Anthropic发布了一份涵盖2025年12月至2026年8月的威胁情报报告，详细记录了该公司识别并中断的Claude AI模型滥用行为1。报告涉及七个主要危害领域：网络运营、影响力操作、监视、诈骗欺诈、生物武器开发、常规武器开发和模型蒸馏，涉及国家支持的黑客、财务动机犯罪分子、商业间谍软件供应商和政治激进分子等多类威胁行为者1。
在网络运营方面，俄罗斯间谍组织GTG-20006在20多个组织中进行目标活动，包括乌克兰政府和北非政府机构；ShinyHunters从一家技术供应商盗取超过1TB数据，从航空公司获取数千万乘客记录；中国行为者GTG-10007针对全球约50个组织进行间谍活动，覆盖政府机构和企业1。影响力操作案例涉及俄罗斯、伊朗、土耳其、海湾国家、南亚、非洲和欧洲，共9个案例针对六大洲受众进行虚假信息传播和政治操纵1。
生物武器相关滥用构成严重威胁1。Anthropic已阻止了可能用于生物武器开发的AI使用尝试2，涉及5个案例包括病原体研究、基因编辑、流感适应性研究和毒素合成1。此外，监视操作方面，中国、伊朗和西非国家利用Claude建设大规模监视系统，包括面部识别、社交媒体分析和异议人士追踪1。常规武器开发涉及6个案例，包括导弹制导软件、反鱼雷系统、无人机群和电子战系统，其中也门行为者进行了制导火箭实地测试1。
在模型蒸馏方面，阿里巴巴、Moonshot、DeepSeek、Zhipu和Xiaomi等中国AI实验室通过欺诈账户进行大规模非法蒸馏，其中阿里巴巴单日峰值约300万次交换1。欺诈活动中，GTG-15001操作超过20个虚假约会应用，部署4700个AI虚拟人物与至少25000个独特个人对话进行诈骗1。
Anthropic已禁止了所有相关账户，与执法部门和行业合作伙伴共享信息，并强化了安全防护措施1。该公司同时成功中断了俄罗斯和中国针对Claude模型的AI活动6。
Anthropic has released a comprehensive threat intelligence report documenting AI abuse incidents identified and disrupted between December 2025 and August 2026, spanning seven distinct harm categories: cyber operations, influence operations, surveillance, fraud, biological weapon development, conventional weapon development, and model distillation.1
The report details multiple threat actors—including state-sponsored hackers, financially motivated criminals, commercial spyware vendors, and political extremists—leveraging Claude for malicious purposes.1 In cyber operations, state-backed groups conducted targeted campaigns globally; Russian-aligned GTG-20006 executed operations across more than 20 organizations including Ukrainian government entities and North African institutions, while a Chinese actor designated GTG-10007 targeted approximately 50 organizations worldwide across government and corporate sectors.1 The group known as ShinyHunters (GTG-50014) stole over one terabyte of data from a technology vendor and obtained tens of millions of passenger records from an airline.1
Influence operations involved nine documented cases across Russia, Iran, Turkey, Gulf states, South Asia, Africa, and Europe, targeting audiences on six continents with disinformation and political manipulation.1 Surveillance operations in China, Iran, and West African nations employed Claude to build large-scale systems incorporating facial recognition, social media analysis, and dissident tracking.1 Five biological abuse cases involved pathogen research and bioweapon-related activities including gene editing, influenza adaptation research, and toxin synthesis.1 Conventional weapons development spanned six cases encompassing missile guidance software, anti-torpedo systems, drone swarms, and electronic warfare systems, with Yemen-based actors conducting field tests of guided rockets.1
Model distillation emerged as a significant abuse vector, with Chinese AI laboratories—Alibaba, Moonshot, DeepSeek, Zhipu, and Xiaomi—conducting large-scale illicit extraction through fraudulent accounts, with Alibaba reaching approximately 3 million exchanges at peak daily usage.1 One fraud operation, GTG-15001, deployed over 4,700 AI virtual personas across more than 20 fake dating applications to conduct romance scams involving at least 25,000 unique individuals.1
Anthropic has prohibited all associated accounts and coordinated information sharing with law enforcement and industry partners while strengthening security protections.1 The company has also disrupted AI campaigns targeting Claude originating from Russian and Chinese actors.6
评论
还没有评论，欢迎留下第一条。