Anthropic claimed that China's Alibaba tried to illegally access its artificial intelligence (AI) model Claude by mobilizing tens of thousands of fake accounts.
Bloomberg said on the 24th (local time) that Anthropic said this in a letter sent to U.S. senators and White House officials.
According to the report, Anthropic claimed that operators linked to the lab that developed Alibaba's AI model Qwen used 25,000 fake accounts to exchange 28.8 million conversations with Claude starting in April.
Anthropic said they targeted Claude's core strengths, including software engineering and agent reasoning.
It added that Chinese corporations such as Alibaba are imitating major U.S. AI models at lower expense through so-called "adversarial distillation" attacks. Distillation is a technique in which developers train their own systems using the outputs of another AI model to achieve similar performance at much lower expense. If used to imitate cutting-edge AI models without permission, it may violate terms of service.
Anthropic added that it is being carried out "illegally, systematically, and on an industrial scale."
On Apr. in a memo, Michael Kratsios, head of the White House Office of Science and Technology Policy (OSTP), suggested that the government would cooperate in cracking down on attempts by Chinese corporations to misuse the outputs of U.S. AI models.
The memo defined these actions as different from legitimate research practices in that they are conducted on an "industrial scale" and rely on thousands of fake accounts.
In the U.S. Senate, Sen. Bill Hagerty (R-Tenn.) and Sen. Andy Kim (D-N.J.) plan to introduce an amendment to the National Defense Authorization Act (NDAA). The amendment is expected to include blacklisting or sanctioning Chinese corporations that inappropriately accessed outputs of U.S. AI models to train competing AI models.