American AI giantAnthropicA scathing letter was recently submitted to the U.S. Senate accusing Chinese technology companies ofAlibabaSystematically stealing its flagship model through a large-scale “model distillation” attackClaudecore competencies. This incident, known as the "largest-scale AI model extraction operation in history," has pushed the Sino-US AI competition to a new level of tension.
28.8 million conversations, 25,000 fake accounts
According to a letter Anthropic submitted to the Senate Banking Committee on June 10, Alibaba’sTongyi Qianwen(Qwen) Laboratory operator, inApril 22 to June 5Within six weeks, by approx.25,000 fake accounts,andClaudecarried out more than28.8 million timesDialogue exchange. This scale is nearly twice the combined size of the other three Chinese AI lab attacks Anthropic discovered in February this year.
What is "model distillation"?
The keywords used by Anthropic areDistillation. This is a technique that uses the output of a powerful model to train a weaker model - without touching the source code, model weights, or training data, the attacker can simply ask the target model millions of carefully crafted questions, record the answers, and then use these question and answer pairs to train his own model.
The key issue is: the cloned modelInherited abilities, but not inherited safety guardrails. originalClaudeSecurity training, usage policies, and access controls built into the model are all lost during the distillation process. This means that the distilled model may have close to theClaudepowerful ability.
Highly precise attack targets
Unlike the previous attack discovered in February, this operation did not generally extractClaudegeneral ability, butAimed accuratelyClaudeCore areas with the most commercial value: Agentic Reasoning, software engineering and long-term task processing. These are key capabilities that Anthropic has invested heavily in research and development and are the most difficult to replicate.
Industry Impact and Controversy
Anthropic's choice to submit the letter to Congress instead of the court itself illustrates the complexity of the matter. The output of current AI models is not as explicitly protected legally as code or copyrighted text, which is why Anthropic is pursuing legislation rather than litigation. Alibaba, meanwhile, has yet to publicly respond to the accusations.
Analysts pointed out that this incident has a direct warning to developers who use APIs to build products: the intelligence you rent may be copied, and the security controls carried with the model may disappear after copying. Chief analyst at Greyhound Research warns,“Enterprise supply chains no longer stop at software, APIs and cloud zones, but now also include rented intelligence – and rented intelligence can be replicated and redeployed.”
Summarize
Regardless of whether Anthropic's specific accusations against Alibaba can ultimately be proven, "model distillation" as an extremely low-cost means of replicating capabilities is escalating at an alarming rate. From 16 million in February to 28.8 million in June, the growth curve of attack scale itself is the clearest warning. For developers relying on cutting-edge APIs, model security is no longer an optional item – it is becoming a core topic in supply chain risk management.
AI Dash — Discover the most useful AI tools and track the latest trends in the AI industry.