OpenAI announced on Wednesday that it identified and neutralized a coordinated campaign aimed at extracting proprietary reasoning from its artificial intelligence models. The company linked the core cluster of this activity to Moonshot AI, the developer behind the Kimi chatbot, intensifying friction in the global AI race.
According to OpenAI, the adversarial activity began in early July. It escalated significantly over a two-day period, generating approximately 16,000 requests from more than 4,000 users. Ultimately, the company traced related behavior to a network exceeding 15,000 users and stated it had fully disrupted the campaign by July 28.
OpenAI characterized the maneuver as “adversarial distillation,” a technique where one AI model’s outputs or underlying reasoning are utilized to train or enhance another system. The company warned that successfully extracting such reasoning could allow competitors to replicate advanced capabilities without bearing the substantial costs associated with developing and safeguarding frontier models, thereby posing risks to both safety and national security.
A spokesperson for OpenAI clarified that the operators did not breach the company’s encryption, databases, or stored user conversations. Instead, they manipulated interactions with the models to reproduce hidden reasoning in a format accessible to the requestor. While OpenAI noted it was unclear if all operators were linked to a single entity, it attributed a central cluster of the activity to individuals associated with Moonshot AI.
The discovery follows a recent accusation by Anthropic, OpenAI’s rival, which alleged that several Chinese AI developers, including Moonshot AI and Alibaba, secretly used its Claude model to assist in training their own systems. These events highlight growing concerns among U.S. AI firms that rivals are leveraging their technology to develop competing products more rapidly and cheaply.
OpenAI has shared its findings with other AI developers through the Frontier Model Forum and via government information-sharing channels. Moonshot AI did not immediately return requests for comment.
OpenAI’s numbers feel exaggerated for PR impact. 16,000 requests? Big tech always inflates threat levels when competitors get too close to parity.
Wait, did they actually hack anything? The article says no encryption or databases were breached. Is this really a cyberattack or just aggressive prompt engineering?
This is alarming. Using another company’s R&D to train your own models without permission feels like intellectual theft disguised as competitive strategy.