OpenAI Ties Reasoning-Extraction Campaign to Moonshot AI, 64 Days After Shutting It Down

OpenAI said on Wednesday that it disrupted a coordinated campaign to extract the hidden reasoning of its models, and that a core cluster of the activity traces to people associated with Moonshot AI, the Chinese developer of the Kimi models. The company shut the campaign down by July 28 and published its account on September 30, a gap of 64 days.

The post, titled "Disrupting a coordinated model-distillation campaign", was published at 10:30 GMT, according to OpenAI's news feed. OpenAI's own page returned a bot-challenge screen when fetched, so the detail below comes from the feed summary and from two outlets that read the full post: The Next Web and FourWeekMBA.

What OpenAI says happened, and when

The activity began on July 1 at low volume. It spiked on July 24 and 25, when OpenAI counted 16,000 requests from more than 4,000 users. The company then found related activity across more than 15,000 users before closing it out on July 28.

Date	What OpenAI reports
July 1	Activity begins at low volume
July 24 and 25	16,000 requests from more than 4,000 users
July 28	Campaign fully shut down
September 30	Public disclosure

OpenAI calls this adversarial distillation, which it defines as "the systematic and unauthorized use of one model's outputs or reasoning to help train, reproduce, or improve another model", as quoted by The Next Web.

How the extraction worked

The operators copied encrypted reasoning from one conversation. They then asked a model in a separate conversation to decrypt and transcribe it. According to both outlets, the operators did not break the encryption, reach a database or read stored user conversations.

FourWeekMBA reports that a footnote describes the figures as attempted extractions, not necessarily successful ones. It also reports that no success rate was disclosed.

Who OpenAI names, and how firmly

OpenAI wrote that it is unclear whether all operators came from a single actor. It attributes "a core cluster of the activity to individuals associated with Moonshot AI", in the words FourWeekMBA quotes from the post.

Moonshot AI has no quoted response in the post or in either outlet's account. The Next Web and FourWeekMBA both note that OpenAI is a party to the dispute and that the account is its own.

Caroline Zier, who leads strategic national security policy at OpenAI, told Bloomberg, in The Next Web's rendering: "Our concern is about violation of our terms of service, not open models or legitimate distillation."

What OpenAI changed

OpenAI said it banned the accounts involved, tightened sign-up checks and added protections for hidden reasoning. It also said it shared findings with the Frontier Model Forum and with government channels, according to The Next Web.

OpenAI did not say how many of the 16,000 requests returned usable reasoning text. It did not say which models were targeted or whether any Kimi release used the output. It has not said why disclosure took 64 days.

Earlier coverage of OpenAI on this site includes its AI math release norms advisory group and its $30 billion round at a $1.4 trillion pre-money valuation. A written response from Moonshot AI is the next fact that would change this story.