이미지: METAL LAB 생성
Summary
- The ChatGPT Work cloud browser can now keep working on tasks even on sites that require login, according to release notes published August 25
- Login credentials aren't exposed to the model or stored anywhere, and the authentication feature is available on Plus and Pro plans
- For hard-to-reverse actions like bookings or payments, ChatGPT always checks with the user in the chat window first
- 발표일
- 2026년 8월 25일, 오픈AI 챗GPT 릴리즈노트
- 대상
- 챗GPT 워크 웹·모바일 클라우드 브라우저
- 인증 이용 조건
- 플러스·프로 유료 요금제
- 클라우드 브라우저 이용 조건
- 프리·고를 제외한 유료 요금제, 지원 지역
- 로그인 방식
- 챗GPT가 로그인 화면 자동 표시, 비밀번호 관리자·2단계 인증 코드 지원
- 보안 장치
- 별도 검토 모델이 피싱·기만 여부 사전 점검, 아이디·비밀번호는 저장 안 됨
- 이전 단계
- 2026년 7월 24일에는 사용자가 직접 클라우드 브라우저를 넘겨받아 로그인해야 했음
- 최초 단계
- 2025년 1월 오퍼레이터 리서치 프리뷰는 로그인·캡차·결제 화면마다 사용자에게 제어권을 넘김
Sites that require login are now fair game too
According to an update OpenAI posted to its ChatGPT release notes on August 25, the ChatGPT Work cloud browser can now complete tasks on sites requiring login, on both web and mobile. Previously, users had to take over the entire cloud browser session to log in themselves. Now ChatGPT surfaces the login screen directly, and once the user enters a username and password, ChatGPT picks up the rest of the task. The authentication feature is available on Plus and Pro plans.
From Operator to here — how it evolved
To understand this update, it helps to look back at the last year and a half of changes. Operator, OpenAI's research preview launched in January 2025, handed full control back to the user every time it hit a login screen, CAPTCHA, or payment page. On July 24, 2026, OpenAI announced via an X post from its developer account that ChatGPT Work agents could now work with sites requiring login — but at that stage, users still had to take over the cloud browser themselves to log in, though the session persisted afterward so repeated logins weren't necessary.
The August 25 update cuts that process down further. Now ChatGPT can display the login screen on its own, so users no longer need to take over the entire browser. Password managers work as usual, and two-factor authentication codes can be entered when needed.
Where to find it
It starts in the ChatGPT Work web or mobile browser. There's no separate menu path for this — you simply ask ChatGPT Work to complete a task, and the cloud browser kicks in automatically. When it reaches a page requiring login, ChatGPT pauses and asks the user to log in.
Step-by-step
- Tell ChatGPT Work what you want done — for example, "Check the X-ray cost on this insurance portal."
- ChatGPT opens the site in the cloud browser and pauses when it hits a screen requiring login.
- Enter your username and password in the secure login window. Password managers work here, and you can enter a security code if needed.
- Before any information is submitted, a separate review model checks the request and the destination for signs of phishing or deception. Users can review the site's URL and a preview of the login form, and can inspect the actual site directly before proceeding.
- Once logged in, ChatGPT continues working within that session. The authentication persists across subsequent tasks until it expires, so there's no need to log in again each time.
- Before hard-to-reverse actions like confirming a booking or making a payment, ChatGPT asks for confirmation again in the chat window.
- Under Settings > Cloud Browser > Browser Data, users can delete browsing history at any time, either entirely or for specific sites.
Who can use it
| Category | Details |
|---|---|
| Plans with authentication feature | Plus, Pro |
| Cloud browser itself | Paid plans excluding Free and Go, in supported regions |
| Note | Availability may vary based on phased rollout and workspace permissions |
What you can try
OpenAI's published examples lean toward everyday tasks. They include looking up utility providers for a new apartment and signing up for the right plan, booking a DMV appointment or filling out passport renewal paperwork, logging into an insurance portal to check the cost of an X-ray or blood test, and finding candidate profiles that match hiring criteria and show interest in moving jobs. The cloud browser documentation also mentions business use cases, like logging into utility accounts to compare rate plans or cross-checking billing records against accounting software.
Access permissions and payments stay with the user
By default, ChatGPT asks for permission before accessing a new site. Under Settings > Cloud Browser, users can choose from "Always check" for manual review every time, "Auto-approve" for sites that look safe, or "Always allow (not recommended)" for blanket access — and permissions can also be set site by site. This access setting is separate from confirmation for critical actions like bookings or payments — granting site access doesn't automatically authorize a payment.
If ChatGPT gets stuck during a task, it will ask the user to take over the cloud browser, or provide a control link if the user requests to take over first. Some websites have security measures that block automated browser access, so sites that work fine in a personal browser may still block ChatGPT from accessing them or completing tasks. OpenAI notes that these restrictions are set by the websites themselves, not by ChatGPT.
The cloud browser is fully separated from the browser on a user's personal device, with its own cookies and login sessions. It doesn't pull in accounts already logged in on the device, saved passwords, or open tab history, and clearing data for a specific site logs the user out of that site.
For more on how ChatGPT Work is being used, see ChatGPT Work: Internal demo shows everything from meeting briefs to strategy decks, which covers how sales, marketing, and strategy teams are actually sequencing the tool in practice.
Editor's take
If you had to sum up this update in one line, it would be: agents have started managing people's internet accounts for them. Back in January 2025, Operator dropped everything the moment it hit a login screen. A year and a half later, login itself has been pulled inside the agent's workflow. This isn't unique to OpenAI, either — it's a signal that the next battleground for browser agents is shifting from "how smart is it" to "how many account-gated sites can it operate on your behalf."
The generational leap is easy to see side by side. Under Operator, a login page meant the task stopped cold and a human had to sit back down at the keyboard. By July, humans still logged in, but the session persisted afterward, cutting down on repeat logins. With this August update, ChatGPT itself pulls up the login screen and plugs straight into a password manager, narrowing the human's role down to a single moment: typing in a username and password. The only intervention point left is confirming bookings or payments — so the next phase of competition will likely be about how carefully that confirmation step gets preserved.
From a practical standpoint, the sweet spot for teams to try this now is repetitive, rule-based login work — utility bills, insurance, government appointments — where the process is fixed and predictable. Tasks that involve shifting negotiation terms are still safer left to human review for now. If rolling this out to a team, starting with "Always check" and gradually moving trusted sites to individual approval is a reasonable way to manage the risk.
Expect rival agents to roll out similar login-integration features in the coming weeks. But which sites choose to block automated browsers will likely be the real determinant of how well this works in practice — logging in itself just got easier, but the next gate is still controlled by the website.
Comments