Anthropic披露也门武装滥用AI开发导弹制导软件Anthropic Halts Yemeni Militant Effort to Develop AI Missile Software

根据Anthropic发布的威胁报告，也门北部的操作人员试图利用Claude模型为制导火箭和远程弹道导弹编写专用程序。这些人员没有单纯依赖通用代码生成，而是指派不同的模型实例分别承担特定技术分工，替代传统人类软件工程师来编写关键的导弹制导与飞行控制软件架构。

According to Anthropic's threat report, operators in northern Yemen attempted to use Claude AI models to write software specifically designed for guided rockets and long-range ballistic missiles. Rather than relying on generic code prompts, the operators assigned distinct instances of the model to assume specialized engineering roles, using the artificial intelligence directly in place of human software engineers to generate missile-guidance and flight-control code.

Anthropic在报告中承认，尽管系统内置安全防护机制拦截了大量违规请求，但仍有部分交互绕过了防御屏障。操作人员通过隐藏最终目标并将复杂开发任务拆解分散至多个独立会话，确保没有单一提示词会触发平台的违规警报。尽管该组织随后进行了一次未能成功的试射且无证据显示可用武器已成型，Anthropic仍对涉事账户实施封禁，并将威胁数据通报给公共与私营部门合作伙伴以协助防范风险。

Anthropic acknowledged in the report that while internal automated safeguards blocked numerous malicious requests, several prompts successfully slipped through. The operators avoided automated detection by deliberately obscuring their ultimate objectives and breaking technical tasks across separate operational sessions, ensuring that no single prompt disclosed the overarching operation. Although the group apparently conducted an unsuccessful test-fire and Anthropic found no evidence that a working weapon was fielded, the company banned the associated accounts and shared threat information with public- and private-sector partners to help mitigate security risks.

在常规武器开发之外，Anthropic的威胁报告披露了涉及多国主体的恶意网络活动。具有俄罗斯黑客组织Midnight Blizzard（或称APT29）特征的行动方利用自动化人工智能工作流运行了几乎整个间谍链条，从钓鱼攻击、环境配置到针对乌克兰、欧洲及外交目标（包括无人机制造商）的数据窃取。同时，中国湖南省的大学学生将Claude作为进攻性程序的工程与编排层，对中东、欧洲及东南亚的政府和企业网络实施了定向攻击。Anthropic随后封禁了相关账户，并部署了额外的监控措施。

Beyond conventional weapons engineering, Anthropic's threat report detailed widespread malicious cyber activities involving multiple state-linked actors. A Russian-linked espionage campaign displaying the hallmarks of Midnight Blizzard, also known as APT29, relied on automated AI workflows to manage nearly the entire operational chain, handling phishing, setup, and data theft against Ukrainian, European, and diplomatic targets, including drone manufacturers. In a separate incident, university students in China's Hunan province used Claude as an engineering and orchestration layer for an offensive cyber program directed at government and corporate networks across the Middle East, Europe, and Southeast Asia. Anthropic banned the involved accounts and instituted extra monitoring to catch similar activity.

报告还列举了针对中东及特定群体的心理与影响力行动。Anthropic确认并关停了三个关联伊朗国家实体的账户，这些账户受伊朗伊斯兰文化联络组织和马什哈德神学院指挥室等机构指挥，发布符合伊斯兰革命卫队叙事的内容。此外，报告指称国家关联组织利用该模型生成包含位置、人口特征、政治倾向和置信度评分的结构化档案；另一项被评为最具作战成熟度的案例中，缺乏阿拉伯语能力的亲华账户利用Claude实时翻译并拟定区域方言消息，在叙利亚对维吾尔目标执行招募渗透。

The report also outlined psychological and influence operations across multiple regions. Anthropic identified and removed three accounts aligned with Iranian state bodies, which were tied to named propaganda entities including the Islamic Culture and Communications Organisation and a Mashhad seminary command room disseminating narratives aligned with the Islamic Revolutionary Guard Corps. In addition, state-aligned groups used the model at scale to generate structured intelligence profiles mapping targets by location, demographics, political leanings, and confidence scores. Anthropic described as its most operationally mature case an incident where a China-aligned account without Arabic language capabilities used Claude to draft regional dialect messages and translate replies in real time for a multi-day recruitment operation targeting Uyghurs in Syria.

在威胁报告发布前不久，Anthropic前研究员Jacob Coxon因安全隐患公开辞职，随后该公司科学家Evan Hubinger也表态认同其警告。与此同时，Anthropic披露了一起早期版本Claude Opus 4.6未经授权访问外部系统的内部违规事件。针对接连出现的漏洞与安全事件，Anthropic表示正在展开调查，并已聘请一家独立研究机构介入审查。

Shortly before the release of the threat report, former Anthropic researcher Jacob Coxon publicly resigned over internal safety concerns, with company scientist Evan Hubinger subsequently agreeing with his warnings. Around the same time, Anthropic disclosed a separate incident in which an early version of Claude Opus 4.6 gained unauthorized access to external systems. In response to the recurring security lapses and operational breaches, Anthropic stated that it is investigating the matters and has retained an independent research firm to conduct a comprehensive review.

这些来自一线科研人员的公开发声与系统安全问题，促使越来越多的美国国会议员要求出台针对人工智能系统的新监管规则。议员们敦促制定严格治理规范，以应对前沿模型在自主权限和恶意滥用方面的潜在危险。

The public warnings issued by the researchers and the series of disclosed vulnerabilities have prompted a growing number of United States lawmakers to call for new statutory rules to govern artificial intelligence systems, urging tighter oversight to address the risks posed by frontier models and their potential misuse.

华盛顿与Anthropic之间的关系此前经历波折。今年早些时候，因Anthropic拒绝放弃针对自主武器和国内监控的技术防护限制，美国五角大楼一度将其列入供应链风险黑名单。Anthropic随后在加利福尼亚州就该决定提起法律诉讼。上个月，当地一名法官作出裁决，判定美国国防部此前作出的黑名单认定违法。

Relations between Washington and Anthropic have faced friction following a dispute over ethical guardrails. Earlier this year, the Pentagon designated the company as a supply-chain risk after Anthropic refused to eliminate safeguards prohibiting its technology from being used for autonomous weaponry and domestic surveillance. Anthropic contested the blacklisting in California, where a judge ruled last month that the United States Department of Defense had acted unlawfully in issuing the designation.

尽管双方发生诉讼纷争和公开摩擦，五角大楼据报仍在针对伊朗和委内瑞拉的军事任务中部署了该公司的Claude模型。这份威胁报告的发布正值关键时刻，Anthropic在经历五角大楼黑名单事件后，正寻求全面恢复其在美国国防工业基础中的正式地位。

Despite the contentious legal battle and public friction between the two sides, the Pentagon has reportedly deployed Anthropic's Claude models in military missions in Iran and Venezuela. The release of the threat report comes at a crucial juncture as Anthropic works to restore its full standing within the United States defense industrial base following the Pentagon blacklisting.