OpenAI accuses Moonshot AI of mass GPT reasoning extraction attempts

By Splitfeed EditorialLast updated Oct 1, 2026, 00:20 UTC

On September 30, 2026, OpenAI publicly accused Moonshot AI, the Chinese company behind the Kimi conversational AI product, of running a coordinated campaign to extract hidden reasoning data from GPT models. OpenAI said the effort involved more than 10,000 distinct attacks and that it successfully disrupted the operation. The accusation was confirmed across multiple outlets including Bloomberg and CNBC, and Futurism reported on the notification emails OpenAI sends to organizations whose infrastructure was implicated in or affected by such campaigns. Moonshot AI had not issued a public response as of the available reporting.

The stakes here extend well beyond a single security incident. Hidden reasoning data, sometimes called chain-of-thought traces, represents some of the most proprietary material a frontier model developer holds. If a competitor could extract it at scale, they would gain detailed insight into how a model structures its thinking, potentially shortcutting years of expensive training and research. The accusation also lands in a charged geopolitical context, adding a concrete technical flashpoint to the broader US-China competition in AI development. What remains genuinely in dispute is the attribution itself: OpenAI has made the link to Kimi, but without a response from Moonshot AI, the question of intent, authorization, and the precise mechanism of the attacks is unresolved in public.

None of the three camps had published reactions as of the available sources. Pro-AI voices would typically treat this as evidence that frontier model security infrastructure is working as intended, pointing to OpenAI's disruption of the campaign as a sign that the industry can defend itself against extraction attempts. Anti-AI voices would likely use the story to argue that the concentration of powerful proprietary systems in a small number of companies creates exactly the kind of high-value target that invites adversarial behavior, raising questions about systemic risk. Middle-ground observers would probably focus on the difficulty of attribution in these cases and call for clearer international norms around model security and intellectual property.

The next meaningful development to watch is whether Moonshot AI issues a formal response to the accusations, and whether OpenAI or any regulatory body moves toward legal or diplomatic action. Any technical disclosure about how the extraction attempts were structured and detected would also significantly sharpen the public understanding of what actually occurred.