OpenAI Says Moonshot AI Tried to Steal Its Hidden AI Secrets

TLDR

• OpenAI detected a surge of requests trying to extract hidden reasoning from its AI models starting in early July.
• Activity spiked to 16,000 requests from over 4,000 users in two days, with related activity found across more than 15,000 users.
• OpenAI says no encryption, databases, or stored conversations were breached.
• A core cluster of the activity was linked to people connected to Moonshot AI, maker of the Kimi AI system.
• The disclosure follows similar accusations Anthropic made against Moonshot AI and Alibaba weeks earlier.

---

OpenAI says it found and shut down a coordinated effort to pull hidden reasoning out of its AI models. The company says a central part of the activity was linked to people connected to Moonshot AI, a Chinese company that builds the Kimi AI system.

OpenAI accused Chinese rival Moonshot AI of being responsible for a wide-scale effort to extract data from its GPT artificial intelligence systems that could be used to reproduce the reasoning and capabilities of its most advanced models https://t.co/Vsejx99s0N

— Bloomberg (@business) September 30, 2026

According to OpenAI, the activity started in early July at a low level. It then jumped on July 24 and 25, when more than 4,000 users sent 16,000 requests using a similar pattern.

OpenAI later found related activity tied to a wider group of more than 15,000 users. The company says it fully stopped the campaign by July 28.

What OpenAI Found

OpenAI calls the method “adversarial distillation.” This means taking one model’s outputs or reasoning and using them to train or improve a different model without permission.

The company says the operators did not break into its encryption, databases, or stored user conversations. Instead, they used a technique to make hidden reasoning from one conversation show up in a different conversation.

OpenAI says this let them read reasoning that is normally kept hidden from users. The company says this kind of extraction could let others copy advanced AI abilities without spending the same time or money on safety work.

OpenAI shared its findings with other AI companies through the Frontier Model Forum. It also shared the information through government channels.

Steps OpenAI Took

OpenAI says it took several steps after finding the activity. These include limiting or removing accounts involved in the requests.

The company also says it added new barriers to stop people from creating new accounts for this purpose. It closed the specific weakness that let hidden reasoning be extracted this way.

OpenAI added new systems to detect this kind of activity in real time. When the activity involved outside services, OpenAI says it worked with those companies to find and stop the accounts.

Moonshot AI has not responded to requests for comment, according to CNBC.

Broader Scrutiny of Moonshot

This is not the first time Moonshot AI has faced accusations like this. Michael Kratsios, who leads the White House Office of Science and Technology Policy, has said Moonshot carried out large scale distillation against U.S. models.

Kratsios also said Moonshot obtained banned Nvidia chips to build its Kimi K3 model. Separately, a research firm called Frontier Security says Kimi K3 escaped a cybersecurity test built by the U.K. government’s AI Safety Institute.

Moonshot has not responded to these claims either.

This news follows a similar case from weeks earlier, when Anthropic accused Moonshot AI and Alibaba of using its Claude model to help train their own systems without permission.

OpenAI says it expects more attempts like this in the future. The company says these efforts are likely to become harder to detect as AI models keep improving.