内容来源：https://lifehacker.com/tech/i-let-chatgpt-sign-into-my-web-accounts-for-me?utm_medium=RSS

内容总结：

AI助手再升级：ChatGPT Work可代登录网站，亲测体验喜忧参半

随着人工智能代理（AI Agent）能力不断增强，聊天机器人ChatGPT近日向其“Work模式”推送了一项新功能——可代用户登录网站，理论上能进一步接管日常网络操作任务。官方宣传中，该功能的应用场景包括为新公寓开通水电、预约医生以及取消改期行程的票务等，话术鼓励用户“凡是平时需要打开浏览器手动点击完成的事，都可以试着交给ChatGPT Work去做”。

不过，将个人登录凭据交由AI托管，这一设想听起来颇为大胆。它是否安全？AI在账号内操作时是否会“翻车”？带着这些疑问，有外媒记者进行了实际测试，授权其登录一些非重要账户以观察效果。目前，该功能已面向ChatGPT Plus、Pro及Business用户，在网页和移动端逐步推送。

关于密码安全：官方称不存储凭据

根据ChatGPT于2026年8月25日发布的更新说明，系统“无法看到用户的用户名或密码，这些信息既不会被模型获取，也不会用于模型训练，更不会被存储”。那么，不存密码又如何实现登录？最基础的方案是：当网站或应用弹出登录提示时，用户通过ChatGPT界面直接输入账号信息完成登录（同时可能需处理双重验证），此后AI会在未来任务中保持该登录状态。

更进阶的省时方案是授权ChatGPT与第三方密码管理器对接。虽然官方未列出具体支持的密码管理器名单，但据测试观察，目前该功能仅限移动端使用——手机上的密码管理器可在ChatGPT需要时自动弹出并填充对应凭据。

这一功能底层依托于ChatGPT内置的一个真实浏览器，AI可自主操作该浏览器。从网站视角看，它如同一台新设备发起登录，部分网站允许用户将ChatGPT标记为“可信操作方”，以便未来免密再次登入。

用户可通过网页端头像菜单中的“设置 > 云浏览器”查看并管理相关数据，例如删除已存储的Cookies，相当于强制AI退出所有已登录站点。官方支持文档称：“用户始终掌控ChatGPT Work可访问的网站范围；在涉及预订或支付等关键操作前，AI会先征求确认。”不过，AI如何精准识别即将进行的操作是否属于“关键”，目前尚不明确。

实测体验：登录流程繁琐，登录后表现尚可

记者先在网页端点击顶部“Work”切换至代理模式，尝试下达指令：“登录我的eBay账户，看看拍卖进展如何”。然而整个过程异常缓慢——等ChatGPT完成登录，记者早已手动登录多次。此外，对话界面信息并不清晰，AI索要验证码时，记者甚至不清楚该去哪里查收。

最终记者不得不切换到ChatGPT的浏览器视图，才完成短信验证码环节。登录后，AI虽在页面间错误跳转了几轮，但最终能正确播报当前拍卖价格。屏幕角落会实时显示AI所操作浏览器的缩略图，方便用户监控。

第二个任务为登录Airbnb，查找加州圣克鲁斯附近、特定价格区间内的合适房源。此次登录环节仍未顺畅——AI称Airbnb网站“不配合”，记者再度返回浏览器视图手动登录。不过一旦获取权限，AI运行较为流畅，能顺利搜索房源并汇总报价。

记者评价道：“现阶段，登录流程实在过于繁琐。网站似乎对AI浏览器有所警惕，整体速度偏慢，且常遇到验证码、通行密钥等阻碍。很多时候仍需像平时一样手动登录，才能让ChatGPT在网页上动起来。”

信任仍是最大门槛

登录成功后，只要指令简单明确，AI的表现尚算不错。在Airbnb上，它能较好理解日期与地点，但未必能根据景观或装修风格等主观感受进行筛选。

记者坦言，最关键的问题在于信任。“我始终不认为AI能比我做得更好，即便它最终可能更快。我想亲自看看Airbnb房源长什么样、谁在eBay上出价、或者修车有哪些可选时段——若全权交给ChatGPT，我始终担心它会漏掉重要信息。”测试中，ChatGPT在成功登录后并未出错或给出错误答案，但这种担忧始终挥之不去。

最终，记者决定短期内继续自行处理这些线上事务，暂不将账户托管给AI。

中文翻译：

人工智能代理正随时间推移变得日益强大，而ChatGPT Work（即ChatGPT的代理功能部分）最新推出的能力是能够登录网站——理论上，这能为你分担更多工作负担。官方给出的使用示例包括为新公寓开通水电煤等公用事业服务、预约医生问诊，以及为改期行程取消机票。其宣传语如是说：“如果某件事你通常需要打开浏览器自己一步步点击操作，不妨试着让ChatGPT Work来做。”

