Key Highlights
- OpenAI said it identified and disrupted a campaign attempting to extract protected reasoning from its AI models, with a core cluster linked to Chinese startup Moonshot AI.
- Activity surged to 16,000 requests from more than 4,000 users over two days, with related activity ultimately traced across more than 15,000 users.
- OpenAI said operators did not breach its encryption, databases, or stored user conversations.
- The findings follow Anthropic's recent accusation that Chinese AI developers, including Moonshot AI and Alibaba, secretly used its Claude model to train their own systems.
A Campaign to Extract Hidden Reasoning
OpenAI said it identified and disrupted a coordinated campaign in which operators attempted to extract protected reasoning from its artificial intelligence models, linking a core cluster of the activity to Chinese startup Moonshot AI, the developer behind the Kimi model. The company described the activity as beginning in early July before surging sharply.
At its peak, the campaign generated 16,000 requests from more than 4,000 users over a two-day span, according to OpenAI. The company ultimately traced related activity across a broader cluster of more than 15,000 users and said it had fully disrupted the campaign by July 28.
OpenAI described the behavior as "adversarial distillation," a technique in which one AI model's outputs or underlying reasoning are used to help train or improve a separate model. The company said extracting such reasoning could allow other developers to reproduce advanced capabilities without making the same investment in developing and safeguarding frontier models, raising both safety and national security concerns.
No Breach of Core Systems, OpenAI Says
OpenAI said the operators involved did not breach its encryption, databases, or stored user conversations. Instead, the company said the activity involved manipulating interactions with its models in an effort to surface hidden reasoning in a form visible to the requester, rather than exploiting a technical vulnerability in its infrastructure.
OpenAI said it remains unclear whether every operator involved in the campaign is connected to a single actor, though the company attributed a core cluster of the activity to individuals associated with Moonshot AI. OpenAI said it has shared its findings with other AI developers through the Frontier Model Forum, as well as through government information-sharing channels.
Part of a Broader Pattern of Cross-Border AI Tension
The disclosure comes just weeks after OpenAI rival Anthropic accused several Chinese AI developers, including Moonshot AI and Alibaba, of secretly using its Claude model to help train their own AI systems. That earlier accusation underscored growing concern among United States AI companies that competitors could use their models to accelerate the development of rival technology more quickly and at lower cost.
Together, the two disclosures reflect a broader anxiety within the American AI industry: that the enormous capital and engineering investment required to build frontier models could be partially circumvented by extracting reasoning or outputs from already-trained systems, effectively allowing rivals to shortcut years of research and development spending.
A Test for Industry-Wide Information Sharing
OpenAI's decision to share its findings through the Frontier Model Forum, an industry group focused on frontier AI safety, points to a growing reliance on collective information-sharing mechanisms as individual companies grapple with these threats. Whether such mechanisms can meaningfully slow coordinated extraction campaigns, particularly those that do not rely on traditional cybersecurity breaches, remains an open question facing the industry.
The overlapping accusations from OpenAI and Anthropic also suggest that adversarial distillation techniques may be increasingly common across multiple frontier AI labs, rather than isolated incidents, intensifying scrutiny on how AI companies monitor and respond to unusual usage patterns on their platforms.






Please wait processing your request...