Alibaba Bans Claude Code Internally: How the Detection-Code Dispute Unfolded

Alibaba Bans Claude Code Internally (What Was Reported)

On July 4, 2026, Hong Kong's SCMP and US-based TechCrunch reported that Alibaba would ban internal use of Claude Code from July 10. Claude Code is an AI coding assistant from Anthropic that runs on the command line (a text-based interface). Let us start with what exactly was reported.

The "high-risk software" designation and internal notice

According to the internal notice reported by SCMP, Alibaba told staff that back-door risks had recently been found in Claude Code, and that after a comprehensive evaluation it had been added to a list of high-risk software with security vulnerabilities. The starting point for understanding this story is that the ban is reported as one company's internal rule, not a government regulation.

The replacement is the in-house "Qoder"

Staff were reportedly directed to Qoder, Alibaba's in-house AI coding tool. Reports say the notice went beyond stopping Claude Code use by July 10 and also instructed employees to delete Claude models from company systems. In other words, it was framed not as a simple tool swap but as removing Anthropic's products from inside the company.

The Trigger: the June 30 Discovery of "China-User Detection Code"

The direct trigger for the ban was the discovery that Claude Code contained code quietly judging whether a user was based in China. Let us look at who found it and what it did.

Detecting China-based users by timezone and proxy

The code came to light through a reverse-engineering analysis posted to Reddit on June 30. According to that analysis, the code had been present since version 2.1.91 (released April 2); it checked whether the system timezone matched Asia/Shanghai or Asia/Urumqi and matched proxy (relay-server) URLs against a list of Chinese corporate networks. That list reportedly included networks and cloud regions tied to Alibaba, Baidu, ByteDance, and Moonshot AI.

What stands out is that rather than transmitting the result directly, the code used steganography (hiding information inside other data). The analysis says it quietly swapped the date format and the apostrophe character inside the system prompt, thereby "watermarking" interactions from users judged to be China-linked.

Anthropic's explanation: a "misuse-prevention experiment"

After the discovery, Thariq Shihipar, an engineer on Anthropic's Claude Code team, explained that the code had started as an experiment to prevent account misuse and distillation attacks, and confirmed its removal. The code was indeed taken out in the July 1 update. What is confirmed is only that the code existed, was uncovered on June 30, and was removed on July 1; the full intent behind it rests on the parties' own accounts.

The Backdrop: Anthropic's "Distillation Attack" Accusation

The detection code did not appear out of nowhere; Anthropic's accusation came first. In the timeline, that accusation is the first point of conflict.

~25,000 fake accounts, 28.8 million exchanges

Anthropic reportedly told the US Senate that operators linked to Alibaba used about 25,000 fraudulent accounts to run more than 28.8 million exchanges with Claude between April 22 and June 5, describing it as the largest known attack of its kind against the company. The aim was "distillation" — using a high-performing model's outputs to train another model and replicate its abilities at lower cost.

If distillation is unfamiliar, it helps to picture it as "having a stronger model answer a huge volume of questions, then using those answers as study material for your own model." Anthropic attributes the campaign to operators linked to Alibaba's AI lab, Qwen.

Mutual exclusion of US–China AI tools, and the impact on everyday users

Laid out in sequence, the events form a pattern of mutual exclusion: Anthropic accuses a party of misuse and responds with detection code, and the exposed Chinese side then removes the US tool from inside the company. In the US Congress, there is also reported movement toward legislation that could sanction those who run large-scale distillation attacks — so the dispute between companies is widening into a policy question.

At the same time, the ban is Alibaba's internal rule; it does not mean everyday users lose access to Claude or Claude Code. On Anthropic's side, identity verification introduced from July 2026 is already in motion as an account-misuse countermeasure. For details, see the guide to Claude's identity verification. A separate strand — the regulatory back-and-forth between Anthropic and the US government — is covered in the Claude Fable 5 restoration article.

Conclusion: Read the Confirmed Facts and the Reporting Separately

Finally, it helps to sort this episode by how firmly each part is established.

So the facts around the detection code are backed by analysis and the parties' acknowledgment, while Alibaba's ban itself rests on reporting of an internal notice, with no comprehensive official statement yet from either company.

The picture of US and Chinese AI companies warily fencing off each other's tools is likely to continue in different forms. There is no direct practical impact for everyday users, but it is an instructive case for understanding how AI tools' terms of use and anti-abuse measures work. For the full picture of Claude's model lineup and pricing, see the guide to Claude.

Free ToolURL to Markdown ConverterConvert any public web page URL to Markdown. Preserves headings, tables, lists, and links — perfect for LLM and RAG preprocessing, research notes, and archiving web articles.Try it now →