美国安全机构指控 DeepSeek 与月之暗面等中国人工智慧企业，透过名为「模型蒸馏」（distillation）的技术，以工业级规模汲取美国顶尖 AI 模型的输出资讯来训练自身系统。美国官方与 Anthropic、OpenAI 等科技公司认为此未授权蒸馏行为损害了商业利益并带来安全风险，而中国商务部则驳斥此类指控毫无根据，并警告若美方以此为借口进行打压将采取反制措施。

蒸馏技术本质上是利用能力较强的「教师模型」输出作为训练资料，来训练规模较小、运算成本较低的「学生模型」。虽然企业内部进行授权蒸馏可达到模型压缩并提升效率的目的，但第三方未经许可的复制行为则引发智慧财产权与商业竞争争议。美国企业担忧投入数千亿美元研发的闭源系统被低成本复制甚至开源，进而造成巨额营收损失，同时也可能因防护机制被移除而衍生资安威胁。

针对此类未授权蒸馏攻击，Anthropic 等公司透过追踪 IP 位址及中继资料，发现了数以千万计利用虚假帐号进行的互动纪录。面对矽谷业者的诉求，美国政府与国会正研议防范措施及制裁机制，美方企业亦加强联防并封锁可疑帐号；然而因蒸馏行为往往依赖间接证据，技术取证与全面防范在实务上仍具相当挑战性。

US security agencies have accused Chinese AI firms, such as DeepSeek and Moonshot AI, of improperly leveraging advanced American models at an industrial scale through a technique called distillation to train their own systems. While US officials and companies like Anthropic and OpenAI argue that unauthorized distillation jeopardizes massive investments and national security, China's Ministry of Commerce dismissed the allegations as groundless and warned of countermeasures against potential US suppression.

Distillation involves using outputs from a larger teacher model to train a smaller student model, replicating key capabilities at a fraction of the original computational cost. Although authorized distillation is a standard industry method for optimizing internal models, unauthorized third-party distillation threatens proprietary business models. US companies argue that low-cost or open-weight replicas undercut their ability to recoup infrastructure investments while also risking the removal of vital safety guardrails.

To detect and counter distillation attacks, companies like Anthropic have tracked millions of anomalous queries originating from fraudulent accounts using IP and metadata analysis. As Silicon Valley urges action, US lawmakers and the administration are considering policy interventions, including potential sanctions and industry-wide defense coordination, though definitively proving and preventing distillation remains technically challenging.