OpenAI Accuses Moonshot AI of Model-Distillation Campaign

OpenAI has accused individuals associated with Moonshot AI, the Chinese company behind the Kimi family of AI models, of carrying out a coordinated model-distillation campaign. OpenAI said the activity was designed to extract protected reasoning from its models and potentially use the information to improve another AI system. The company said it identified and disrupted the activity after detecting unusual interactions with its models.

According to OpenAI, the activity initially appeared at relatively low levels in early July 2026 before increasing sharply later that month. The company said it recorded about 16,000 requests from more than 4,000 users on July 24 and 25, eventually identifying similar activity involving more than 15,000 users. OpenAI said it had fully disrupted the activity by July 28, although it was not clear whether all of the activity was connected to a single organisation.

The allegations centre on a technique known as model distillation, in which outputs from a more capable AI model can be used as training material for another model. OpenAI said some of the activity attempted to obtain protected reasoning by manipulating conversations and transferring encrypted reasoning information between sessions. The company described this as a security concern because extracted reasoning could potentially be used to train another model while avoiding some of the safeguards applied to the original system.

The accusations follow similar claims from Anthropic involving Moonshot AI and other Chinese AI developers. Anthropic said in September that it had identified activity in which Moonshot allegedly routed customer requests to Claude and subsequently extracted reasoning information from the resulting exchanges. Anthropic reported more than 23 million exchanges associated with Moonshot between May and July, while also identifying separate campaigns involving Alibaba and DeepSeek. These are allegations from the companies involved and have not been independently established as facts by the reports cited here.

The dispute highlights growing tensions over how AI companies protect proprietary model capabilities as competition between frontier AI developers intensifies. Distillation itself is a legitimate machine-learning technique, but companies increasingly distinguish between authorised distillation and attempts to obtain model outputs through large-scale or deceptive methods. OpenAI said adversarial distillation could create security and national-security risks, while the broader industry continues to develop technical measures to prevent unauthorised extraction of model capabilities.