Safety and security
OpenAI says it disrupted a campaign to extract its models' reasoning
OpenAI said on September 30, 2026 that it has disrupted a coordinated campaign to copy the protected reasoning of its models. This is the company's account, and for now it comes without details.
The essentials
3 according to the source · 3 open questions
- According to the sourceOpenAI says it has disrupted a coordinated campaign to distill its models, according to the text it published on September 30, 2026.1
- According to the sourceAccording to OpenAI, the campaign aimed to extract the protected reasoning of its models.1
- According to the sourceOpenAI says it is strengthening its defenses against adversarial distillation, though the available summary does not explain how.1
Why it matters
If it happened the way OpenAI tells it, this suggests two things. The first is that the company treats its models' reasoning as something that needs protecting. The second is that someone tried to get hold of it in an organized way, not as a one-off case.
With only the company's account, the scale cannot be measured. We do not know whether the campaign was large or small, or whether it was stopped before or after it got what it was after.
The details
On September 30, 2026, OpenAI published a text in which it says it has disrupted a coordinated campaign to distill its models1. Distilling, in this context, means using an AI model's answers to train another model that imitates it. According to the company, the campaign sought to extract the protected reasoning of its models1.
OpenAI adds that it is strengthening its defenses against what it calls adversarial distillation, meaning distillation done against the wishes of whoever created the model1. All of this is the company's account. Maunga has only been able to see the summary of the announcement, which does not say who was behind the campaign, how long it lasted or what it managed to obtain1.
What we don't know
- Who organized the campaign and from where: the summary of the announcement does not say.
- How much reasoning was extracted, over how long, and how OpenAI detected it.
- What the new defenses consist of, and whether anyone outside the company has verified what happened.
Background
Maunga's archive records three OpenAI security incidents in 2026, but they were of a different kind: they involved the company's own internal agents. What OpenAI describes now is an attempt to extract information from its models1.
Sources
Was this story useful? Yes Not really
This story was written in Spanish by an artificial intelligence system from the sources listed above, and translated automatically. Before publication, a program checks that every figure and every name appears in the sources, and that the translation keeps the same facts, figures and sources. It has not been reviewed by a person. Spanish original · How we work (in Spanish)