China’s National Vulnerability Database (NVDB) warned on 8 July that versions 2.1.91 to 2.1.196 of Anthropic’s AI coding tool Claude Code contain what it described as a “security backdoor” capable of transmitting users’ location and identity data to remote servers without consent. The state-affiliated platform advised immediate uninstallation or upgrade. Within days, Alibaba told staff it would prohibit use of the tool from 10 July, citing security concerns, according to people familiar with the matter.

Anthropic engineer Thariq Shihipar responded on social media that the monitoring code was part of a March experiment to prevent account abuse by unauthorised resellers and to protect against model distillation—a practice the San Francisco startup has previously accused Chinese AI labs, including Alibaba, of conducting. Shihipar said stronger mitigations were since deployed and the code would be fully removed in the next release. The episode unfolds against a backdrop of tightening access: the US government restricted foreign access to some advanced Anthropic models in June, while Anthropic itself blocks users in China, though VPNs and proxy services circumvent the ban.

The NVDB alert landed as the UN convened a Global Dialogue on AI Governance in Geneva, where a preliminary report from its Independent International Scientific Panel on AI underscored the asymmetric concentration of computing power. The United States holds 75 per cent of the capacity of the world’s 500 most powerful AI supercomputers, with China at 15 per cent. The panel also flagged risks to information integrity, noting that AI enables large-scale production of persuasive content designed to deceive, potentially eroding the shared reality needed for democratic debate. Separately, the Future of Life Institute, a US think tank, released a safety scorecard giving Anthropic the highest overall grade—a C+—while concluding that the global industry is failing to address existential threats, including the possible misuse of models for cyberattacks.

Anthropic said the monitoring code would be rolled back in the next software release. The Geneva dialogue, which concluded on 7 July, is expected to feed into ongoing multilateral discussions on AI governance frameworks. Chinese authorities have urged organisations to strengthen network traffic monitoring to prevent unauthorised data leakage, while the UN panel’s final report is due later this year.