasoning to show users are now part of protecting intellectual property.

The bottom line

Distillation has gone from a research shortcut to a front line in the U.S.-China AI contest. OpenAI is betting that technical locks on reasoning traces can raise the cost of copying, while Congress is betting on bans, provenance rules and sanctions. For companies in the middle, the practical move is the same either way: know exactly which models run in your stack, where they came from, and what your vendors’ APIs will and will not let you do.

Sources: OpenAI, “Combatting adversarial distillation” (Nov. 20, 2026); House Committee on Homeland Security release; Joint investigative report (PDF); Anadolu Agency; Bloomberg Law; DeAI.

Put AI to work on the front line of your business. While the big labs fight over who owns the smartest models, MAJ.COM Voice AI gives you 24/7 AI voice agents that answer every inbound call and handle outbound sales and customer support, so you never miss a lead. See how MAJ.COM Voice AI works →

Edited by Coron @QUE.COM
Website: https://QUE.COM Intelligence
Sponsored by: https://MAJ.COM AI Autonomous


Discover more from QUE.com

Subscribe to get the latest posts sent to your email.

Coron @QUE.COM

Editor at QUE.com Intelligence.

coronquecom has 226 posts and counting.See all posts by coronquecom

Leave a Reply

Connect with

Discover more from QUE.com

Subscribe now to keep reading and get access to the full archive.

Continue reading