"DeepSeek and Moonshot's Main Target for 'Distillation' Is Astra"

▲ Astra

OpenAI's state-of-the-art artificial intelligence (AI) model, "GPT-6 Astra," released in September, has become a primary target for so-called "distillation" attacks by leading Chinese AI companies such as DeepSeek and Moonshot AI, U.S. tech media outlet The Information reported on the 28th (local time), citing sources.

A "distillation" attack refers to an act where competitors systematically and repeatedly query a specific AI using proxy services and tens of thousands of fake accounts to save costs and time in developing their own AI models.

Through this method, they reverse-engineer and steal the reasoning patterns and capabilities of the target model.

Sources stated that Moonshot AI distilled Astra during the development process of "Kimi K3.1 Pro," which was released last week, and DeepSeek benefited from Astra distillation while creating its "V4.1-Flash 1017" version released on the 17th.

Sources added that while Chinese AI companies had until recently mainly copied Anthropic's Claude models, their target shifted to OpenAI, which is relatively easier, as Anthropic took strong countermeasures against such attacks.

Anthropic and OpenAI have repeatedly accused Chinese developers, including Moonshot and DeepSeek, of distilling their models.

Distillation is a commonly used technique in AI development, and Anthropic and OpenAI have themselves used distillation techniques when creating smaller, faster, and cheaper versions of their models.

However, external companies targeting these models is controversial.

Both OpenAI and Anthropic prohibit other companies from distilling their models, and they do not make their own models available within China.

Anthropic announced earlier this month that new security measures restricting access to Claude's reasoning process have led to a sharp decrease in information extraction attempts, stating that "the amount of data (reasoning tokens) extracted by distillation actors has also dropped dramatically to one-tenth to one-hundredth of previous levels."

Anthropic stated that it has made it difficult for attackers to repeatedly probe Claude's reasoning process, and has identified and permanently blocked tens of thousands of malicious accounts linked to distillation attacks.

An OpenAI spokesperson said, "The company takes this issue very seriously and continues to work on resolving it."

※