Kimi悄悄将完整的用户提示发送到ByteDance位于中国的分析管道
我截获了一条逐字的 Kimi Work 3.2.7 提示，该提示包含在一个 kimi_work_send_message 事件中并被上传到 Volcano Engine 的 DataFinder 分析系统，事件中还包含用户、会话和设备标识符。该公司针对 analytics SDK 的具体披露列出了设备、网络、位置和应用程序信息，但并未明确列出会话文本。
By Ryan Merket · Published
RUNTIMEWIRE INVESTIGATION — Exclusive
Original reporting by RuntimeWire, based on documents, testing, reverse engineering.
为何重要
AI 提示可能包含公司的核心机密：源代码、凭证、合同、客户数据和未发布的计划。将这些文本复制到分析工具中会把一次对话变成额外的行为数据记录，从而扩大围绕那些用户可能以为只是为回答他们而被处理的资料的隐私和安全暴露。
Reporting record
Finding
RuntimeWire established that Kimi Work 3.2.7 copied a complete user prompt into a kimi_work_send_message product-analytics event sent to gator.volces.com.
How we verified
Methods: documents, testing, reverse engineering.
A controlled HTTPS capture showed the synthetic prompt verbatim in the query property of a kimi_work_send_message event; static analysis of the Kimi Work 3.2.7 app.asar traced input.query to that event and the gator.volces.com analytics endpoint; Volcano Engine documentation identifies the endpoint as a DataFinder reporting address.
RuntimeWire unpacked and analyzed the Kimi Work 3.2.7 app.asar archive, traced its analytics code, routed the Windows client through an HTTPS inspection proxy, entered a unique synthetic prompt in a fresh conversation, and inspected 213 outbound HTTPS requests, including the captured analytics payload.
Tested versions:
- Kimi Work 3.2.7
Reproduction
RuntimeWire independently reproduced the core finding.
Use the tested Kimi Work 3.2.7 application package, recorded by SHA-256 as c0f672d96c12d4a5028913c426ce847523ee78bfb78e8c39321f355ce590c4f4. Route the Windows application through an HTTPS inspection proxy, enter the synthetic prompt containing RUNTIMEWIRE-KIMI-PRIVACY-35352689 in a fresh conversation, and inspect outbound traffic for a POST to https://gator.volces.com/list containing a kimi_work_send_message event with the prompt in its query property.
Company response
RuntimeWire requested comment; the company had not responded by publication time.
本文为我为 RuntimeWire 的原创报道，基于逆向工程、受控网络拦截和一手文档。
我一直把 Kimi 当作日常主要使用的工具。
本周情况改变了。
一天之前， Anthropic 披露 Moonshot AI（Kimi 背后的公司）曾秘密将部分 Kimi 客户请求通过 Claude 转发。Anthropic 表示在一个十天的期间内观测到近 300,000 条客户请求被中转，其中包括包含监控材料、公司内部源代码和实时凭证的请求。Anthropic 称 Moonshot 至少保存了这些交流中的部分，并将其用于链式思考提取管道（chain-of-thought extraction pipeline）。
在阅读那份报告后，我拆解了最新的 Kimi Windows 客户端，并提出了一个更狭窄的问题：
当我在 Kimi Work 中输入内容并按下发送时，这段文本还会被发送到哪里？
我发现了另一个目的地。
在对 Kimi Work 3.2.7 的受控测试中，我在一个新的对话里输入了一条唯一的合成消息：
我将该 Windows 应用通过运行在我自己测试机器上的 HTTPS 检查代理进行转发。…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.