All news

1 October 2026

OpenAI says people tied to China's Moonshot AI tried to copy its models' thinking

OpenAI says it shut down a campaign that tried to pull hidden reasoning out of its AI models, and linked the core of it to the maker of the Kimi chatbot.

Building a top AI model costs a fortune. Copying one, it turns out, might just take a lot of very well-chosen questions. OpenAI says it caught a coordinated group trying to do exactly that, and it has pointed the finger at people linked to Moonshot AI, the Chinese startup behind the Kimi chatbot.

What happened

According to OpenAI, the activity started in early July. At its peak it reached 16,000 requests from more than 4,000 users in just two days. Digging further, the company found related activity across a wider group of more than 15,000 users. It says the campaign was fully shut down by July 28.

The goal was to get OpenAI's models to reveal their reasoning, the step-by-step working a model does before it gives an answer, which OpenAI normally keeps hidden. OpenAI calls this "adversarial distillation." Distillation means using one model's answers to train another; the adversarial part is doing it without permission.

OpenAI stressed what did not happen: nobody broke its encryption or got into its databases or stored user chats. The operators simply tried to steer conversations so the hidden reasoning came out in plain view.

The company said it could not be sure everyone involved was working for one group, but tied the core cluster to individuals associated with Moonshot. It has shared its findings with other AI labs through the Frontier Model Forum, an industry safety group, and with government channels. Moonshot did not immediately respond to CNBC.

Why it matters

This is the second such accusation in a few weeks. Rival Anthropic recently accused several Chinese developers, including Moonshot and Alibaba, of secretly using its Claude model to help train their own systems.

For investors, the issue is the moat. US AI companies are spending enormous sums on chips, data centers and research on the bet that their models stay ahead. If a competitor can shortcut that work by learning from the finished product, it could build similar tools faster and more cheaply, and that lead shrinks. OpenAI also frames it as a safety and national security risk, since copied capabilities would arrive without the safeguards built around the originals.

What to watch

  • Whether Moonshot responds to the claims.
  • Whether other AI labs report similar campaigns after OpenAI's findings were shared.
  • Any response from US officials, given OpenAI routed its findings through government channels.

Based on reporting by CNBC: OpenAI links China’s Moonshot AI to extraction attempt. Facts, not advice.

More news