大家好，我是：
- 农村程序员（2020年回农村，至今未出村）
- 独立开发者（致力于开发 1000 个小产品）
- 技术观察者（分享编程技术，行业资讯等）
- 前端之虎陈随易（自封的）
- 我的官网是：https://yicode.tech
Postman 又翻车了，这次不是因为涨价，是因为它动了别人家的东西。
10 月 24 日傍晚，一位开发者打开电脑上的 Postman，看了几分钟就关了。
第二天他发现，自己常用的两个 AI 编程助手不太对劲。
一个老报登录失败，另一个接到写接口的活儿，会先去翻 Postman 的使用说明。
他查了大半天，最后在 Postman 自己的日志里找到了答案。
日志写着安装成功，装进去的是给 Claude Code 和 Codex 用的插件。
问题在于，他从来没点过安装。
一份写着已安装的日志
他用的是 Postman 12.33.4。
日志里那行触发条件，直译过来就是一条专门给老用户的升级流程。
也就是说，只要你是登录状态的老用户，打开一次，它顺手就把事办了。
留下来的东西有三样。
第一样在他的 Codex 配置里，多了一个插件市场，一段每次启动自动读的指令，还顺手把这条指令标成了可信。
第二样在 Claude Code 的设置里，Postman 的插件从那一刻起就是打开状态。
第三样在他的家目录，多了一个 postman 命令，还留了一条指向官方的远程连接，只是连不上。
那句启动指令的内容，大意是遇到接口相关的工作，先加载 Postman 的规范。
听着不重，落点却很深。
它等于在别人家的工具里写了一句话，以后这类活儿默认找我，而且不用你点头。
十天没人应，一夜全网知
9 月 11 日，Postman 上线了这个插件仓库。
10 月 24 日，第一次有人在日志里抓到自动安装。
10 月 26 日凌晨，这位开发者写了一份很长的报告，标题里带着 “严重” 两个字。
报告里什么都有，日志原文、配置截图、安装器的源码链接，六条诉求一条条列清楚。
然后就没有然后了。
整整八天，仓库里没有一句官方回复，报告也没被打上任何标签。
11 月 3 日清晨，第一条评论才出现，接下来一天涌进五十条。
11 月 4 日凌晨，有人把它搬上了 Hacker News。
到我去核对的时候，那条帖子一千四百多分，四百六十多条评论。
当天白天，中文圈也跟上了，一条汇总帖几个小时就有十几万阅读。
骂声里有三种人
第一种最多，是收拾东西走人的。
他们的话很朴素，打开一个接口调试工具，结果它改了我的编程助手。
这些帖子底下最常出现三个名字，Bruno、Insomnia 和 Milkman，都是同类工具，能免费离线用。
第二种人更冷静，指向也更准。
他们说真正的漏洞不在 Postman，而在于我们给 AI 的信任只是一行配置，谁都能写。
Postman 只是碰巧当了那个写它的软件。
第三种人很少，但很关键。
他们提醒，那个传得到处都是的标题说的是 “试图共享全部 API 数据”，这一句站不住。
这不是突然发生的
Postman 的口碑已经跌了好几年。
去年 10 月，它砍掉免费版的小团队协作，一个团队只能留一个人，当时就有一轮凭什么还用它。
今年 6 月，它自己的安全页面承认，有人借一个第三方服务的账号，从它的销售系统里拿走了客户名单和报价。
那次是服务器被别人攻破，这次是自己动手。
同类的先例也有，GitKraken 在今年 4 月因为差不多的动作公开道过歉，官方博客里写的是，他们增强了 AI 工作流，但漏了关键步骤。
那句共享数据，站不住脚
把现有材料摊开，其实只有三层。
第一层是坐实的。
老用户打开一次就会被装插件，两个助手的配置会被改，启动指令会被标成可信。
两位互不相识的人，在两个版本上都复现了。
第二层是没说清的。
那条指令到底有没有真的跑起来，两个人说法不一致，一个说有记录，一个说没有。
第三层是没有证据的。
上传数据这件事，没有任何一份日志能支撑。
材料只到留了一条远程连接，而且连不上这一步。
写报告的人在文末专门声明，他不主张数据外泄。
想确认自己有没有中招
三条命令就够了。
grep -i postman ~/.codex/config.toml
grep -i postman ~/.claude/settings.json
ls ~/.local/bin/postman ~/.copilot/settings.json
要清理，客户端菜单里有卸载入口，在帮助下面的 “Postman CLI and Plugins” 里。
不过已经有人反馈，Copilot 那边卸不干净，配置里还留着。
截至 11 月 4 日，Postman 的官方账号还在发性能测试的内容，没就这件事表态。
写报告的人只提了一个要求，把它当成产品边界的事故来处理，别当成文档问题。
说到底，这件事让人不舒服的地方，不是一家公司多装了一个插件。
是我们早就默认，把能改配置的权限，交给了每一个愿意伸手的软件。
你机器上有没有中招，欢迎在评论区说说。
issue 地址：https://github.com/postmanlabs/postman-app-support/issues/14370。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Hello, I'm:
- Rural programmer (back to the countryside in 2020, not yet out of the village)
- Independent developers.
