美国众议院美中战略竞争特别委员会的民主党领衔议员正要求美国几大领先的人工智能实验室提供信息，以了解北京涉嫌从其公司窃取模型权重的行为。
The leading Democrat on the House’s China Select Committee is calling on the U.S.’s leading AI labs for information on Beijing’s alleged efforts to steal model weights from their firms.
众议院美国共产党战略竞争特别委员会排名第一的民主党众议员罗·卡纳（Ro Khanna，加利福尼亚州民主党人）于周四向 OpenAI、Anthropic、Meta 以及谷歌母公司 Alphabet 索取数据，并表达了对中国共产党窃取行为可能“轻而易举地侵蚀美国在人工智能领域领先地位”的担忧。模型权重是人工智能系统内部存储机器学习知识并决定其输出结果的数值参数。
Rep. Ro Khanna (D-Calif.), the ranking member of the House Select Committee of the Chinese Communist Party, requested data on Thursday from OpenAI, Anthropic, Meta and Alphabet, the parent company of Google, expressing concerns the CCP’s theft could “erode America’s AI lead with the stroke of a keyboard.”Model weights are the numerical parameters inside AI systems that store the learning machine’s knowledge and determine its outputs.
卡纳在《国会山报》获得的一封信中写道：“如果中国通过国家资助的黑客手段从美国领先的前沿实验室窃取专有模型权重，中国共产党将瞬间绕过多年昂贵的研发过程，直接获得美国政府长期以来一直试图对华封锁的高级功能。”
“If the PRC were to steal a proprietary model weight from a leading U.S. frontier lab through state-sponsored hacking, the CCP would instantly bypass years of expensive research and obtain immediate access to advanced capabilities that U.S. Government policy has long sought to deny the PRC,” Khanna wrote in letters obtained by The Hill.
今年早些时候，白宫科学和政策办公室主任迈克尔·克拉齐奥斯（Michael Kratsios）在一份备忘录中写道，美国政府了解到中国实体正在利用数千个代理账户和越狱策略来获取专有信息。
Earlier this year, Michael Kratsios, the director of the White House Office of Science and Technology, wrote in a memo that the U.S. government has learned Chinese entities are leveraging thousands of proxy accounts and jailbreaking tactics to access proprietary information.
这一过程被称为蒸馏，在此期间，较强系统的输出被用于训练能力较弱的系统。它们虽然无法达到原始系统的全部性能，但却能让外国行为者以更低廉的价格发布看似具备可比性的产品。
This process is referred to as distillation, during which the outputs of a stronger system are used to train less capable ones. They do not meet the full performance capabilities of the original but allow foreign actors to release products that appear comparable for a cheaper price.
在激烈的人工智能竞赛中，美国一直在努力限制美国技术向中国转移。去年年底，特朗普总统因决定允许英伟达向中国出售其 H200 芯片而遭到本党同仁的猛烈抨击。卡纳的信件最早由路透社报道。
The U.S. has sought to limit the transfer of U.S. technology to China amid the tight AI race. President Trump came under backlash from his own party late last year over his decision to allow Nvidia to sell its H200 chips to China. Khanna’s letters were first reported by Reuters.
加州民主党议员称，即便中国方面不窃取完整的模型权重，也可能从美国企业的系统中获取“价值［知识产权］”，并接触敏感研究。
The California Democrat argued that, even if Chinese actors do not steal a full model weight, they could obtain “value [intellectual property]” from the American companies’ systems and access sensitive research.
他要求这些企业提交一份清单，列明过去两年中国或其他敌对行为体试图获取模型权重和知识产权的所有已知案例，以及企业为加强网络安全所做的努力。
He asked the companies to hand over a list of all known instances of China or another hostile actor trying to obtain model weights and IP over the past two years, along with the firms’ efforts to boost cybersecurity.
“正如你们的企业有责任防范模型被用于潜在危险用途，并应对人工智能对人类和社会造成的影响一样，你们也有责任确保公司发生的网络安全事件不会危及美国国家安全，”卡纳写道。这些企业和中国驻华盛顿大使馆均未立即回应置评请求。
“Just as your companies have a responsibility to protect against potentially dangerous uses of your models and the impacts of AI on humanity and society, you have a responsibility to ensure a cybersecurity incident at your company does not endanger American national security,” Khanna wrote. Neither the companies nor the Chinese Embassy in Washington immediately responded to request for comment.