OpenAI says there was an attempt to obtain protected AI model reasoning — CNBC

OpenAI said it had identified and stopped a coordinated campaign involving attempts to obtain the protected reasoning of its artificial intelligence models. The company linked a key cluster of this activity to individuals associated with China’s Moonshot AI startup, the developer of Kimi, CNBC Top News reports.

Thousands of requests in two days

According to OpenAI, the activity began in early July and later intensified sharply: 16,000 requests were submitted by more than 4,000 users over two days. The company eventually identified related activity in a cluster of more than 15,000 users and said it had fully stopped the campaign by July 28.

OpenAI stressed that the operators did not breach the company’s encryption, databases, or stored user conversations. Instead, they manipulated interactions with the models in an attempt to reproduce hidden reasoning in a form accessible to the person submitting the request.

More current news is available on the UA.News Telegram channel Telegram.

What adversarial distillation is

The company called this approach “adversarial distillation.” It involves using the responses or reasoning of one AI model to train or improve another. OpenAI believes that obtaining such reasoning could allow others to replicate advanced model capabilities without comparable investments in their development and protection, creating potential security and national security risks.

OpenAI said it was unable to determine whether all operators were linked to a single entity. The company said it shared its findings with other AI developers through the Frontier Model Forum and government information-sharing channels. Moonshot AI did not provide CNBC with a comment at the time of publication.

Several weeks earlier, Anthropic accused Chinese AI developers, including Moonshot AI and Alibaba, of covertly using the Claude model to train their own systems.