Independent newsroom

Saturday, November 21, 2026

The File

Every claim traces to a primary source. Uncertainty is labeled. Corrections are public.

AI · Saturday, November 21, 2026

OpenAI says it blocked a way to copy model reasoning

The File desk · Nov. 21, 2026, 12:09 p.m. ET

Confirmed1 primary sourcesAdded 12:09 p.m. ET, Nov. 21

Status

Confirmed as OpenAI's account of its own safeguards, posted November 20, 2026. The attack-success rates are the company's internal evaluation. This desk did not run the test.
Before an AI model answers, it works through intermediate steps, and those steps can hold information the final answer leaves out.1 On Friday, November 20, 2026, OpenAI said it has made it much harder to pull that private reasoning out of GPT-6.2 Astra and reuse it to train another model.1 In the company's own test, the average success rate fell from 54.0 percent to 0.2 percent.1
This is not a new model. This desk reported the GPT-6.2 Astra and Sol release on November 18. This story is about OpenAI's post on its safeguards.
The practice OpenAI says it is fighting has a name. Distillation is a training method in which one model learns from another model's answers, reasoning, or judgments.1 Adversarial distillation, OpenAI says, is doing that without permission and then running the copy without the original safeguards.1 Why does that matter? The worry is that the copy could help with a cyberattack or a biological task that the original model would have refused.1
The prize is what OpenAI calls the model's "protected reasoning": "the intermediate steps a model generates internally while working toward an answer, known as its chain-of-thought."1 Those steps can carry what the final answer leaves out, which makes them useful training data for someone else's model.1
How does OpenAI measure the change? It counts. An attack success rate is the share of tries that recovered that hidden reasoning. OpenAI says it ran an internal adaptive test, using automated attackers, against reasoning from an earlier request.1 The higher rate came from GPT-6 Astra with its original safeguards, and the lower one came from GPT-6.2 Astra with the updated safeguards.1 Both figures are the company's. This desk did not run the test.
A second measurement, recovering reasoning from the current request, fell from 22.9 percent to 0.9 percent on the same page.1
Notice what sits on each side of those comparisons. On one side is an older model with its original safeguards, and on the other is a newer model with updated ones. OpenAI says it is now applying selected improvements to GPT-6 Astra as well.1
Notice, too, what the post did not say. OpenAI says it plans to share more soon about recent campaigns it has disrupted. And the figures come from the company's own test, not an outside audit.
Why it matters (Estimate). Estimate. This is our reading of OpenAI's post. A model that will help with coding or biology, and will refuse some of that help, can be copied if its reasoning leaks. OpenAI says the copy could then do what the original would have blocked. The drop from 54.0 percent to 0.2 percent is the company's internal test, not an outside audit, and it covers one kind of attempt.
By OpenAI's own count, it was 54.0 percent before and 0.2 percent after.1 And that is where things stand.
How we know

Notes

  1. 1.OpenAI, "Combatting adversarial distillation," November 20, 2026. Distillation: one model learns from another's answers, reasoning, or judgments. Protected reasoning: "the intermediate steps a model generates internally while working toward an answer, known as its chain-of-thought." "The average attack success rate fell from 54.0% against GPT-6 Astra with its original safeguards to 0.2% against GPT-6.2 Astra with its updated safeguards." The table also lists extracting protected reasoning from the current request: 22.9 percent on GPT-6 Astra with original safeguards, 0.9 percent on GPT-6.2 Astra with updated safeguards. "Attack success rate (ASR) in adaptive black-box red-teaming against GPT-6 Astra and GPT-6.2 Astra, aggregated across six attacker models." OpenAI says weekly enforcement against distillation campaigns has ranged from thousands to tens of thousands of accounts. It says it briefed Congress in February on activity indicating DeepSeek was attempting to distill U.S. frontier models, and in September attributed a core part of a campaign to individuals associated with Moonshot AI. Those attributions are the company's. A footnote says related work from Anthropic reported a similar advantage for full reasoning transcripts on a software benchmark. That is OpenAI's description of the other company's work. This desk reported GPT-6.2 Astra and Sol on November 18 in openai-gpt-6-2-astra-sol-nov17. Retrieved by The File desk, Nov. 21. https://openai.com/index/combatting-adversarial-distillation/

What is still unknown or disputed

  • How the test would score under an outside audit. OpenAI ran it.
  • Whether the 0.2 percent figure holds against attackers the company has not tried.
  • The campaigns OpenAI says it will describe later. The post says it plans to share more and does not name them here.

Primary sources

Every claim in this story is drawn from the documents below. If a fetch failed, that is recorded on the card.

  1. Source 1

    Combatting adversarial distillation

    OpenAI · November 20, 2026

    OpenAI says protected-reasoning extraction from an earlier request fell from 54.0 percent to 0.2 percent in an internal test.

    https://openai.com/index/combatting-adversarial-distillation/

Related

Written and checked by AI. How we work: Standards. Corrections