OpenAI says attackers’ average success at pulling protected reasoning out of its models fell from 54.0% to 0.2% after it upgraded safeguards in GPT-6.2 Astra, according to a post published on 20 November. That is a 270-fold drop, by our arithmetic.
The same post says OpenAI’s weekly enforcement against distillation campaigns has ranged from thousands to tens of thousands of accounts, “with the rate increasing significantly this past summer.” It repeats that a core part of a September campaign was attributed to people associated with Moonshot AI, the developer of Kimi, which we reported at the time.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer. It teaches the general idea only and does not discuss this report.
The key insight: OpenAI says its updated GPT-6.2 Astra safeguards cut reasoning-extraction success from 54.0% to 0.2% in its own red-team tests, and that its weekly enforcement actions against distillation campaigns have ranged from thousands to tens of thousands of accounts.
What OpenAI Measured
The test was an internal adaptive black-box evaluation: automated red-teamers adjusted their attacks based on how the system responded, with the same view an external user would have, OpenAI says. Results were aggregated across six attacker models.
Extracting protected reasoning from an earlier request succeeded 54.0% of the time against GPT-6 Astra with its original safeguards and 0.2% against GPT-6.2 Astra with updated safeguards. Extracting it from the current request fell from 22.9% to 0.9%, about 25 times lower by our arithmetic.
OpenAI says it is now applying selected safeguard improvements to GPT-6 Astra as well.


Why OpenAI Guards the Reasoning
Protected reasoning is “the intermediate steps a model generates internally while working toward an answer, known as its chain-of-thought,” the post says, and these traces “can contain information the final answer leaves out, making them especially valuable training data for another model.”
In an internal study, a student model trained on exact reasoning traces from Astra-level teachers generally outperformed one trained on reasoning summaries after supervised fine-tuning. After reinforcement learning, the exact-trace student still scored higher at the largest inference budget tested, while the summary-trained student scored higher at lower budgets.
OpenAI’s conclusion: protecting reasoning traces “can limit cyber capability transfer, even when the underlying task is one allowed under our usage policies.” Anthropic reported a similar advantage for full transcripts, the post notes; we covered Anthropic’s own anti-distillation measures in October.
How the Defenses Work
At the request level, OpenAI uses adversarial training so models refuse to disclose protected reasoning, input classifiers to spot extraction attempts, and output classifiers that check both answers and tool-call arguments for leaks.
With GPT-6.2 Astra and GPT-6.2 Sol, a model’s protected reasoning is tied to the conversation history that produced it; if an earlier message is rewritten, the API drops the reasoning that followed. Older model families with weaker safeguards are blocked from reading the new models’ protected reasoning.
At the campaign level, OpenAI says it enforces against networks of accounts and shares indicators of compromise with other model providers. The September disclosure described attackers moving encrypted reasoning from one conversation into another and asking the model to reveal it.
The National Security Pitch
OpenAI says it briefed Congress in February on activity indicating DeepSeek was attempting to distill US frontier models, and that it plans “to share more soon about recent campaigns we’ve disrupted.”
It says it supports legislation that would give the US government more tools to identify and sanction foreign actors that illicitly extract capabilities, and has told Congress it supports the Chip Security Act, the AI OVERWATCH Act and the MATCH Act.
The post lands in a week of US pressure on Chinese models: SpaceX-owned Cursor told lawmakers it is weighing phasing out Chinese open-weight models, and Treasury Secretary Scott Bessent linked AI attacks to Chinese models (our piece).
The Structural Read
As we read it, OpenAI is treating a model’s hidden reasoning the way it treats its weights: as the asset to protect, because that is what a rival would most want to copy.
As we read it, pairing the security numbers with three named bills turns a safety post into a lobbying document; the figures are still OpenAI’s own.
OpenAI, ‘Combatting adversarial distillation’
“Our weekly enforcement actions against these campaigns have ranged in scale from thousands to tens of thousands of accounts, with the rate increasing significantly this past summer.”
Three Implications
FOR RIVAL LABS OpenAI says it shares indicators of compromise with other model providers and recommends checking tool-call arguments for leaks.
FOR CONGRESS OpenAI says it supports the Chip Security Act, the AI OVERWATCH Act and the MATCH Act.
FOR THE RECORD the success rates are OpenAI’s internal evaluations, not independent tests.
What Is Not Established
The figures are OpenAI’s own internal evaluations; no outside party has published a replication. The post names no new labs beyond the earlier DeepSeek and Moonshot AI attributions and gives no account totals for specific campaigns. It does not include responses from DeepSeek or Moonshot AI.
The Bottom Line
OpenAI says its updated safeguards cut extraction of earlier-request reasoning from 54.0% to 0.2% and that its weekly enforcement actions have ranged from thousands to tens of thousands of accounts, while asking Congress for new powers against foreign distillers.
97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. All figures and quotes are from OpenAI’s 20 November post and are the company’s own internal evaluations. Multiples (270-fold, about 25 times) are our arithmetic. Earlier attributions are as OpenAI stated them.
Sources: OpenAI, ‘Combatting adversarial distillation’, 20 Nov 2026: 54.0%->0.2% and 22.9%->0.9% ASR tables; weekly enforcement thousands to tens of thousands of accounts; Moonshot/DeepSeek attributions; legislation · FourWeekMBA 235111 (30 Sep 2026): OpenAI names Moonshot-linked reasoning extraction · FourWeekMBA 237083 (5 Oct 2026): OpenAI’s distillation campaign disclosure · FourWeekMBA 257334 (13 Oct 2026): Anthropic’s anti-distillation measures









