A single US artificial intelligence company says it has identified nearly 29 million attempts to steal its core technology. The accused: Alibaba, the Chinese e-commerce giant. The method: a technique called “distillation,” where one AI model’s responses are harvested to train another.
Anthropic, the San Francisco firm behind the Claude AI model, laid out the accusation in a June 10 letter to US Senators Tim Scott and Elizabeth Warren. The company told lawmakers the operation involved thousands of fraudulent accounts.
It called it the largest extraction campaign of its kind. The stakes are concrete. Distillation attacks let a less capable model absorb the reasoning and decision-making methods of a more advanced one.
In this case, Anthropic says the operators targeted Claude’s most valuable functions: handling lengthy, complex tasks and its internal logic chains. Those are not trivial features.
They are the product of years of research and enormous computing expense. Once extracted, they can be repackaged into a competitor’s system at a fraction of the cost. Anthropic’s letter argued these attacks are happening at “industrial scale.” The implication is clear.
If Chinese companies can routinely siphon off the capabilities of US AI models, the technological edge the US has spent billions to build erodes quickly. The military dimension is direct. Anthropic cited other alleged attacks it said posed a threat to the US military.
An AI model that can handle complex reasoning is not just a commercial asset. It is a tool for logistics, intelligence analysis, and autonomous systems.
For anyone watching US-China tech rivalry, this fits a pattern. Mutual suspicion has defined the relationship for years. Washington accuses Beijing of theft.
Beijing accuses Washington of containment. This accusation moves the fight from semiconductors and 5G into the heart of AI.
The technology is newer. The methods of theft are harder to detect. A distillation attack leaves no stolen hardware.
It leaves no smuggled blueprint. It leaves only a pattern of queries — millions of them — against a public-facing model. Anthropic urged Congress to impose penalties on organizations that run such campaigns.
That is a direct call for legislation. It asks lawmakers to treat AI model extraction as a form of intellectual property theft with national security consequences.
Whether Congress acts remains an open question. The letter is a formal document, but it is also a political move. It puts the issue on the record.
It forces a response. Alibaba has not publicly responded to the specific accusation in the letter.
The company faces a serious allegation from a US firm with a track record in AI safety research. Anthropic was founded by former OpenAI employees. It has positioned itself as a responsible developer of advanced AI.
That credibility gives its accusation weight. The numbers are stark. Twenty-nine million interactions.
Thousands of accounts. A single target.
The scale suggests an organized operation, not random probing. If true, it means Alibaba — one of the world’s largest technology companies — ran a sustained campaign to strip a US competitor of its core intellectual property. This story is not about a minor breach.
It is about the fundamental architecture of how AI capability is acquired. Distillation attacks are not new.
But the accusation that a state-linked company is deploying them at this scale changes the conversation. It shifts the debate from hypothetical risk to documented theft. The technology relationship between the US and China has always been tense.
Now it has a new flashpoint, and it is not in a factory or a lab. It is in the data logs of a single AI model.
