China's vulnerability database warns of backdoor in Anthropic's Claude Code versions 2.1.91–2.1.196

Full text

A cybersecurity firm associated with the Chinese government reportedly has concerns about Anthropic's Claude AI models and has warned users against trusting it, at least before updating to the latest version.

As reported by Reuters , the Chinese National Vulnerability Database (NVDB) claims that recent Claude Code versions have a monitoring mechanism that transmits sensitive information to remote servers without consent from the user. This allegedly includes geographic location data and user-identifying information. This was all posted via the agency's official WeChat account, according to the outlet.

It reportedly affects users on version 2.1.91 and goes up to version 2.1.196. 2.1.196 was eventually replaced by 2.1.200 on July 3. The NVDB encourages users to either uninstall if they use the impacted version, or simply update.

As well as this, the NVDB has taken this opportunity to urge users to tighten controls and strengthen monitoring techniques for the data they allow Claude Code access to. I have reached out to Anthropic for comment on this story, but have not yet received a response.

This is not the first time that Anthropic's coding tool has drawn ire in China this month. Last week, Alibaba reportedly banned its employees from using Claude at work due to fears that it could identify Chinese users with information fed to it.

(Image credit: Getty Images) Anthropic's story with Alibaba goes deeper than this, with the Claude creator accusing Alibaba of training its AI models on Anthropic's work. Anthropic claims that Alibaba, between the end of April and the start of June, generated over 28 million exchanges with Claude through close to 25,000 fraudulent accounts to harvest information.

Interestingly, Anthropic has been blocking access to Chinese companies for some time, and those companies have been reportedly getting around this restriction through overseas subsidiaries and even plain old VPNs. Anthropic is, according to the Financial Times , looking to crack down on these workarounds in the future.

The relationship between Anthropic and China seems pretty strained right now, and it seems likely that Chinese agencies are looking to swap to homegrown AI models sooner rather than later.