- Technical observers (sharing of programming techniques, industry information, etc.)
- The Tiger of the Front.
- My network is: https://yicode.tech
Postman flipped over again, not because the price went up, because it moved other people's things.
On the evening of October 24th, a developer opened Postman's computer and closed it for a few minutes.
The next day he found out that there was something wrong with his two AI programming assistants.
One of the old newspapers failed to log in, and the other received a writing interface, who went to see Postman's instructions first.
He searched for half a day and finally found the answer in Postman's own journal.
The log says the installation was successful and it was loaded with plugins for Claude Code and Codex.
The problem is he never ordered the installation.
An installed log
He used Postman 12.33.4.
The trigger for the line in the log is a process of upgrading specifically for older users.
In other words, as long as you're an old user with a login status, open it once, and it does it.
There are three things left behind.
The first one, in his Codex configuration, had an additional plugin market, a command to start automatic reading each time, and it had been marked as credible.
The second is in Claude Code's settings, where Postman's plug-in opens from that moment on.
The third is in his home directory, with an additional postman command and an official remote connection, which is simply not connected.
The contents of the start-up instruction are to the effect that you encounter work related to interfaces and load the Postman code first.
It doesn't sound heavy, but it's deep.
It's like writing a sentence in someone else's tools, and then these kinds of jobs come to me without you noding.
It's not for anyone in 10 days.
On September 11th, Postman turned on the plug-in warehouse.
On 24 October, for the first time, automatic installation was caught in the log.
In the early hours of October 26, the developer wrote a long report with the word "serious" in its title.
Everything is in the report. The original log, the configuration screenshot, the source-code link of the installer, and the six claims are clearly listed.
Then there's no more.
For eight days, there was no official reply from the warehouse and the report was not labelled.
On the morning of 3 November, the first comment appeared, and the next day, 50 articles came in.
November 4th morning, someone moved it on Hacker News.
And when I went to check, the post was more than 1400, and more than 460 comments.
During the day, the Chinese circle also followed, with 100,000 readings in a few hours.
There are three of them.
The first is to pack and leave.
Their words were simple and opened an interface debugging tool, which changed my programming assistant.
The three most common names under these posts, Bruno, Insomnia and Milkman, are all similar tools that can be used free of charge by offline.
The second person is cooler and more pointing.
They say the real hole isn't in Postman, but the trust we give AI is just a line of configuration that anyone can write.
Postman just happened to be the software that wrote it.
The third is very few, but crucial.
They reminded me that the popular title was "Trying to share all API data," and it didn't hold.
It's not sudden.
Postman's reputation has fallen for years.
Last October, it cut off the free version of the small team, and there was only one team left, and there was one round of it.
In June of this year, its own security page admitted that a third-party service account had been borrowed to take customer lists and quotations from its sales system.
That was when someone broke down the server, this time by themselves.
The same precedent is that Git Kraken publicly apologized in April of this year for similar actions, and the official blog wrote that they strengthened AI workflow, but missed the key steps.
It's not gonna hold.
There are only three layers of material available.
The first level is solid.
The old user will be plugged in once, the configuration of the two assistants will be modified and the start-up command will be marked as credible.
Two people who don't know each other, they both appear in both versions.
The second floor is not clear.
Whether or not the order really ran up, there was no agreement between the two, one with a record and the other with no.
There is no evidence on the third floor.
The uploading of data is not supported by any log.
The material had been left with only one remote connection and could not be connected to it.
The author of the report specifically stated at the end that he did not advocate the release of data.
I'm trying to make sure I'm not hit.
Three orders will suffice.
grep -i postman ~/.codex/config.toml
grep -i postman ~/.claude/settings.json
ls ~/.local/bin/postman ~/.copilot/settings.json
To clean up, there is an unloaded port in the client menu, in the "Postman CLI and Plugins" below the help.
But there's been some feedback that Copilot's uncleaned, and it's still in the configuration.
As of November 4, Postman's official account number was still running the performance test, and no statement was made on the matter.
The author of the report made only one request to treat it as an accident at the product boundary, not as a document problem.
At the end of the day, it was not an extra plug in a company that was uncomfortable.
It's a default that we've given permission, which can be reconfigured, to every software that's willing to reach out.
Did you get a hit on the machine? Welcome to the comment section.
Address: https://github.com/postmanlabs/postman-app-support/issues/1437.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.