我完全赞成在行政事务上节省时间，但把我的登录凭据交给ChatGPT，这感觉像是迈出了一大步。我能信任ChatGPT确保我的信息安全无虞吗？一旦它进入我的账户，出乱子的概率又有多大？为了探个究竟，我让它登录了我一些不太重要的账户，看看它到底能做什么。

这项功能现已在网页端和移动端应用中向ChatGPT Plus、Pro和Business账户用户逐步推出。

ChatGPT会存储你的用户名和密码吗？

根据2026年8月25日的ChatGPT发布说明，“ChatGPT无法看到你的用户名或密码，这些信息从未被模型看到，也未用于模型训练。ChatGPT不会存储你的用户名或密码。”那么，如果你的用户名和密码从未被存储，它是如何运作的？这里最基本的选项是，在网站或应用提示时直接通过ChatGPT登录（同时完成任何双因素认证验证），然后让AI保持登录状态以用于未来的任务。

更高级的版本——也是能帮你节省更多时间的方式——是授权ChatGPT与第三方密码管理器对接。ChatGPT并未列出支持的密码管理器清单，但据我所知，这仅适用于移动端：你手机上的任何密码管理器都可以在ChatGPT需要时弹出，提供相关凭据。

支撑这一切的，是内置在ChatGPT中的一个真实浏览器，它可以自行访问和操作。你的账户会将其视为一次新的登录（就像你从新设备登录一样），有些网站和应用会允许你将ChatGPT标记为“受信任”的操作者，允许它未来再次登录。

你可以控制这个浏览器持有的数据：从网页版ChatGPT中，点击你的个人资料图标（左下角），然后进入设置 > 云端浏览器。例如，你可以从这里删除已存储的cookie，这实际上会将AI从它能访问的所有网站中踢出。根据ChatGPT的支持文档：“你可以掌控ChatGPT Work能访问哪些网站，它会在执行重要操作（如完成预订或支付）前始终征求你的确认。”然而，它究竟如何判断即将执行的操作是否重要，这一点并不明确。

当我给ChatGPT布置一些任务时，发生了以下情况

如果你想在网页版ChatGPT中使用此功能，请点击屏幕顶部的“Work”切换到代理模式，然后告诉AI你想做什么，比如：“登录我的eBay账户，告诉我我的拍卖进行得怎么样”或“登录我的记账应用，提交这些发票”。

我先试了eBay那个任务——不知何故，整个过程慢得令人发指，等到ChatGPT终于获得访问权限时，我用常规方式登录都能登录好几次了。还存在ChatGPT对话往往会混淆正在发生的事情的问题：它要求输入验证码，而我不知道该去哪里查看。

你可以随时切换到ChatGPT浏览器本身，我最终也这么做了，以通过验证码阶段（验证码是通过短信发送的）。AI在最初笨拙地浏览了几个错误页面后，能够准确告诉我我所竞拍物品的当前拍卖价格——屏幕角落有一个实时浏览器的缩略图，所以你可以看到正在发生什么。

接下来，我让ChatGPT登录我的Airbnb账户，在加州圣克鲁斯附近寻找一些价格区间内的好房源。同样，我不得不切换到实际浏览器视图进行登录（系统提示我Airbnb网站不太配合），但一旦我让AI获得了访问权限，它就运行得相当顺畅，查找房源并给我列出了价格概况。

目前，登录流程实在太不稳定了。网站似乎对AI浏览器心存警惕，一切都非常缓慢，而且你还会遇到验证码、通行密钥之类的麻烦。你通常还是得像平时一样手动登录，才能让ChatGPT在某个网站上运行起来。

不过，一旦这一关过了，只要你的指令比较容易理解和执行，一切就能以相当令人印象深刻的方式运转。ChatGPT在Airbnb上处理日期和地点似乎驾轻就熟，但我不确定它能否根据风景或整体装饰氛围来筛选房源。

也许最重要的是，我仍然不相信AI能比我做得更好，即使它（最终）能做得更快。我想亲眼看看Airbnb房源是什么样子，想知道谁在eBay拍卖中出价，或者想看看汽车保养预约有哪些可用时段——如果我把这些交给ChatGPT，我会太担心它遗漏了什么（我对让亲朋好友帮我做事也有同样的顾虑，但那就是另一个问题了）。

在我的测试中，ChatGPT一旦获得了这些网站的访问权限，在我交给它的所有任务中都没有失败，也没有给我任何不准确的答案，但如果我完全改用这种方式，这些担忧总会萦绕在我心头。目前，我还是打算自己来处理这些在线行政事务。

