OpenAI** has added signed-in website support to ChatGPT Work’s browser on web and mobile, letting it continue working after the user logs in and steps away. For users, that creates a clear decision: which jobs are defined enough to hand off, and what controls they need before Work continues without them.

OpenAI launched ChatGPT Work in July as an agent for longer, more involved tasks. Its remote cloud browser already existed, but OpenAI’s launch documentation said it could not sign in to websites or complete payments. The update removes that login barrier from the existing browser, allowing Work to reach more of the online jobs that previously stopped at an account login page.

When Work reaches a supported sign-in page, it pauses while the user enters a username, password, and any required security code through OpenAI’s secure sign-in flow. OpenAI says those credentials go directly to the remote browser, are not visible to the model, and are not stored by ChatGPT. Its release notes also say they are not used in model training. Once authentication is complete, Work resumes the delegated task through the signed-in session and can keep working after the user closes the computer or turns off the phone. That lets the user focus elsewhere while Work continues the defined job. The same independence creates the central concern: whether users can keep Work inside the assignment, retain approval over consequential actions, and end the signed-in session when they choose. The feature is most useful for personal and business tasks that involve several steps after login, because those jobs can save users the most time and attention.

ChatGPT Work’s signed-in website support lets users personally complete a supported login, then hand an account-based task back to its remote cloud browser to continue after they step away.

A website login is often only the first step in a longer online task. OpenAI says users can sign in, hand a task back to Work, and leave while its remote browser continues working on a supported website. The task can keep running after the user closes the computer or turns off the phone, pausing when it needs input, another sign-in, or confirmation. That is most useful when Work can handle several steps after login, such as searching, comparing options, entering information, and preparing something for the user to review.

A DMV appointment makes that handoff easier to picture. OpenAI says Work can find an appointment and prepare the booking for approval. A user can specify the location, acceptable times, and other constraints, let Work search, then review the proposed appointment before approving it. OpenAI suggests similar uses for comparing utility plans, checking insurance costs, saving apartment listings, and preparing reimbursement paperwork.

OpenAI also applies the idea to business work. Work could find job candidates and draft outreach, move invoices from email into accounting software, fill out permit applications, or add client-call actions to a vendor portal. These jobs involve several steps between the initial request and a reviewable result, which gives Work more useful work to take off someone’s plate.

The examples extend Work beyond software development into everyday administration and business operations. However, completing a quick, one-step action directly may still be faster. Delegation makes more sense when the job has clear limits, can be reviewed, and requires enough work to save meaningful time and attention. That ability to leave the task running comes from the cloud browser, which serves a different purpose from the built-in browser in the desktop app.

ChatGPT’s built-in browser is the better fit when a user wants to watch and guide the work. The user opens it inside the desktop app, where they and ChatGPT can see the same page, and can sign in when a website asks. But that browser depends on the computer running the desktop app. OpenAI says Work’s cloud browser runs on a separate remote computer, so the task can continue after the user closes their computer or turns off their phone. Its purpose is not to hide the browsing. It is to let the job continue without keeping the user’s device running.

The two browsers also work differently. The built-in browser does not automatically use the person’s Chrome profile, cookies, open tabs, or existing website logins. It has its own browser state, and the user opens it from the desktop app before asking ChatGPT to work on the current page. With Work on web or mobile, the user describes the result they want instead of choosing the cloud browser themselves. OpenAI says ChatGPT decides whether the task requires a connected app or plugin, the cloud browser, or both. The cloud browser also keeps its own cookies and signed-in sessions rather than inheriting them from the user’s regular browser.

OpenAI introduced ChatGPT Work on July 9 alongside Work on web and mobile and the built-in desktop browser. The cloud browser existed then too, but OpenAI’s launch documentation said it could not sign in to websites or complete payments. The August 25 update adds supported website logins to that existing remote browser; it does not introduce a new browser or replace the built-in one.

Previously, reaching a website login was where the cloud browser’s work stopped. It could not ask the user to sign in and then resume the task inside the account. Now, when Work reaches a supported login page, it pauses and asks the user to complete the sign-in. The user still has to be there for that step, but can hand the task back afterward and leave while Work continues through the signed-in pages. At that point, the remote browser has an authenticated session it can use to complete the assignment. That makes the difference between protecting the user’s password and controlling Work’s access to the signed-in account especially important.

When Work reaches a supported sign-in page, it pauses and asks the user to authenticate through OpenAI’s secure sign-in form, including any required two-factor authentication or security code. OpenAI says another model first reviews the request and destination for signs of phishing or deception. The user can also check the website address, preview the sign-in form, and inspect the live site before entering anything.

OpenAI says the username and password entered through that form go directly to the remote browser and are not visible to the model. The company also says ChatGPT does not store those credentials, while its release notes state that they are not used in model training.

That protection applies to the login credentials, not everything behind the login. Once the user signs in, Work can use the account pages, information, and controls available through the session to carry out the task. OpenAI does not say the browser is technically confined to only the page, record, or field named in the request. In other words, the user’s instructions tell Work what job to perform, but the account may make more information and controls available than that job requires. What Work can actually reach still depends on the account’s permissions, the website, and ChatGPT’s controls.

Two separate things can continue after the user signs in. First, the job the user already assigned can keep running while they are away, pausing if it needs more information, another sign-in, or confirmation. Second, the remote browser can stay signed in after that job ends. Staying signed in does not mean Work is still working, and it does not authorize a new task. It only means that if the user later assigns another task involving the same website, they may not need to log in again. The session remains available until it expires or the user clears that site’s cloud-browser data. Because account access can remain available even when no task is running, users need clear control over which websites Work can enter, which actions require approval, and when the browser is signed out.

