In the first week of August, an IT service desk gets the same ticket three times in one day. Atlas has stopped working, which browser should I use now? One colleague has already sorted it out: the ChatGPT extension is installed in Chrome, and at the first permission prompt they chose to allow it on every site.
Atlas has gone, but what it did has not. It has moved, and one of the places it moved to is the browser your organisation already manages.
#What happened to ChatGPT Atlas
OpenAI released Atlas in October 2025 as a standalone browser for macOS, with an agent mode that could carry out tasks while you browsed. According to OpenAI's announcement (Checked 2 November 2026), it was available on Free, Plus, Pro and Go, in beta for Business, and for Enterprise and Edu if an admin turned it on.
In the ChatGPT release notes (Checked 2 November 2026), OpenAI said it was deprecating Atlas as it brought browser-based agentic capabilities into ChatGPT and Codex, with Atlas scheduled to stop working on 9 August 2026. Bookmarks, open tabs and browsing history would not transfer automatically. Users could export cookies and passwords to the ChatGPT desktop app and bookmarks to Chrome. ChatGPT conversations are separate from Atlas and stayed where they were.
OpenAI's help article on the retirement (Checked 2 November 2026) confirms that Atlas has been retired and no longer gets updates. It includes a short list for workspace admins: share the article with members who used Atlas, help them move to an approved, supported browser, update internal documentation that still mentions Atlas, and remind people not to share passwords, cookies or session files. A browser that gets no updates gets no security fixes either, so any Mac that still has Atlas on it is worth cleaning up.
#Where Atlas's browser features went
The release notes name two successors: the ChatGPT desktop app for deeper browser work, and the ChatGPT extension or sidebar in Chrome for help while browsing. OpenAI's documentation on the browser (Checked 2 November 2026) adds a third, a cloud browser. They work differently, and the difference decides what ChatGPT gets to see.
The built-in browser is the closest thing to Atlas: a window in which you and ChatGPT see the same page, and in which ChatGPT can click and type. OpenAI says ChatGPT asks for confirmation before sensitive actions such as submitting information. The cloud browser is available on paid plans except Go. On Enterprise an admin has to switch it on first, and signing in to websites through the cloud browser is not available in Enterprise or Edu workspaces.
#What the ChatGPT extension can see in Chrome and Edge
The extension is where things change for an organisation. Atlas was a separate browser that someone had to choose to install. The extension sits in the browser where the work already happens, alongside the sessions for the mailbox, the CRM and the case management system.
According to the extension guide (Checked 2 November 2026), ChatGPT can read or act on sites where you are already signed in, and OpenAI gives LinkedIn, Salesforce, Gmail and internal tools as examples. At installation, Chrome may ask for permissions such as reading and changing all your data on all websites. Side chat is available in Chrome, Edge, Brave and Vivaldi; from the desktop app, ChatGPT can use tabs and control the browser in all five.
By default ChatGPT asks before using each new website, offering Allow once, Allow for this site, Allow for all sites or Decline. Allow for all sites carries an elevated risk warning. Access to browsing history is requested every time. OpenAI does not keep a separate full log of browser actions, but it does keep what enters the conversation: page text, screenshots and the steps ChatGPT took.
OpenAI is candid about the risk. The guide asks you to treat page content as untrusted, warns that misleading content raises the risk of ChatGPT copying browser history data somewhere unintended, and advises keeping secrets and highly sensitive data out of browser tasks. Researchers showed the same kind of risk earlier with Perplexity's Comet browser. How far an extension's permissions reach, AI or not, is covered in Browser extensions are AI tools too.
#The settings admins have
For managed devices, OpenAI documents managed configuration for the desktop app (Checked 2 November 2026). Among other things, you can disable the built-in browser and restrict browser control and Computer Use. The guide to local security (Checked 2 November 2026) lists a separate setting that disables agent-driven browser control through supported extensions. The same page says policies can restrict site access, uploads, downloads and browsing history access, and that a user's approval cannot override an admin's denial.
Two caveats from that documentation. The built-in browser, browser control and Computer Use are separate settings, and OpenAI recommends checking each path on its own. And when ChatGPT operates a browser through Computer Use, the site and capability policies do not apply. Switching off the built-in browser alone does not cover everything.
The allow and block lists in the desktop app, under Settings > Computer Use, are managed by the employee; an admin's denial in policy takes precedence. Whether the extension may be installed at all is something you control with the same extension policy you use for any other extension: ExtensionInstallBlocklist and ExtensionInstallAllowlist in Chrome, and policies of the same name in Edge (both checked 2 November 2026).
#What your organisation should decide about AI in the browser
An AI browser is no longer one product. It is three paths with different reach, and each deserves its own entry in your AI register.
Atlas. Take it off the approved software list and have admins remove it from Macs. It gets no more updates, and OpenAI's own help article asks admins to move people to an approved browser.
The desktop app's built-in browser. Part of your decision about ChatGPT itself, with a condition of its own. Decide whether the browser is on, and which sites ChatGPT may work on.
The ChatGPT extension. The weightiest decision, because it works inside your people's signed-in sessions. Record whether it is allowed, in which browsers, and what an employee may let it do: summarising a page is not the same as letting it fill in a form in the CRM. Allow for all sites usually does not belong on that list.
The cloud browser. It works without the cookies and passwords from the employee's own browser, which makes it the least intrusive of the three. On Plus and Pro, people can still sign in to websites in it themselves. Whatever it brings from a page into the conversation falls under the same terms as the rest of ChatGPT.
To make those decisions you need to know which AI tools are already being opened in the browser. The BeeSensible extension in Chrome and Edge notices when someone opens an AI tool and reports only the domain, at most once a day per device, with no path, no page content and no user id. See shows this as counts and patterns, and in Decide you record what is allowed for each tool, under which conditions and who owns the decision. Employees get a notice when they open a tool the organisation has restricted or not allowed. For security teams putting this next to their own browser management, there is AI governance for security teams.
