OpenAI said it detected and halted a coordinated campaign to extract protected step-by-step reasoning from its models, with a central cluster traced to people connected to Moonshot AI, the Chinese company behind Kimi. Activity began at low volume on July 1 before more than 4,000 users submitted 16,000 requests on July 24 and 25; OpenAI identified similar behavior across more than 15,000 users and said the campaign was fully stopped by July 28. OpenAI called the method adversarial distillation, in which outputs or hidden reasoning from one model are harvested to develop or improve another, but said there was no breach of encryption, databases or stored conversations. The disclosure adds to Anthropic’s allegation that roughly 300,000 requests were routed through about 5,380 fraudulent proxy accounts in Singapore and Japan, primarily targeting Claude Opus. U.S. officials, including White House Office of Science and Technology Policy Director Michael Kratsios, have accused Moonshot of distilling data from Anthropic’s Fable model to build Kimi K3 and of obtaining banned Nvidia chips. Moonshot has denied the allegations in earlier statements, although it did not respond to CNBC’s request for comment on the latest disclosure. Kimi K3, an open-weight model with 2.8 trillion parameters, was released in July 2026 and has drawn attention for coding and reasoning; the accounts give conflicting release dates of July 16 and July 27. No specific sanctions or other government measures have been confirmed.