China Issues Warning on Security Risks from Anthropic’s Claude Code AI Tool
Table of Contents
You might want to know
Could widely used AI coding assistants transmit sensitive user data without consent?
How are authorities and companies responding to potential security vulnerabilities in foreign AI tools?
Main Topic
China has issued a formal warning about potential security vulnerabilities in Anthropic’s AI coding assistant, Claude Code. According to a statement from China’s Ministry of Industry and Information Technology, national cybersecurity monitoring detected a "back‑door" issue in certain releases of the tool that could allow data to be transmitted to remote servers without the user’s knowledge or permission. The advisory highlights concerns that the tool might expose user identity and location data among other sensitive information.
The advisory specifically identified affected releases by version numbers and recommended that users either uninstall or upgrade from the impacted builds. The cybersecurity notice covered versions 2.1.91 through 2.1.196, which correspond to updates issued between early April and late June, according to Anthropic’s release history. Anthropic’s platform indicated a later build was available at the time of the advisory.
The warning arrived amid heightened tensions in the broader U.S.‑China technology competition. Last month Anthropic accused the Chinese technology company Alibaba of attempting to extract aspects of its AI capabilities — a claim Alibaba did not publicly address at the time. Separately, Chinese tech firms and developers have continued to find ways to access U.S. AI services, and some local developers have reported active use of Claude Code in China. In response to the evolving security concerns, Alibaba instructed employees to stop using Anthropic’s tools for work starting July 10, according to media reports.
From a cybersecurity perspective, the alleged vulnerability is notable because coding assistants often interact with local codebases, development environments, and configuration files. If a tool were to transmit snippets of code, metadata, or environment details to an external endpoint without explicit consent, it could create both intellectual property and privacy risks for organizations and individual users. This possibility — unauthorized exfiltration of sensitive information — is the core concern prompting official advisories.
Anthropic had not immediately provided a public comment in response to requests for clarification at the time of the advisory. Meanwhile, authorities urged affected users to take remedial action: remove the vulnerable versions or apply updates that address the reported security issue. Organizations that depend on third‑party AI tools are being reminded to include such tools in their security assessments, maintain software inventory and version control, and enforce usage policies that reduce exposure to unknown vulnerabilities.
Key Insights Table
| Aspect | Description |
|---|---|
| Reported Risk | A "back‑door" vulnerability in Claude Code that could send data to remote servers without user consent. |
| Affected Versions | Versions 2.1.91 through 2.1.196 (released between April 2 and June 29). |
| Recommended Action | Uninstall or upgrade from the affected versions; organizations should enforce tool usage policies and monitoring. |
| Context | The advisory comes amid intensified U.S.‑China tech competition and prior accusations around attempts to extract AI capabilities. |
| Vendor Response | Anthropic had not immediately responded to requests for comment at the time of reporting. |
Afterwards...
Looking forward, the episode underscores the importance of rigorous supply‑chain and third‑party risk management for AI tools. Enterprises and governments should prioritize secure deployment practices, continuous monitoring, and transparent vendor communications. Further investment in tool verification, runtime inspection, and provenance tracking for AI models and applications can help reduce the likelihood of covert data flows. Greater emphasis on software bill of materials (SBOMs), formal verification where feasible, and routine security audits are practical steps to improve resilience.
At a policy level, the incident may accelerate dialogues about cross‑border data flows, export controls for advanced AI capabilities, and international standards for AI security. For developers and security teams, adopting a zero‑trust approach to third‑party integrations and treating AI assistants as components subject to the same controls as other networked software will help manage risk. Continued research into AI system transparency and robust telemetry that preserves privacy while enabling security investigation will also be valuable as reliance on such tools grows.