Anthropic accuses China's Alibaba of stealing Claude's AI capabilities

[DragonSlayer101]

In a nutshell: Anthropic has accused Chinese e-commerce giant Alibaba of illegally extracting capabilities from its Claude AI model using 25,000 fake accounts. According to the San Francisco-based AI startup, accounts operated by Alibaba's Qwen AI team engaged in more than 28.8 million conversations with Claude between April 22 and June 5, 2026, as part of a coordinated "distillation attack."

In a letter sent to Senators Tim Scott and Elizabeth Warren on June 10, Anthropic accused the Chinese tech giant of carrying out what it described as the largest recorded corporate espionage campaign against the company. It alleged that Alibaba sought to illicitly extract Claude's capabilities in order to train a smaller, less capable AI model to better respond to a wide range of queries.

A distillation attack, also known as a model extraction attack, is a form of intellectual property theft in which operators affiliated with an AI firm query a rival company's model through public APIs and use the outputs to train their own system. The attacker typically interacts with the target AI chatbot like an ordinary user, but at an industrial scale, using millions of carefully crafted prompts.

The letter further alleged that the attacks were designed to help Alibaba's Qwen AI model access some of the advanced capabilities of Anthropic's frontier "Mythos Preview" model, which the company claims represents a significant improvement over older systems like Claude Opus, particularly in areas such as coding and digital security.

Anthropic alleged that the Chinese government was complicit in the attacks as part of the country's broader push to assert global dominance in AI, machine learning, and related technologies. The company added that, if successful, such attacks could pose an existential threat to the United States and its allies worldwide.

The recipients of the letter, Senators Scott and Warren, are the chair and ranking member, respectively, of the US Senate Committee on Banking, Housing, and Urban Affairs. The committee held a two-day hearing on artificial intelligence titled "AI and the American Dream: Promoting Innovation, Affordability, and American Dominance" on June 11 and 12.

Two days after the letter was sent, the Commerce Department imposed strict restrictions on Anthropic's "Mythos" and "Fable" AI models over concerns that China and other blacklisted countries could use their advanced capabilities against US interests. Anthropic has since disabled public access to both models globally, pending clarification from the federal government.

It is worth noting that this is not the first time Anthropic has accused a Chinese AI company of attempting to steal its intellectual property. In February, the company claimed to have identified a campaign by operators affiliated with three Chinese AI startups, including DeepSeek, coordinating distillation attacks to improve their own AI models.

Permalink to story:

Anthropic accuses China's Alibaba of stealing Claude's AI capabilities

 

[Kirby1]

Sounds about right.

 

[LordVile95]

Oh no the AI company got its data stolen? Oh my how sad I am for them. It’s not like their entire business model is scraping the internet to steal, sorry, learn from others data or anything.

 

[Endymio]

Oh no the AI company got its data stolen? Oh my how sad I am for them. It’s not like their entire business model is scraping the internet to steal, sorry, learn from others data or anything.

I wonder if you'll still make light of the Chinese theft of trillions of dollars of IP and attempted domination of world industry when you're sitting in a CCP reeducation camp, being trained to stitch sneakers instead of playing Counterstrike.

 

[NumberNine]

The models use stolen data anyways.
This isn't art, it's a tool.

 

[Burty117]

I wonder if you'll still make light of the Chinese theft of trillions of dollars of IP and attempted domination of world industry when you're sitting in a CCP reeducation camp, being trained to stitch sneakers instead of playing Counterstrike.

Just to confirm, when China steals trillions of dollars of IP, China bad.

But if a USA firm effectively steals trillions of dollars of IP for AI training, all cool?

 

[Theinsanegamer]

I wonder if you'll still make light of the Chinese theft of trillions of dollars of IP and attempted domination of world industry when you're sitting in a CCP reeducation camp, being trained to stitch sneakers instead of playing Counterstrike.

Absolutely none of this whataboutism addresses the comment in question.

These AI tools took everything they could find to build themselves, scooping up publicly available data. So the Chinese made public accounts and scraped from Anthropic....and we're supposed to care?

 

[maxxcool7421]

One AI bro stole another AI bro's tech who stole from artists and coders on the internet without permission .. die in a fire Anthropic .. you are child killing scum.

 

[Endymio]

Absolutely none of this whataboutism addresses the comment in question.

These AI tools took everything they could find to build themselves, scooping up publicly available data. So the Chinese made public accounts and scraped from Anthropic....and we're supposed to care?

