China Warns of a 'Backdoor' in Claude Code, Demanding Restrictions Over Alleged Location Data Leaks

According to Reuters, Chinese authorities have spent the past month meeting with major domestic AI firms -- Alibaba, ByteDance, and Zhipu AI (Z.ai) among them -- to discuss whether Beijing should restrict overseas access to China's most advanced AI models, including ones not yet released. Unlike ordinary content moderation, this marks a real reversal for Chinese AI, which built its global following on being cheap and easy to access. Alibaba and ByteDance are reportedly set to shut down user-created agents on July 15th.

On July 8th, China's National Vulnerability Database (NVDB), operated by the industry ministry, warned that Claude Code versions 2.1.91 through 2.1.96 carry a serious backdoor risk. The warning claims Claude Code has a built-in monitoring mechanism capable of transmitting sensitive data -- including geographic location and identity identifiers -- to remote servers without user consent. NVDB told organizations to review affected systems, uninstall the impacted versions or upgrade to a secure release with the alleged code removed, and tighten external network access and traffic monitoring for development tools.

Anthropic did not immediately respond to Reuters' request for comment. The warning followed shortly after Alibaba had already banned employees from using Claude Code, citing features that could identify China-linked users.

In a separate story, test results show Fable 5 acting as an 'orchestrator' -- handling only the high-level judgment calls while delegating the heavy lifting to cheaper Sonnet 5 sub-agents. On SWE-Bench Pro, this setup reached about 92% of standalone Fable 5's performance at roughly 63% of the cost. On a fact-checking task across 10 major US national parks, a single Fable 5 run cost about $4 and took 68 seconds, while the orchestrated setup cost about $161 and took 194 seconds -- but pushed over 80% of tokens to cheaper worker models.