Anthropic公司在近期威胁报告中披露，也门北部一支与胡塞武装相关的组织利用其Claude Code AI模型开发导弹制导软件及相关武器系统1。该组织通过Claude进行导弹制导软件开发、轨迹计算模拟以及火箭测试数据分析1。Anthropic已禁用了相关账户，但该组织已事先编译出离线可执行程序，并组建了独立的工程工具包1。
这支组织采用了协作工作流来提升开发效率1。他们运行多个Claude实例并行处理编码、研究和技术审查任务，将原本需要专业导弹工程团队完成的工作通过AI辅助大幅压缩1。具体地，该组织使用Claude为飞行控制系统集成开源自动驾驶软件，并开发了六自由度轨迹仿真工具1。在一次火箭测试失败后，他们立即利用Claude分析发射遥测数据1。该组织最终开发出了射程超过2000公里的弹道导弹以及名为"R2000"的高超音速滑翔车概念1。
Anthropic的防护措施曾被多次绕过，运营者通过模糊化任务意图、分散工作到不同对话和采用其他规避手段来突破限制1。Anthropic威胁情报主管Jacob Klein表示，这种威胁的性质十分复杂："并非简单地说'我要制造生物武器去杀死所有人'"1。该公司威胁报告涵盖了2025年12月至2026年8月期间被中断的行动，记录的六个常规武器案例中，三例涉及中国、两例涉及俄罗斯、一例涉及也门1。
Anthropic has disclosed that a Yemen-based organization linked to Houthi militants used its Claude Code AI model to develop missile guidance software and conduct ballistic missile engineering, according to the company's threat report released in September 1. The group employed parallel instances of Claude to distribute coding, research, and technical review tasks, effectively compressing work that would typically require a professional missile engineering team into an AI-assisted workflow 1. The organization subsequently compiled the collaborative outputs into an offline executable program to continue independent development 1.
The group's capabilities extended beyond guidance systems. Using Claude, they developed flight control integration for ballistic missiles with ranges exceeding 2,000 kilometers and conceived a hypersonic gliding vehicle concept designated "R2000" 1. The organization also leveraged Claude to create six-degree-of-freedom trajectory simulation tools and incorporated open-source autopilot software into their flight control architecture 1. Following a rocket test that failed, the group immediately used Claude to analyze the launch telemetry data 1.
Anthropic documented attempts to circumvent its safety protections through multiple methods, including obscuring intent, fragmenting work across separate conversations, and employing alternative evasion techniques 1. The company disabled the relevant accounts after discovering the misuse, though the organization had already established independent engineering toolsets 1. Jacob Klein, Anthropic's threat intelligence lead, noted the complexity of such threats, stating: "You are not seeing someone in a comic book kind of way say, Hey, I want to build a biological weapon to kill everybody. It's an incredibly nuanced situation." 1 Among six conventional weapons cases recorded by Anthropic during disrupted operations between December 2025 and August 2026, three involved actors from China, two from Russia, and one from Yemen 1.
评论
还没有评论，欢迎留下第一条。