All this "data theft" vapidity has been long since addressed. Training an AI model on publicly available data isn't "stealing", no more than you're stealing a CNN story by watching it, then repeating what you heard to friends and family members. Do you not realize that -- with the exception of reviews by @Steve -- every article on this and most other tech sites -- is "stolen" in the exact same manner as AI models are trained?

And yes: if and when the CCP dominates AI and begins using it for military purposes -- yes, you'll care then. Of course, by then it'll be too late.

 

[J spot]

If Alibaba did this, is there an actual law that says that it's illegal? Is it actual espionage, or this language being used to make it sound very bad for better legal positioning?

 

[Burty117]

All this "data theft" vapidity has been long since addressed. Training an AI model on publicly available data isn't "stealing", no more than you're stealing a CNN story by watching it, then repeating what you heard to friends and family members. Do you not realize that -- with the exception of reviews by @Steve -- every article on this and most other tech sites -- is "stolen" in the exact same manner as AI models are trained?

And yes: if and when the CCP dominates AI and begins using it for military purposes -- yes, you'll care then. Of course, by then it'll be too late.

But it wasn't just free publicly available data was it?

All those books, movies, music, art, photo's etc.. most of it was essentially stolen, even Googling it, Googles own Gemini admits it's a contentious issue "bridging copyright law and technological development".

So the answer from you and big tech then is yes, it's absolutely fine to steal anything if it means AI training gets done.

 

[pnntmp]

I don't understand why Anthropic is whining about China stealing from them in letters to senators who can't do anything, instead of addressing the problem themselves.

If you know someone is stealing, just cut their access. Or silently switch to an older and less capable model. Or something else. China will not stop stealing, no matter how many letters are circulating.

 

[Endymio]

But it wasn't just free publicly available data was it?

You're confused. Any data that wasn't publicly available was purchased, including books, photos, and music tracks. Feeding a book you own into an computer algorithm doesn't violate copyright law, until or unless that algorithm reproduces it to others in its original form. Copyrights protect the particular expression of an idea, not the idea itself.

Nor have you acknowledged that every human professional -- no matter what profession they're in -- is "trained" on countless copyrighted materials. Your college textbooks are all copyrighted material ... do you in your job use the knowledge you obtained from them? Is that "stealing"?

Every musician, every novelist, every sculptor, every painter, every graphic artist -- they all base their own works on what they learned from those before them. What of this very article we're posting to? Do you think Techspot interviewed Anthropic and Alibaba directly? Or did they locate a publicly available story and, using a "trained" human, recast it into a slightly different form? Is that "data theft"?

 

[pnntmp]

These AI tools took everything they could find to build themselves, scooping up publicly available data. So the Chinese made public accounts and scraped from Anthropic....and we're supposed to care?

So, according to you, digging for gold, and stealing the gold someone else already extracted - these thing are equivalent? Seriously?

Anyone can collect the same publicly available data, process it, and train their model.

 

[Theinsanegamer]

So, according to you, digging for gold, and stealing the gold someone else already extracted - these thing are equivalent? Seriously?

Nope. Get a better argument or go back to the circus you came from.

Anyone can collect the same publicly available data, process it, and train their model.

Oh, so collecting publicly available data is OK? You mean like what China did, making accounts that anyone int he public could do and recording the results?

Hmmmm.........

All this "data theft" vapidity has been long since addressed. Training an AI model on publicly available data isn't "stealing", no more than you're stealing a CNN story by watching it, then repeating what you heard to friends and family members. Do you not realize that -- with the exception of reviews by @Steve -- every article on this and most other tech sites -- is "stolen" in the exact same manner as AI models are trained?

OK, so if the Chinese make accounts and record the data they get from Claude's output, then what is the issue? They recorded and repeated that data, so that isnt stealing. So where's the problem?

And yes: if and when the CCP dominates AI and begins using it for military purposes -- yes, you'll care then. Of course, by then it'll be too late.

So, again, if it is OK for the USA to do it, why is it not OK for China to do it?

 

[maxxcool7421]

All this "data theft" vapidity has been long since addressed. Training an AI model on publicly available data isn't "stealing", no more than you're stealing a CNN story by watching it, then repeating what you heard to friends and family members. Do you not realize that -- with the exception of reviews by @Steve -- every article on this and most other tech sites -- is "stolen" in the exact same manner as AI models are trained?

And yes: if and when the CCP dominates AI and begins using it for military purposes -- yes, you'll care then. Of course, by then it'll be too late.

If anthropics data is available to steal then its public enough ... nobody cares because trump is doing more damage than the ''Chinese'' ever will ..

