Campaign Identified
OpenAI revealed it has dismantled a coordinated operation seeking to extract the internal reasoning of its artificial intelligence models. The central part of the activity was attributed to the Chinese startup Moonshot AI, responsible for the Kimi chatbot.
The action began in July and peaked at 16,000 requests from more than 4,000 accounts in just two days. In total, the company tracked more than 15,000 involved users before completely ending the offensive on July 28.
Method Used
Instead of invading servers or encryption, the operators manipulated conversations with the models to force the exposure of reasoning steps that normally remain hidden. This process, called “adversarial distillation”, allows training rival systems with less cost and time.
Context of Tensions
The episode occurs weeks after Anthropic accused Chinese developers, including Moonshot, of using its Claude model to accelerate its own AI training. The case reinforces American concerns about unauthorized transfer of advanced technology.
Risks Pointed Out
OpenAI warned that extracting reasoning can compromise the security of frontier models and have implications for national security. The company shared the findings with the Frontier Model Forum and governmental intelligence channels.
Moonshot has not yet officially commented on the allegations.