.jpg?w=1024)
The AI race is no longer only about who can build the strongest model. It is also about who can stop rivals from copying one.
Anthropic accused Alibaba-linked operators of using thousands of fake accounts to extract millions of responses from Claude, according to a letter submitted to US senators and seen by Reuters. The company said the activity was part of an attempted AI distillation campaign, a method used to train one model on another model’s outputs.
The allegation adds a sharper edge to the US-China AI fight, where frontier model access, safety controls, and national security concerns are increasingly tangled together.
An AI distillation attack
On June 10, two days before Fable 5 was restricted, Anthropic presented a letter to US Senators Tim Scott and Elizabeth Warren ahead of a hearing on artificial intelligence. In the letter, the Claude maker claimed that actors linked to Alibaba’s AI lab, Qwen, engaged in illicit interactions with Claude from April 22 to June 5, 2026.
According to Anthropic, almost 25,000 accounts engaged in the act, generating more than 28.8 million conversations with Claude during that period.
The technical term for that is AI distillation: training another AI model using data from a superior model. Put simply, it works by presenting a model with complex queries, collecting the responses, and using that to train another model. The company noted that this allows China to develop models with capabilities similar to those of Mythos Preview.
Alibaba, caught in US tensions
Alibaba is a Chinese e-commerce platform with an international footprint and a vast portfolio of subsidiaries. However, the company is now grappling with a US blacklist that could undermine its popularity and revenue in the region.
After the US designated Alibaba as a “Chinese military company,” the company sought legal redress, noting that the claims had "no basis in fact or law.”
On top of that, the company now has to deal with Anthropic’s case, which could worsen its already strained relationship with US officials.
More AI theft claims
Alibaba is Anthropic’s latest case on AI distillation from China.
In February, the company announced that Chinese AI startups DeepSeek, MoonShot, and MiniMax used Claude’s data to improve their models. Together, the trio allegedly created 16 million interactions using 24,000 fraudulent accounts.
DeepSeek has also reportedly targeted OpenAI for the same purposes.
The dangers behind distilling AI models
Aside from revenue loss, these allegations point to a growing challenge for both AI companies and their customers.
For AI developers, successful distillation could allow competitors to close the technology gap without matching the years of research, computing resources, and engineering required to build frontier models from scratch.
For users, the risk is less visible. While a distilled model may deliver similar answers, Anthropic argues it may not inherit the same safety alignment and safeguards built into the original system, raising new questions about reliability and risk in enterprise deployments.
Anthropic’s ask is simple: that the US government continue to help in combating these kinds of attacks, which can also have national security implications.
Related reading: Anthropic's latest allegations come as the company is already navigating a tense relationship with Washington. Read our full timeline of that conflict.