Users can decide how much permission Work has to enter websites. By default, ChatGPT asks before opening a new site in the cloud browser. The user can keep that Always ask setting, choose Auto approve so ChatGPT reviews the address and pauses only if it appears unsafe, or select Always allow—an option OpenAI says it does not recommend. These choices can be changed later under Settings → Cloud browser, where users can also allow or block individual websites. Permission to enter a site does not automatically approve everything Work might do there.

OpenAI says ChatGPT is also designed to request confirmation before actions that are difficult to reverse or create financial, legal, account, or other real-world commitments, such as confirming a booking or making a payment. That is OpenAI’s description of the product’s safeguard, not an independent guarantee that every consequential action will be identified.

Users can also end a website’s stored access or take control of the browser themselves. Under Settings → Cloud browser → Browser data, they can clear one site’s data to sign Work out of that account or clear all browser data to sign it out across websites. Changing those settings controls future access but does not reverse an action Work has already completed. If Work gets stuck, it can ask the user to take over. Users can also request takeover in the chat, and ChatGPT will provide a link that opens the remote browser on their phone or computer.

These controls do not mean every signed-in task will work without intervention. Some websites block automated browsers, and Work may be unable to complete certain logins, CAPTCHAs, transactions, or final steps on its own. OpenAI says it tests the system against prompt injection, phishing, and unintended actions, but warns that those safeguards do not eliminate every risk. The company also says information Work uses from websites is handled according to the user’s ChatGPT data-control settings.

OpenAI says cloud browsing is available on paid ChatGPT plans other than Free and Go. Availability of website sign-in may still vary by rollout and workspace settings, and Enterprise administrators must enable cloud browsing for their workspaces. Availability determines who can try the feature; its controls and limits determine whether a particular account-based task is worth delegating.

For users who have access, the choice is whether a clearly defined job will save enough time and attention to justify giving Work access to that particular account.

Website authentication changes the category of work ChatGPT Work can complete. Public browsing can gather information, but many personal and business tasks happen inside account portals, forms, scheduling systems, and operational records. Crossing that login boundary allows Work to move beyond research and handle more of the steps needed to prepare or complete an account-based job.

ChatGPT Work may offer the clearest benefit to people, small businesses, and operations teams that regularly complete multi-step work inside online accounts. What begins as a simple request can require several rounds of searching, comparison, data entry, and preparation before there is anything useful to review. A clear outcome and constraints give Work a defined chain of steps it can carry forward while the user focuses their attention elsewhere. A quick action may still be faster to complete directly, while a defined job that can continue until it reaches a review or approval point offers more meaningful delegation value.

Delegating signed-in tasks to ChatGPT Work is an account-by-account decision rather than a blanket judgment about whether the product can be trusted everywhere. Because Work can continue without supervision, three choices become especially important: what job Work is authorized to complete, which actions require approval, and whether the login should remain available after the task ends.

In short, the best starting point for ChatGPT Work is a bounded, reviewable job that saves enough time and attention to justify the access it requires.

As AI agents move further into signed-in systems and complete more work without supervision, their usefulness will depend on whether people can set boundaries, step away, and still trust what they find when they return.

Q: What changed with ChatGPT Work’s website sign-in?
A: On August 25, OpenAI added supported website sign-in to ChatGPT Work’s existing cloud browser on web and mobile. When Work reaches a supported login page, it pauses for the user to sign in and then resumes the delegated task, which can continue after the user steps away.

Q: Can ChatGPT Work see or store my password?
A: OpenAI says usernames, passwords, and required security codes entered through its secure sign-in flow go directly to the remote browser and are not visible to the model or stored by ChatGPT. OpenAI’s release notes also say the credentials are not used in model training.

Q: What can ChatGPT Work access after I sign in?
A: After authentication, ChatGPT Work can use the account pages, information, and controls available through the signed-in session to complete the assigned job. OpenAI’s documentation does not describe a separate technical fence restricting Work to one page, record, or data field, so what it can reach also depends on the account’s permissions, the website, and ChatGPT’s controls.

Q: What’s the difference between ChatGPT’s built-in browser and cloud browser?
A: ChatGPT’s built-in browser runs inside the desktop app and is designed for work the user can watch, guide, and annotate. ChatGPT Work’s cloud browser runs remotely, keeps separate browser data and signed-in sessions, and can continue a task after the user closes their computer or turns off their phone.

Q: Will ChatGPT Work keep working after I leave, and does it stay signed in?
A: The assigned task can continue while the user is away, and the cloud browser can separately remain signed in after that task ends. A retained login does not keep the task running or authorize ChatGPT Work to start another job; it only lets a separately authorized future task reuse the session until it expires or the user clears that website’s browser data.

Q: How do I control or stop ChatGPT Work’s access to a website?
A: Users can control how ChatGPT Work receives permission to enter websites, allow or block individual sites, take over the remote browser, and clear a site’s cloud-browser data to sign out. OpenAI says ChatGPT is designed to request confirmation before difficult-to-reverse actions such as bookings and payments, but not every website or step will work, and OpenAI warns that its safeguards do not eliminate every risk.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing support, AEO/GEO/SEO optimization, image concept development, and editorial structuring support from ChatGPT, an AI assistant. All final editorial decisions, perspectives, and publishing choices were made by Alicia Shapiro.