 

[Endymio]

Oh, so collecting publicly available data is OK? You mean like what China did, making accounts that anyone int he public could do and recording the results?

Wrong again. China created several thousand fraudulent accounts to conduct thirty million data exchanges, violating not just Anthropic's TOS but federal law (18 U.S. Code § 1030.) China then used those results to calculate weights -- hidden, private, proprietary data -- in Anthropic's model. Anthropic's claim doesn't rest on copyright law, but on theft of trade secrets, fraudulent system access, and national security concerns.

Got it straight now?

 

[Notmare]

Whaaaburgers and french cries all around!

 

[Theinsanegamer]

Wrong again. China created several thousand fraudulent accounts to conduct thirty million data exchanges, violating not just Anthropic's TOS but federal law (18 U.S. Code § 1030.) China then used those results to calculate weights -- hidden, private, proprietary data -- in Anthropic's model. Anthropic's claim doesn't rest on copyright law, but on theft of trade secrets, fraudulent system access, and national security concerns.

So like how Anthropic copies works protected by federal copyright law that can be found online and used them to train their model?

If China can reverse engineer your private proprietary data from how your respond to inquiries, you suck at engineering and deserve to have your trade secrets stolen. They paid for those accounts, so cry me a river, Anthropic got their money.

Got it straight now?

Yeah yeah we got it, if USA steals things thats good, but if China does it that's evil.

 

[LordVile95]

So, according to you, digging for gold, and stealing the gold someone else already extracted - these thing are equivalent? Seriously?

Anyone can collect the same publicly available data, process it, and train their model.

Publicly available doesn’t mean copyright free, for example on my personal computer I have a copt of Fusion 360 under their personal, non commercial licence. Google for example would not be able to get that software for free like I can because the licensing terms do not allow it, they would have to pay for a licence.

Just because information is publicly available doesn’t mean it’s free for commercial use

 

[protodeus]

When they scrape (steal) data they call it fair use. When a foreign company does the same thing they call it theft. And they don't think we notice the double standards?

 

[protodeus]

Qwable and Qwythos use this dataset I believe.

 

[Ramzz]

You're confused. Any data that wasn't publicly available was purchased, including books, photos, and music tracks. Feeding a book you own into an computer algorithm doesn't violate copyright law, until or unless that algorithm reproduces it to others in its original form. Copyrights protect the particular expression of an idea, not the idea itself.

Nor have you acknowledged that every human professional -- no matter what profession they're in -- is "trained" on countless copyrighted materials. Your college textbooks are all copyrighted material ... do you in your job use the knowledge you obtained from them? Is that "stealing"?

Every musician, every novelist, every sculptor, every painter, every graphic artist -- they all base their own works on what they learned from those before them. What of this very article we're posting to? Do you think Techspot interviewed Anthropic and Alibaba directly? Or did they locate a publicly available story and, using a "trained" human, recast it into a slightly different form? Is that "data theft"?

I think you are the one confused if you think they purchased those books to train their chat bots. Is this why settled for 1.5 billion with the authors to avoid going to trial?

https://www.theguardian.com/technology/2025/sep/05/anthropic-settlement-ai-book-lawsuit

 

[Liranan]

25 thousand paying accounts are fake? The desperate propaganda is of such low quality it's terrible. Anyone can do what they accuse 'China' of doing: pay and then record the reasoning. You can do it with Chinese models too but that's just competition, amirite?

Wrong again. China created several thousand fraudulent accounts to conduct thirty million data exchanges, violating not just Anthropic's TOS but federal law (18 U.S. Code § 1030.) China then used those results to calculate weights -- hidden, private, proprietary data -- in Anthropic's model. Anthropic's claim doesn't rest on copyright law, but on theft of trade secrets, fraudulent system access, and national security concerns.

Got it straight now?

You can do it with any model, genius. Recording reasoning and calculating the weight isn't that hard.

 

[Liranan]

I wonder if you'll still make light of the Chinese theft of trillions of dollars of IP and attempted domination of world industry when you're sitting in a CCP reeducation camp, being trained to stitch sneakers instead of playing Counterstrike.

The biggest IP thieves are the US (Operation Paperclip) through the NSA (Perkins) and the Mossad.
Guy Galanti (Mossad) was recently arrested for stealing semiconductor trade secrets from US companies. Where is your faux outrage?

If the Chinese are so good at stealing secrets why haven't they reverse engineered ASML machines? Why not reverse engineer or copy chips rather than license from AMD and try to create their own? I've asked these questions before but you propagandists refuse to answer because the answer is obvious.