Alibaba Orders Full Uninstall of Claude Code After China's Security Agency Flags Backdoor

Follow my public account: [编程朝花夕拾] to get first-release content.

01 Introduction

I didn't pay attention when I saw the news two days ago that Alibaba required uninstalling Claude Code. Today I have time to dig into the whole story. What made Alibaba order all employees to uninstall Claude Code?

02 Origin

On July 3, 2026, multiple media outlets cited sources inside Alibaba, saying that Alibaba had issued a notice to all employees: starting July 10, all Claude products from Anthropic (including models like Sonnet, Opus, Fable, and agent tools like Claude Code) would be completely banned, and Alibaba's self-developed Qoder was recommended as a replacement.

Just 5 days later, on July 8, the NVDB (National Information Security Vulnerability Database) platform under the Ministry of Industry and Information Technology officially issued a risk warning, publicly "naming" Claude Code.

03 Event Timeline

Timeline information comes from the internet:

• 2026.4: Anthropic released Claude Code version 2.1.196 (allegedly containing reporting code targeting Chinese users)
• 2026.6: A Reddit user "reverse-engineered" Claude Code, revealing a built-in monitoring mechanism that silently transmits user location, identity, and other information back to servers
• 2026.7.1: The head of Claude Code publicly responded, arguing it was "only to prevent model distillation"
• 2026.7.3: Alibaba internally issued a notice: starting July 10, all Claude products are completely banned; no official public response yet
• 2026.7.8: MIIT NVDB issued a risk warning, publicly naming Claude Code as having a "backdoor implantation" risk
• 2026.7.10: Alibaba's ban officially took effect

04 Reasons for Alibaba's Ban

4.1 The Core Trigger

Claude Code was exposed for having a "hidden backdoor." A Reddit user reverse-analyzed the Claude Code client and found that versions 2.1.91 to 2.1.196 contained a segment of monitoring code specifically targeting Chinese users:

• Silently collected sensitive information such as local code, project files, directory structures, and development logs
• Transmitted user location, identity identifiers, and other data to remote servers without user authorization
• This code was only triggered by Chinese IPs, Chinese time zones, and Chinese user characteristics, leading the industry to describe it as "spy code"

Anthropic's official explanation was "to prevent Chinese teams from conducting model distillation"—but this statement not only failed to quell the controversy but actually confirmed the fact of "targeted surveillance": identifying Chinese users specifically to intercept distillation is itself an overreach in means.

As a large company that takes data security extremely seriously, measures were naturally taken.

4.2 MIIT NVDB Risk Warning

The regulatory level delivered the final word. On July 8, the MIIT NVDB platform officially issued the "Circular on the Security Risks of Claude Code," with a very severe characterization:

• Confirmed existence of a monitoring mechanism: Transmitted sensitive information to overseas servers without user consent
• Clearly identified affected versions: 2.1.91 to 2.1.196
• Disposal recommendations: Relevant units should immediately investigate, uninstall or upgrade to versions where the backdoor code has been removed, and strengthen external connection permission controls and traffic monitoring for development tools

MIIT publicly named it + recommended uninstallation—the weight of this signal far exceeds any single company's internal ban. For Alibaba, a top-tier internet company, source code security is a "zero tolerance" red line. After the regulatory naming, the ban went from "optional" to "mandatory."

Original MIIT text:

Address: https://www.nvdb.org.cn/publicAnnouncement/2074681830578630657

4.3 Industry Dynamics

This is not an isolated incident. Since 2026, Anthropic has continuously filed "anti-distillation" lawsuits against Chinese AI companies and proactively blocked all channels allowing Chinese users to use Claude. Banning Claude Code is Alibaba's counter-move in this confrontational landscape:

• Refuse to continue contributing any collectible data to Anthropic
• Cut off the channel for foreign vendors to "reverse-engineer Chinese developers' work content through IDE clients"
• Use this opportunity to promote their own Qoder (Alibaba Cloud's Agentic Coding platform), strengthening domestic substitution

05 Insights

In the big data era, possessing more data has profound significance whether for training AI or making decisions. Especially for large companies, once core business code, algorithm models, and internal architecture are leaked, the losses are immeasurable. Data security is extremely important.

The rise of AI originated abroad, and during use, many developers may also feel that foreign models or products are better. But there are certain risks in using them. Domestic alternatives have also been launched one after another, such as Alibaba's Qoder, ByteDance's Trae, Tencent's CodeBuddy, Huawei's CodeArts... Domestic Agentic Coding tools are experiencing a concentrated explosion.

Enterprise-level AI governance will make auditability, private deployment capability, and "data not leaving the domain" hard indicators for tool selection.

Folks, are you still using Claude Code?