刘宁Ning Liu
← 文章← Articles

供应商说「资料安全看你用哪家模型」,只说对了一部分A Vendor Says Data Security Depends on Which Model You Use. That Is Only Partly Right

2026-10-23

最近有人问我一个问题:公司要对接一款 AI 产品,问供应商核心资料怎么保护,对方说,这取决于你们用的是哪家的模型;如果不放心,也可以自己买服务器来跑模型。

这个回答只说对了一部分。

模型那一层,确实要看用哪家、签的什么条款

资料最后交给哪家的模型处理、按什么条款处理、会不会被拿去训练,确实取决于公司用的是哪家、开的是什么账号。之前那篇《公司的资料,到底能不能发给 AI?》专门讲过怎么看这些条款。

被跳过的,是供应商自己的那一段

资料从公司的系统到模型,中间还要经过供应商的产品。这一段发生的事,跟用哪家模型关系不大:

所以,就算模型那一层的条款写得很清楚,资料也可能在中间这一段被多读一些、多存一份。这部分应该由供应商自己回答,不能用「取决于模型」带过去。

「可以装在你们本地」,要问清装的是什么

不少产品会说,可以部署在客户自己的服务器上。这时要多问一句:装在本地的是连接服务,还是模型本身?

如果只有连接服务在本地,模型还在云上,资料照样要发到云上的模型去处理。

自己买服务器跑模型,解决的是另一件事

自己部署模型,解决的是「资料不交给外部的模型公司」这一件事。之前那篇也算过这笔账:投入不小,要有人维护,效果也不一定比云端的好。

这里再补两点。一是自己部署之后,服务器的权限、补丁、日志都要自己管,安全责任跟着转到了公司自己身上,并不会因为在本地就自动更安全。二是供应商产品那一段的问题,不会因为换了模型就消失:它照样要读公司的系统,照样可能在自己那里存一份。

对接之前,关于产品或服务,向供应商问清楚如下五个问题:

  1. 它会读哪些资料?权限能不能按项目或目录来限制?
  2. 资料会不会经过它自己的服务器?存不存、存多久、谁能看到?它自己会不会拿这些资料去改进产品?
  3. 最后交给哪家的模型、按谁的合同处理?能不能换成公司自己的模型账号?
  4. 它能不能修改系统里的内容?修改之前,有没有人工确认这一步?
  5. 合作结束后,留在它那里的资料怎么删?

这五个问题,问的都是资料从进来到离开,经过了谁的手、在谁那里留了多久。个人认为,能答清楚、也愿意把答案写进合同的供应商,比只说「取决于模型」的更靠谱。

Someone recently put a question to me. Their company was about to connect an AI product, and when they asked the vendor how their core data would be protected, the vendor said it depended on which company's model they used; and if they were still worried, they could buy their own servers and run the model themselves.

That answer is only partly right.

At the model layer, it does depend on whose model and which terms

Which company's model ends up processing the data, under what terms, and whether the data may be used for training do depend on which provider the company uses and what kind of account it has. An earlier article, Can company material go into AI?, looked at how to read those terms.

What gets skipped is the vendor's own part

On its way from the company's systems to the model, the data passes through the vendor's product. What happens in that stretch has little to do with which model is used:

So even when the terms at the model layer are perfectly clear, the data may still be read more widely, or copied once more, in this middle stretch. This part is for the vendor to answer; it cannot be brushed aside with "it depends on the model".

"It can be installed on your premises": ask what exactly is installed

Many products say they can be deployed on the customer's own servers. That is the moment to ask one more question: is what runs locally the connector service, or the model itself?

If only the connector runs locally and the model is still in the cloud, the data still has to be sent to the cloud model to be processed.

Running your own model solves a different problem

Deploying a model yourself solves one thing: your data is not handed to an outside model company. The earlier article also did the sums on this: the investment is considerable, someone has to maintain the setup, and the results are not necessarily better than in the cloud.

Two more points. First, once you run it yourself, you have to manage the servers' access rights, patches and logs; the security responsibility moves to the company itself, and being on premises does not make it safer automatically. Second, the issues in the vendor's part do not go away when the model changes: the product still has to read the company's systems, and may still keep a copy on its side.

Before connecting, ask the vendor these five questions about its product or service:

  1. Which data will it read? Can its access be limited by project or folder?
  2. Will the data pass through the vendor's own servers? Is it stored, for how long, and who can see it? Will the vendor use the data to improve its own product?
  3. Which company's model does the data end up with, and under whose contract? Can the company's own model account be used instead?
  4. Can it change what is in the system? Is there a step where a person confirms a change before it is made?
  5. When the relationship ends, how is the data left with the vendor deleted?

All five questions ask the same thing: from the moment the data comes in to the moment it leaves, whose hands it passes through and how long it stays with whom. In my view, a vendor that can answer them clearly, and is willing to put the answers in the contract, is more reliable than one that only says "it depends on the model".

刘宁 Ning Liu|帮传统企业把 AI 跑通的人。工程师出身:浙大计算机本科、中科大硕士,在微软、MicroStrategy 做过工程师,2018 年起带队做企业 AI 落地,主导过集团级 AI 中台。技术之外,还有两段经历:长江商学院 MBA(含密歇根 Ross 海外交换),能把技术和公司战略对齐;创过业,又在教育、地产等传统行业做了 8 年 CTO,从技术配合业务,做到技术带着业务走。现为企业提供 AI 落地诊断、战略规划、项目落地 Sprint(带企业团队一起交付)与 Fractional AI Lead(长期在位的 AI 负责人)服务。聊聊你的 AI 落地 →
Ning Liu — Your Fractional CTO for AI. An engineer by background (Zhejiang University, USTC; Microsoft, MicroStrategy), group AI-platform lead, hands-on with enterprise AI since 2018. Beyond the engineering, two things: an MBA from CKGSB (with an exchange semester at Michigan Ross), so I can align technology with company strategy; and I founded a company, then spent 8 years as a CTO in traditional industries such as education and real estate — moving from tech that supports the business to tech that leads it. I offer AI Audit, Strategy, Delivery Sprint and Fractional AI Lead. Let's talk about your AI →