英文来源：

AI agents are becoming more capable over time, and the latest ability rolled out to ChatGPT Work (that's the agentic bit of ChatGPT) is being able to sign into websites—and in theory, taking even more of your workload off you. The official examples of how you might want to use this include setting up utilities for a new apartment, organizing an appointment with a doctor, and canceling tickets for a rescheduled trip. "If it's something you’d normally have to open a browser and click through yourself, try asking ChatGPT Work to do it," says the sales pitch.
I'm all for saving time on admin, but handing over my login credentials to ChatGPT feels like a bold step. Can I trust ChatGPT to keep my details safe and secure? And what are the chances of it going haywire once it's in my accounts? In the interests of finding out, I let it log into some of my less important accounts to see what it could do.
This feature is rolling out now in the web and mobile apps for those on ChatGPT Plus, Pro, and Business accounts.
Will ChatGPT store your usernames and passwords?
According to the ChatGPT release notes for August 25, 2026, "ChatGPT cannot see your username or password, and they are never seen by the model or used in model training. ChatGPT does not store your username or passwords." So how does it work if your usernames or passwords are never stored? The most basic option here is to simply sign in through ChatGPT when prompted by a site or app (and clear any two-factor authentication challenges at the same time), and then have the AI stay logged in for future tasks.
The more advanced version, which is where you save a more significant chunk of time, is to give ChatGPT permission to interface with a third-party password manager. ChatGPT doesn't specify a list of supported password managers, but as far as I can tell, this is just for mobile: Any password manager on your phone can pop up when ChatGPT needs it and supply the relevant credentials.
Underpinning all this is an actual browser, built into ChatGPT, that it can access and operate itself. Your accounts will see it as a new sign in (like they would if you signed in from a new device), and some websites and apps will let you mark ChatGPT as a "trusted" operator that can log in again in the future,
You can control the data held by this browser: From ChatGPT on the web, click your profile icon (bottom left), then Settings > Cloud browser. You can delete stored cookies from here, for example, which will effectively kick the AI out of any sites it has access to. As per ChatGPT support: "You stay in control of which websites ChatGPT Work can access, and it will always ask for confirmation before consequential actions, such as completing a reservation or payment." Exactly how it can tell when it's about to do something consequential, however, isn't clear.
Here's what happened when I gave ChatGPT some tasks to do for me
If you want to make use of this inside ChatGPT on the web, click Work at the top of the screen to switch to agent mode, then tell the AI what you want to do: Something like "sign into my eBay account and tell me how my auctions are doing" or "sign into my accounting app and submit these invoices."
I gave the eBay one a go first—the process proved to be incredibly slow for whatever reason, and I could've signed in multiple times in the normal way by the time ChatGPT got around to gaining access. There are also problems where the ChatGPT conversation tends to obfuscate what's going on: It asked for a verification code, and I didn't know where to check for it.
You can, at any point, switch to the ChatGPT browser itself, which I eventually did to get through the verification code stage (it was sent over text). The AI was then able to correctly tell me what the current auction prices were on the items I was bidding on, after initially stumbling through several wrong pages—there's a thumbnail of the live browser in the corner of the screen, so you can see what's happening.
Next up, I tasked ChatGPT with getting into my Airbnb account and finding some great spots close to Santa Cruz, Calif., within a certain price range. Again, I had to switch to the actual browser view to sign in (I was told the Airbnb site wasn't cooperating), but once I had given the AI access, it worked smoothly enough, looking up properties and giving me a rundown of prices.
At the moment, the sign-in flows are just too finicky. Sites seem suspicious of the AI browser, everything is really slow, and you run into complications with verification codes, passkeys, and the like. You'll often have to manually sign in as you would anyway in order to get ChatGPT up and running on a site.
Once that's done, though, everything works reasonably impressively, as long as you're giving commands that are fairly easy to interpret and carry out. ChatGPT on Airbnb seems competent with dates and locations, but I'm not sure it could filter through places based on the view or the general vibe of the decor.
Perhaps most importantly, I still don't trust the AI to do these jobs better than I can, even if it can (eventually) do it more quickly. I want to see for myself what an Airbnb looks like, or who's bidding on an eBay auction, or what the available time slots are for an appointment to get my car serviced—if I left it to ChatGPT, I'd be too worried about it missing something (I have the same reservations about letting friends and family do things for me as well, but that's another issue).
In my testing, ChatGPT didn't fail in any of the tasks I gave it once it had gained access to these sites, or give me any inaccurate answers, but that would always be at the back of my mind if I switched over to this completely. For now, I'm going to stick to these online admin tasks myself.