
中国警告Anthropic编码工具存在后门,中美AI技术摩擦升级
中国网络安全机构称Claude Code可未经同意传输敏感数据,阿里巴巴随即禁用该工具,凸显人工智能领域日益加剧的地缘政治裂痕。
中国国家信息安全漏洞库(NVDB)于7月8日发布警告,称美国人工智能公司Anthropic的编码工具Claude Code在2.1.91至2.1.196版本中存在“后门风险”,可在未经用户同意的情况下将地理位置、身份标识等敏感信息传回Anthropic服务器。该机构隶属于中国工业和信息化部,建议相关单位立即排查并卸载或升级至已移除后门代码的最新安全版本,同时加强网络流量监控。中国科技巨头阿里巴巴随后通知员工,自7月10日起禁止使用Claude Code,理由为安全顾虑。
Anthropic工程师Thariq Shihipar在社交媒体上回应称,相关代码是3月启动的一项实验,旨在防止未授权经销商滥用账户并防范“蒸馏”——即利用一个AI模型的输出训练另一个模型。他表示团队已部署更强的防护措施,并计划在次日发布中彻底移除该代码。此前,Anthropic曾公开指责阿里巴巴等中国AI实验室对其模型进行非法蒸馏,而Claude等西方AI服务在中国大陆需通过VPN等方式才能访问,官方并未批准其广泛使用。
这一事件是中美AI技术摩擦的最新注脚。6月12日,美国政府突然限制外国用户访问Anthropic部分最先进模型,将AI系统本身变为战略资产。联合国在日内瓦举行的全球AI治理对话中发布的初步报告指出,全球AI采用“加速且不对称”:美国拥有全球500台最强AI超级计算机中75%的算力,中国占15%。与此同时,美国智库未来生命研究所的安全评估显示,包括Anthropic在内的九家主要AI公司均未能在防范“存在性威胁”方面获得A级评分,Anthropic虽以C+的总分居首,却因参与美国军方在委内瑞拉和伊朗的行动而受到质疑,尽管五角大楼近期因AI安全分歧曾对其下达禁令。
联合国独立国际科学委员会的报告还警告,AI可大规模生成具有说服力的欺骗性内容,逐步侵蚀信息完整性,削弱公众信任与民主辩论。在此背景下,中国加速推进DeepSeek、智谱、豆包等本土模型,以低价和开源策略争夺市场。Anthropic承诺的代码回滚将成为短期内的观察节点,而更广泛的全球AI治理框架仍在日内瓦对话中缓慢成形。
| 拉丁美洲媒体 | −0.30 | critical |
|---|---|---|
| 俄罗斯及独联体媒体 | 0.00 | neutral |
| 东南亚媒体 | 0.00 | neutral |
| 大西洋/英语圈媒体 | 0.00 | neutral |
Latin America places China's accusation within a context of US technological hegemony and global safety failures, calling for a more balanced AI ecosystem.
The bloc universalizes the Chinese claim by linking it to UN data and global safety reports, turning a bilateral incident into a matter of global governance.
It omits that Anthropic already blocks access from China, which could undermine the credibility of the accusation.
Russia reports the Chinese discovery as a technical fact, defining the backdoor as an intentional loophole, without taking a stance.
The bloc adopts a dry, technical report, citing authoritative sources (CNNVD, WSJ) to present the news as neutral and verified.
It omits any geopolitical context or Anthropic's response, isolating the fact from the broader dispute.
Southeast Asia highlights the irony: China denounces a backdoor in a tool it already blocks, yet the tool is still accessible via VPN.
The bloc adds the VPN detail, creating an implicit tension between access restrictions and security accusations, without explicit judgment.
It omits the UN data on AI asymmetry or global safety reports, which would lend more weight to China's position.
The Atlantic reports the Chinese warning as an official statement, without adding context or skepticism, maintaining a detached observer stance.
The bloc adopts a neutral news tone, presenting the claim as a fact without questioning its validity, but also without endorsing it.
It omits that China itself has restrictions on access to Claude Code, which could make the accusation appear politically motivated.