OpenAI Accuses Moonshot AI of Coordinated Data Extraction Campaign
OpenAI disrupted a campaign by Chinese developer Moonshot AI to extract protected reasoning from GPT models to train competing AI systems.
OpenAI disrupted a coordinated campaign between July 1 and July 28, 2026, aimed at extracting protected reasoning from its artificial intelligence models. The company identified a process called adversarial distillation, where operators manipulated model interactions to reveal internal records of how tasks are processed—information typically withheld from final answers. This activity peaked on July 24 and 25 with 16,000 requests from over 4,000 users, and OpenAI eventually identified a cluster of more than 15,000 suspect accounts.
OpenAI attributed a core cluster of this activity to Moonshot AI, the Chinese developer of the Kimi AI. The attackers allegedly used novel methods, such as copying encrypted reasoning from one conversation and asking a model in another to decrypt and transcribe it. OpenAI stated that while no databases or stored conversations were breached, the extracted data could be used to reproduce the capabilities of its most advanced models to train the Kimi K3 model without the necessary investment in frontier development.
In response, OpenAI banned fraudulent accounts, strengthened sign-up controls, and patched bugs that exposed hidden reasoning. The company shared its findings with government channels and other developers. This follows similar accusations by Anthropic, which claimed Moonshot routed hundreds of thousands of requests to its Claude models for training. Separately, Moonshot AI launched an internal review after the UK-based firm Mindgard demonstrated that Kimi models could be jailbroken to generate information on biological weapons and assassination planning.