← All news·2026-06-25·5 min read

Alibaba deployed 25k fake accounts to replicate Claude

Alibaba attempted to copy 28.8M Claude conversations (across almost 25,000 fake accounts) in order to steal Claude’s advanced reasoning abilities and coding skills to improve their own AI model.

anthropicclaudealibabaбезопасностьdistillation

From April 22nd until June 5th, Alibaba and their Qwen AI lab attempted to engage in 28.8 million conversations with Claude using almost 25,000 fake accounts. This was the biggest known effort yet to distill a cutting edge AI model. In a letter to US senators dated June 10th, Anthropic claimed. The goal of this attack was to steal Claude’s advanced reasoning abilities, multi-step tasks, and software engineering skills.

According to Anthropic, the attackers made use of proxy networks and obfuscation methods to hide their activities. They also made use of what is known as ‘distillation’. Essentially, this involves asking a strong AI model many questions and gathering its responses, then using those responses to train another model which replicates the behavior of the first model, but without having to pay for any of the R&D. Previously, there had been a similar effort by DeepSeek, Moonshot, and MiniMax which involved generating over 16 million Claude interactions using around 24,000 accounts. This recent effort by Alibaba was close to doubling that number.

Why this technique is so hard to stop

In order to carry out a distillation attack, you don’t need to find any bugs in the software or gain access to physical devices. All you need is some accounts (thousands of them), and time. The attackers will act exactly like regular customers, only on a much larger scale. It’s difficult for us to distinguish between someone who wants to experiment with Claude’s features and an automatic system trying to systematically extract information from it. This is why we need to look at statistics across millions of sessions rather than just putting up a firewall.

Anthropic positioned this issue within a wider context. Any time someone successfully pulls off a distillation attack, they take hundreds of billions of dollars worth of American AI research and turn it into a free dataset which anyone can use to train their own models. In their letter, they asked Congress to consider three steps. The first was to modify antitrust laws so that AI companies are allowed to share information about threats with one another. Right now, collaborating on security measures between competitors is considered legally ambiguous. Secondly, they wanted stricter regulations on exporting high performance semiconductors to make it more difficult to train models on the data collected. Finally, they wanted to introduce severe penalties for labs that perform such attacks, thus making them less profitable.

What it means for people who build with Claude

For those of you using Claude via API or Claude Code, this won’t affect how you interact with Claude. But there are serious issues happening underneath the surface. Our main advantage is our research: all the safety checks, training methods, alignment ideas. With distillation attacks, other companies can get pretty much the same results without spending much money. This means that we have to charge more for our services because we have to spend more money on security and detection. It also leads to a kind of arms race where once a model reaches a certain level of capability, it gets targeted immediately after launch.

Meanwhile, Alibaba is also suing the Trump administration over another classification, claiming that they have no connections to the military. These two cases are happening independently of each other. According to Anthropic, they have “confidential evidence” regarding this matter which they felt comfortable sharing with Congress, although they have not yet brought a civil action against Alibaba.

ℹWhat I'd actually do

If you run a product on top of Claude, this is a reminder to respect the terms of service and build as if Anthropic can see exactly what you are doing — because increasingly, it can. If you are worried about the stability of your Claude-based product given geopolitical dynamics, it is worth investing in abstraction layers that let you swap models without rewriting your entire integration. The capability gap between Claude and open-weight models is narrowing precisely because of stories like this one.

Source: arstechnica.com

Free course

Stop reading about AI — start building with it

The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.

Start free →
EAEvgenii Arsentev

Author

Evgenii Arsentev

PhD · Founder & CEO, Dad4Kids