Anthropic alleges – Anthropic says Alibaba-affiliated operators carried out what it calls the largest known distillation attack against it—using nearly 25,000 fraudulent accounts to run 28.8 million requests for Claude between April 22 and June 5. The company’s head of policy has
By June 10, Anthropic had put a number on the theft it says happened—28.8 million exchanges with its Claude model—then tied it to a method the company says is designed to copy advanced capabilities without paying the cost of building them.
In a letter to South Carolina Sen. Tim Scott and Massachusetts Sen. Elizabeth Warren. Sarah Heck. Anthropic’s head of policy. accused one of China’s biggest technology companies—Alibaba—of carrying out what she called “the largest known distillation attack” to date. Heck wrote that Alibaba-affiliated operators tried to “illicitly extract Claude’s capabilities” in order to train Alibaba’s own models.
Heck said the activity took place between April 22 and June 5, carried out through almost 25,000 fraudulent accounts. She described it as systematic and industrial in scale—an effort, in Anthropic’s view, to harvest U.S. AI capabilities across frontier labs and then repackage them as the operators’ own.
Distillation attacks, Heck wrote, use an advanced AI model to train and improve a less advanced one. Anthropic argues the key distinction is that these attacks are often conducted “illicitly. systematically. and at industrial scale.” In her letter. Heck also said the attacks could help Chinese models reach “Claude Mythos Preview-level capabilities” sooner. Mythos is described by Anthropic as one of its most advanced large language models. capable of detecting software vulnerabilities and outperforming humans on cybersecurity tasks.
Heck’s letter didn’t just lay out the allegation. It also pressed lawmakers for action, asking for more legislation aimed at preventing further distillation attacks. She suggested steps including limiting China’s access to advanced U.S. computing infrastructure and penalizing Chinese entities that launch such efforts.
The push comes after additional U.S. restrictions and penalties aimed at advanced AI capabilities. Several weeks earlier, the U.S. government slapped an export control on Anthropic’s latest Fable 5 model, barring foreign individuals from accessing it and citing national security risks.
For Alibaba, the accusations arrive as the company absorbs other recent blows. Alibaba was also recently added to a Pentagon blacklist—a list of businesses the defense department linked to the Chinese military. On Tuesday, Alibaba sued the U.S. government over that designation.
As of press time on Thursday, Alibaba’s share price had dropped more than 4%. Representatives for Alibaba did not respond to a request for comment.
Anthropic Alibaba Claude Sarah Heck Tim Scott Elizabeth Warren distillation attack Mythos Fable 5 export control Pentagon blacklist cybersecurity Qwen Alibaba Cloud
So basically they copied Claude using fake accounts? That sounds shady as hell.
I don’t even know what “distillation” means but 28.8 million sounds like a lot. If it’s true then why is this even allowed on the internet.
Wait so Anthropic built it, then Alibaba trained on it, then the govt is like “ban the computers”?? Also the article says export control on Fable 5, but people already use models online right? Sounds like everyone just copies everyone.
This is why I hate all these AI companies. One minute they’re “innovating” the next minute it’s like industrial espionage with 25,000 fake accounts. Tim Scott and Elizabeth Warren always get involved once it’s a scandal, like they weren’t aware before. Also Alibaba suing the U.S. makes me think they’ll just say “it’s fine” and drag it out forever. Not saying Anthropic is wrong, but it feels like politics and trade stuff mixed in.