📚 New to this topic? Read our full guide: Generative AI Explained.
Anthropic Accuses Alibaba of AI Model Theft: 28.8 Million Claude Chats Harvested
Anthropic, the US company that makes the Claude AI chatbot, has accused China’s Alibaba of a huge case of AI model theft. In a letter to US lawmakers, Anthropic said operators linked to Alibaba’s AI lab secretly held about 28.8 million chats with Claude. They used roughly 25,000 fake accounts to do it. Anthropic calls this the biggest theft attempt of its kind it has ever seen.
This is not normal copying. The accusation is about a trick called “distillation.” In simple words, distillation means training a weaker AI by feeding it the answers of a smarter AI. So instead of building a clever model from scratch, you copy how a better model thinks. Anthropic says Alibaba did this to Claude without permission.
What Anthropic Says Happened
Anthropic sent a letter dated June 10, 2026, to the US Senate Banking Committee. It went to the chairman, Tim Scott, and the senior Democrat, Elizabeth Warren. The letter came just before a congressional hearing on AI.
In the letter, Anthropic says the campaign ran from April 22 to June 5, 2026. That is about six weeks. The company blames Alibaba’s Qwen lab, the team that builds Alibaba’s own AI models. Anthropic says the operators used around 25,000 fake accounts. They also used commercial proxy services. A proxy service hides where internet traffic really comes from. This let them dodge the rule that blocks Chinese users from using Claude.
Across those weeks, the fake accounts ran about 28.8 million chats with Claude. Anthropic says the goal was to collect Claude’s answers in bulk and use them to train a rival model. This is what makes it an “adversarial distillation” attack. The word “adversarial” just means it was done against the rules, on purpose.
Key facts
| Item | Reported figure |
|---|---|
| Claude chats harvested | ~28.8 million |
| Fake accounts used | ~25,000 |
| Campaign window | April 22 – June 5, 2026 (about 6 weeks) |
| Letter date | June 10, 2026 |
| Sent to | Sen. Tim Scott, Sen. Elizabeth Warren (Senate Banking Committee) |
| Accused lab | Alibaba’s Qwen AI team |
How This Compares to Past Cases
This is not the first time Anthropic has flagged such activity. Earlier, the company pointed to three other Chinese AI labs. But the Alibaba case is far bigger than all of them put together.
| Lab flagged earlier | Reported interactions |
|---|---|
| DeepSeek | 150,000+ |
| Moonshot AI | 3.4 million |
| MiniMax | 13 million |
| Alibaba (Qwen) | 28.8 million |
What it means: the Alibaba figure alone is more than twice the next-biggest case. That is why Anthropic told the Senate this is the largest such campaign it has ever recorded.
The Export Control Backdrop
This fight is part of a bigger US-China AI battle. The US Commerce Department has put limits on Anthropic’s most advanced models, called Mythos and Fable. (Export controls are government rules that block certain technology from being sold or shared abroad.) Officials worry these powerful models could end up with military or intelligence users in China.
So while the US tries to keep its best AI inside trusted hands, Anthropic claims rivals are trying to copy that AI through the back door. This is also why other countries are now negotiating their own deals for access. India, for example, is in talks with the US to reach Anthropic’s Fable 5 model through proper channels.
FAQ
What is AI distillation in simple words?
It is training a smaller AI using the answers of a smarter AI. The weaker model learns to copy the better one without building that skill itself.
How did the accounts get past the China block?
Anthropic says the operators used about 25,000 fake accounts and proxy services. A proxy hides the real location of internet traffic, so the system thought the users were elsewhere.
Has Alibaba responded?
As of the reports, Alibaba had not replied to requests for comment. Anthropic also did not give extra comment beyond the letter.
Why It Matters (Especially for India and Founders)
For Indian startups and founders, this case carries two clear lessons. First, the data and outputs of an AI product are valuable assets. If you run an AI service, others may try to copy your model by flooding it with prompts. Rate limits, account checks, and abuse monitoring are no longer optional.
Second, access to top AI models is now a geopolitical issue. The best models may be locked behind export rules. Indian companies that want frontier AI may need to go through government-to-government deals, not just sign up online. That makes home-grown AI work and reliable partnerships more important than ever.
The takeaway is simple. AI capability has become something nations guard like gold. As models get more powerful, the line between fair learning and outright theft will be tested again and again. The Anthropic-Alibaba clash is likely just the first of many such fights.
Sources: Benzinga, CNBC and Financial Express.
