Distillation: OpenAI says it cut reasoning extraction to 0.2%, the day a House report targeted Chinese open-weight models
On November 20, OpenAI detailed its defenses against “adversarial distillation”, which means training a rival model on the outputs of its own: the success rate of one tested attack falls from 54% to 0.2% on GPT-6.2 Astra. The same day, two House committee chairmen released a 50-page report recommending that Chinese models be kept out of federal systems, citing the cases of Airbnb, Cursor and DoorDash.

Two publications came out on the same day, Friday, November 20, in Washington and San Francisco. They tell the same story. OpenAI detailed how it defends itself against “adversarial distillation”: a competitor trains its own model on GPT’s outputs and picks up its capabilities without the safeguards. The same day, two House committee chairmen released a 50-page investigative report. It calls for keeping Chinese open-weight AI models out of federal systems, and argues that much of their progress comes precisely from this kind of distillation.
What OpenAI protects: hidden reasoning
The main target is protected reasoning, the intermediate steps a model such as GPT-6.2 Astra produces before answering, without the user seeing them. According to OpenAI, these traces contain information the final answer leaves out, which makes them especially valuable training data for a “student” model.
OpenAI publishes an internal study to show this. A student model was trained on coding and defensive cybersecurity tasks solved by an Astra-level model. After supervised training, the student fed with exact traces generally did better than the one fed with the reasoning summaries visible to users. After a round of reinforcement learning, the advantage only held at the largest compute budget tested.
The attack test results, measured by OpenAI with six automated attacker models:
| Attack success rate (internal evaluation) | GPT-6 Astra, original safeguards | GPT-6.2 Astra, updated safeguards |
|---|---|---|
| Extracting reasoning from an earlier request | 54.0% | 0.2% |
| Extracting reasoning from the current request | 22.9% | 0.9% |
On the API side, reasoning is now tied to the history that produced it: if a developer rewrites an earlier message, the API drops the reasoning that followed. In September, OpenAI had described attacks that moved encrypted reasoning into another conversation to get it revealed; it says it has since strengthened the walls between users, workspaces and organizations. Older, less protected model families can no longer read GPT-6.2’s reasoning. OpenAI also says it acts every week against networks ranging from thousands to tens of thousands of accounts. It recalls briefing Congress in February on distillation attempts attributed to DeepSeek, then attributing part of a campaign in September to individuals associated with Moonshot AI, the developer of Kimi.
The House report: Airbnb, Cursor and DoorDash as examples
The report is signed by Andrew Garbarino, chairman of the Homeland Security Committee, and John Moolenaar, chairman of the Select Committee on China, both Republicans. The investigation opened in April and draws on three U.S. companies:
| Company | Chinese models cited | Use described in the report |
|---|---|---|
| Airbnb | Qwen 2.5, 3 and 3.5 (Alibaba) | Customer assistant (about 80,000 interactions a day) and summaries of about 250,000 support cases and calls a day |
| Anysphere (Cursor) | Kimi K2.5 (Moonshot AI) | Base of Composer 2, Cursor’s in-house model, a fact that became public after an independent developer identified it |
| DoorDash | Qwen, DeepSeek, Kimi, GLM, MiniMax, ByteDance | 25 model families authorized and deployed in a non-trivial manner (at least $5,000 of compute); replacing one of them could take up to six months |
The report measures the scale of the trend using public sources. On OpenRouter, the share of open-weight models in tokens from U.S.-originating requests reportedly rose from 26% in September 2025 to 60% in August 2026, mostly to the benefit of Chinese models. Chinese models reportedly accounted for 41% of downloads on Hugging Face over a year, and the Qwen family has more than 151,000 derivative models. According to the report, the best American open-weight model is only 10th in the latest Artificial Analysis index, behind six Chinese models. In the September ranking, the best American open-weight model was Thinking Machines Lab’s Inkling, whose initial training used synthetic data generated notably by Kimi K2.5, according to the lab as cited by the report. The same lab has just released Inkling 2 Mini.
The 12 recommendations go beyond a ban. The report calls for banning Chinese models from federal systems and contracts, removing them from high-risk functions of critical infrastructure and blocking federal cybersecurity grants from being used to buy them. It also proposes a national initiative for American open-weight models, a model provenance standard with an AI “bill of materials”, sanctions against distillers, and a review of sales of American training data to Chinese labs.
Our take
The two texts answer each other: OpenAI supplies the technical argument, Congress draws political conclusions. They still need to be read for what they are. OpenAI’s attack rates are measured by OpenAI, and distillation attributions rest on statements by companies and agencies that cannot be publicly verified. The report, for its part, does not have the force of law: these are recommendations from two majority committee chairmen, and they only concern the United States. The report also acknowledges the real driver of adoption: companies pick these models because they are cheap, customizable, and there is no American equivalent at the same price. For a European company using Qwen or Kimi, nothing changes in the short term. The demand for traceability, knowing which model runs where and where it comes from, could however become a requirement in contracts with American customers.
Sources
- OpenAI — Combatting adversarial distillation (November 20, 2026) Primary source
- House Committee on Homeland Security — joint report press release (November 20, 2026) Primary source
- Joint report of the House committees on PRC AI models (PDF, November 20, 2026) Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.


