nshot AI）有关联的人员策划了一场协调一致的行动，试图提取其模型的隐藏推理过程。月之暗面是中国Kimi模型的开发公司。OpenIO于周三通过一篇博客文章公布了调查结论。
Credit: Bangla press via Shutterstock.com OpenAI says people linked to Moonshot AI, the Chinese developer of Kimi, were behind a coordinated campaign to extract the hidden reasoning of its models. The company set out its findings in a blog post on Wednesday.
该活动于7月1日开始，起初请求量较低。7月24日和25日，请求量激增，共有超过4,000名用户发起16,000次请求。OpenAI随后发现，超过15,000名用户存在相关活动。公司称，截至7月28日，已全面终止这场活动。OpenAI无法将每名操作者都追溯到同一个主体，但将“核心团伙”认定为与月之暗面有关联的人员。
The activity began on 1 July at low volume. It spiked on 24 and 25 July, with 16,000 requests from more than 4,000 users. OpenAI then found related activity across more than 15,000 users. It says it fully shut the campaign down by 28 July. It could not tie every operator to one actor, but it attributes “a core cluster” to people associated with Moonshot.
OpenAI写道：“这种活动符合对抗性蒸馏的特征，即系统性地、未经授权地利用一个模型的输出或推理过程，帮助训练、复现或改进另一个模型。”
“This activity is consistent with adversarial distillation: the systematic and unauthorized use of one model’s outputs or reasoning to help train, reproduce, or improve another model,” OpenAI wrote.
提取活动的运作方式OpenAI会隐藏模型在作答前进行的推理过程。公司表示，操作人员从一段对话中复制了经过加密的推理内容，随后在另一段对话中要求模型对其进行解密并转录。
How the extraction worked OpenAI hides the reasoning its models go through before they answer. The operators copied that encrypted reasoning from one conversation, the company said. They then asked a model in another conversation to decrypt and transcribe it.
OpenAI补充称，这些人员并未破解加密系统、侵入数据库或获取已存储的用户对话。此后，OpenAI封禁了相关账户，收紧了注册验证措施，并为隐藏推理过程增加了防护。OpenAI还通过前沿模型论坛（Frontier Model Forum）以及政府渠道，将调查结果与其他人工智能实验室共享。
They did not break its encryption, get into a database or reach stored user conversations, OpenAI added. It has since banned accounts, tightened sign-up checks and added protections for hidden reasoning. It also shared its findings with other AI labs through the Frontier Model Forum and with government channels.
OpenAI负责战略国家安全政策项目的卡罗琳·齐尔（Caroline Zier）告诉彭博社：“我们关注的是违反服务条款的行为，而不是开放模型或合法的蒸馏。”
“Our concern is about violation of our terms of service, not open models or legitimate distillation,” Caroline Zier, who leads strategic national security policy initiatives at OpenAI, told Bloomberg.
指控与反驳据彭博社报道，这是OpenAI首次指控月之暗面。彭博社指出，Anthropic在上个月发布的一份报告中，也对该公司提出了类似指控。
Accusations and pushback This is the first time OpenAI has accused Moonshot, Bloomberg reported. Anthropic made a similar claim against the company in a report last month, Bloomberg noted.
美国政府机构也提出了同样指控。本月发布的一份通告点名了六家中国企业，以及每家企业试图提取的美国模型。中国驳回了这份通告，称相关指控毫无根据。
US agencies have made the same charge. An advisory this month named six Chinese firms and the US models each one targeted. China rejected the advisory as unfounded accusations.
并非所有人都相信这些说法。据彭博社报道，特朗普政府前人工智能负责人大卫·萨克斯（David Sacks）称，这类报道其实是为了迫使美国禁止使用竞争对手开发的开源人工智能模型。
Not everyone accepts the claims. David Sacks, Trump’s former AI czar, has called such reports a push to get the US to ban rival open models, Bloomberg reported.
“Moonshot”项目在国内也面临着监管机构的审查。据《The Information》本月报道，中国监管机构正在对“DeepSeek”和“Moonshot”这两个项目进行调查。
Moonshot also faces scrutiny at home. Chinese regulators are investigating DeepSeek and Moonshot, The Information reported this month.