# Anthropic vs Alibaba: what the Claude 'distillation attack' claim means

> Anthropic accuses Alibaba-linked operators of copying Claude's skills via 25,000 fake accounts; Alibaba denies the still-unproven claim.

*Anthropic says a rival copied Claude's brain using thousands of fake accounts. Alibaba says it didn't. Here's what's actually being alleged, in plain English.*

By The SuggestedTech Team · SuggestedTech
Canonical: https://suggestedtech.com/news/anthropic-alibaba-claude-explained

If you've seen headlines saying **Anthropic accused Alibaba of the 'largest known AI distillation attack'** and weren't sure what any of that meant, this is for you. In short: Anthropic, the company behind the Claude chatbot, told a **US Senate committee** on **10 June 2026** that operators linked to the Chinese tech giant **Alibaba** secretly used Claude on a huge scale to copy its abilities. Anthropic says it was about **28.8 million** conversations through roughly **25,000 fake accounts** over **44 days**. Two things to hold onto from the start: these are **accusations**, and **Alibaba says they aren't true**.

## First, what is 'distillation'?

'Model distillation' sounds technical, but the idea is simple. Imagine you can't get into an expensive tutor's class, so instead you send in thousands of students who each ask the tutor questions, write down every answer, and then use all those answers to train a cheaper copycat tutor at home. The copy never studied the way the real tutor did — it just learned to **imitate the outputs**. That's distillation: you train a smaller, cheaper AI model on a powerful model's answers so it behaves like the powerful one, for a fraction of the cost.

> **Info:** **In plain English:** distillation is copying an AI by studying its answers, not its blueprint. You don't need to see how Claude was built — you just collect enough of what it says and train a cheaper model to say similar things.

> Anthropic accused operators tied to Alibaba of moving to 'brazenly' and 'illicitly' extract Claude's most advanced capabilities, calling it the biggest such campaign it has detected.
> — [CNBC](https://www.cnbc.com/2026/06/24/anthropic-alibaba-distillation-campaign.html), 2026-06-24

## Why Anthropic is so worried about it

You might think the worst part is losing money or trade secrets. For Anthropic, the scarier bit is **safety**. When Anthropic builds Claude, it doesn't just teach it to be clever — it also spends a lot of effort teaching it what **not** to do: refuse dangerous requests, follow usage rules, and sit behind access controls. Anthropic's argument is that when you distil a model from Claude's answers, the copy learns the **skills** but not those **safety brakes**. So you could end up with a model that's nearly as capable as Claude at things like coding or cybersecurity, but far more willing to be misused. As they put it: the capability transfers, the guardrails do not.

## What's confirmed, and what isn't

This is the part to be careful about, because a lot of coverage blurs it. What's **confirmed** is that Anthropic made this accusation, in a letter to the Senate, and that a senator — **Elizabeth Warren** — repeated it, calling it the largest known distillation attack on Anthropic so far. Anthropic also says this one alleged campaign is bigger than three earlier ones it blamed on other Chinese labs (DeepSeek, Moonshot and MiniMax) back in February. What is **not** confirmed is the accusation itself: the 28.8 million figure, the 25,000 accounts, and Alibaba's involvement are all **unverified**, and **Alibaba denies them**. So the honest summary is: a big, serious claim that hasn't been proven and is being disputed.

> Coverage noted Anthropic told lawmakers Alibaba used roughly 25,000 fake accounts to distil Claude — a figure based on Anthropic's own account rather than independent confirmation.
> — [Forbes](https://www.forbes.com/sites/jonmarkman/2026/06/26/anthropic-says-alibaba-used-25000-fake-accounts-to-distill-claude/), 2026-06-26

One more piece of context worth knowing: this came out during a tense week when the US government was already pressuring Anthropic over how its most powerful models get released. That doesn't tell you whether the accusation is true — but it does explain why a dramatic claim about a Chinese rival landed in a letter to Congress rather than a quiet security report. When you read about it, keep the two apart: 'distillation is a real risk' is one thing; 'Alibaba definitely did this at this scale' is another, and that second part is still unproven. The sensible stance is neither to panic about a grand conspiracy nor to wave the whole thing away: take the safety concern seriously, treat the headline numbers as claims until someone independent checks them, and watch whether Alibaba or a regulator ever puts real evidence on the table.

## Key takeaways

- Anthropic says operators linked to Alibaba secretly used Claude to copy its abilities — about 28.8 million conversations through roughly 25,000 fake accounts over 44 days. These are accusations, and Alibaba denies them.
- The technique is called model distillation: you train a cheaper model on a stronger model's answers so it learns to imitate it — a bit like a student copying a brilliant tutor's every reply.
- Anthropic's real worry is safety: a copied model can get Claude's skills without Claude's safety training, rules on misuse or access limits — 'the abilities copy over, the safety brakes don't.'
- Anthropic says it aimed at Claude's coding, reasoning-agent and cybersecurity skills in a new model it calls 'Mythos Preview' — the most valuable and sensitive abilities to steal.
- None of the numbers has been independently confirmed, and Alibaba rejects the whole claim — so treat it as a serious but unproven dispute, not a settled fact.

## FAQ

### What exactly is Anthropic accusing Alibaba of?
Anthropic says operators linked to Alibaba secretly used Claude at huge scale to copy its abilities — about 28.8 million conversations through roughly 25,000 fake accounts over 44 days, aimed at Claude's coding, agent and cybersecurity skills. It calls this the largest known distillation attack on Claude. These are allegations, and Alibaba denies them.

### What is model distillation in simple terms?
It's copying an AI by learning from its answers. You collect lots of a powerful model's responses and use them to train a cheaper model to imitate it — like training a copycat tutor from a brilliant tutor's answers, without ever seeing how the original was built.

### Why does Anthropic think it's dangerous, not just unfair?
Because a copied model can pick up Claude's skills without Claude's safety training, misuse rules and access controls. Anthropic's line is that the capabilities copy over but the safety guardrails don't — so the copy could be capable but far easier to misuse.

### Are the numbers actually true?
They haven't been independently verified. The 28.8 million exchanges, the 25,000 accounts and Alibaba's involvement are Anthropic's claims, made to a US Senate committee. Alibaba denies them, so it's best treated as a serious but unproven accusation.

### Has this kind of thing happened before?
Anthropic says yes. In February 2026 it accused three other Chinese labs — DeepSeek, Moonshot and MiniMax — of large-scale distillation. It says the alleged Alibaba campaign is bigger than all three of those combined.

## Sources

- [Anthropic accuses Alibaba of campaign to 'brazenly' and 'illicitly' extract AI capabilities](https://www.cnbc.com/2026/06/24/anthropic-alibaba-distillation-campaign.html) — CNBC, 2026-06-24
- [Anthropic accuses Alibaba of 'largest known distillation attack' on Claude](https://asia.nikkei.com/business/technology/artificial-intelligence/anthropic-accuses-alibaba-of-largest-known-distillation-attack-on-claude) — Nikkei Asia, 2026-06-24
- [Anthropic Says Alibaba Used 25,000 Fake Accounts To Distill Claude](https://www.forbes.com/sites/jonmarkman/2026/06/26/anthropic-says-alibaba-used-25000-fake-accounts-to-distill-claude/) — Forbes, 2026-06-26