AI browser risk review

ChatGPT Atlas Security Concerns for Businesses

Atlas is retired and gets no security updates, yet copies still hold work logins on Macs. Its agent features moved into the ChatGPT app and a Chrome extension, so the questions for IT moved with them.

StatusDiscontinued in 2026
Security updatesNone since retirement
Where the risk lives nowChatGPT app and Chrome extension

Next step

Close the gap Atlas left with the AI Agent Blocklist

Identifiers and domains of AI browsers, agent apps and agent extensions, so leftover installs and their successors are found and blocked.

From $99 a month for one organisation

Is ChatGPT Atlas safe? Plain answers

These are the questions people type about Atlas, answered in one or two sentences each. The rest of the page explains the detail behind them.

Is ChatGPT Atlas safe?

Retired

It no longer receives security updates, so it does not belong on work devices. Plan the removal of leftover copies.

Is ChatGPT Atlas dead?

Yes

OpenAI announced its discontinuation on 9 July 2026 and set 9 August 2026 as the date it stops working.

Is ChatGPT Atlas going away?

Already gone

It has gone as a product. Installed copies stay on Macs until someone removes them.

Is ChatGPT Atlas out yet for Windows?

Never

Atlas shipped only for macOS on Apple silicon. Windows and mobile versions were announced but never released.

Is ChatGPT Atlas free?

Was free

It was a free download for ChatGPT users on macOS. Agent mode needed a paid plan.

Is ChatGPT Atlas good for business?

Retired

For a company the test is central control and regular updates. A retired product offers neither, so decide on its successors instead.

What is the ChatGPT Atlas app?

A browser

A Chromium based Mac browser by OpenAI with ChatGPT built in: a sidebar that read pages and an agent that acted in websites.

Does Atlas still need attention?

Yes

Leftover copies keep their profile data on disk. Agents that connect to tools raise related questions: see MCP security risks.

How Atlas handled pages, logins and data

To judge the risk that is left, it helps to know what Atlas did. Each feature below still exists somewhere in the ChatGPT product family, shown in the last line of each card.

The ChatGPT sidebar

Users could ask ChatGPT about the page they were on. To answer, the page content went to ChatGPT, including pages from internal tools.

Now: the ChatGPT extension for Chrome and the app's built-in browser.

Agent mode

On paid plans, an agent could click, type and complete tasks in websites, including sites where the user was signed in.

Now: ChatGPT Work and browser use in the ChatGPT desktop app.

Browser memories

An optional feature that let ChatGPT remember facts from sites the user visited. Retention followed the terms of the account in use.

Now: check the memory settings of the ChatGPT account in use.

A normal browser profile

Tabs, history, saved passwords, cookies and sessions, stored in folders on the Mac like any browser.

Now: left behind. This data did not transfer when Atlas was retired.

Enterprise workspaces: Atlas was off by default for ChatGPT Enterprise, and an admin had to turn it on. Admins could also switch agent mode on or off. If your workspace never turned Atlas on, members could still have used it with personal accounts.

The risk changed on 9 August 2026

Retirement removed some risks and created others. Compare the two columns before you decide that a dead browser needs no attention.

Before While Atlas was supported

  • Prompt injection: text on a page could try to steer the agent.
  • Agent actions inside logged-in sessions, with the user's rights.
  • Page content sent to ChatGPT when users asked about a page.
  • Regular security updates to the browser engine.
  • Workspace admins could leave it off or switch agent mode off.

After Since retirement

  • No security updates. Fixes released for the browser engine after retirement never reach leftover copies.
  • Saved passwords, cookies and sessions sit in an unmaintained app.
  • OpenAI said it may no longer open, browse or run agent tasks, but nothing removes it.
  • The same agent features run in the ChatGPT app and Chrome extension.
  • Users who liked Atlas may install another AI browser instead.

Why "it may not open" is not enough: a browser that fails to start still holds its profile data on disk. Removal and, where needed, password changes are the only clean finish.

Risk patterns that apply to AI browsers in general

Researchers showed in 2025 that AI browser assistants in general can be steered by what they read. These patterns describe the category, not one product, and they explain why security teams grew cautious about any agent that reads web pages.

  • Hidden text

    Instructions hidden in a comment or page can reach the assistant when a user asks for a summary. The assistant then works with whatever the user is signed in to.

  • Address bar

    Text pasted into the address bar that only looks like a web address can be handled as a request from the user. Ask staff not to paste text from unknown sources there.

  • Memory

    Assistants that remember across sessions can store a planted instruction and act on it later. Review the memory settings of every account in use.

  • December 2025

    Gartner published its advisory "Cybersecurity Must Block AI Browsers for Now", which covers AI browsers as a category.

  • Today

    AI browser makers say publicly that prompt injection remains an open problem for the industry, so the controls you own still matter most.

The mechanics of these attacks, and the defences that work, are on our page about AI browser prompt injection. For the wider picture see AI browser security risks.

The Atlas risks that moved to its successors

OpenAI moved the Atlas features into the ChatGPT desktop app, the ChatGPT extension for Chrome and a cloud browser that OpenAI's agent runs on its own side. The older desktop app was renamed ChatGPT Classic. Here is what each successor can do and what to check.

Built-in browser

The desktop app has its own browser with tabs, downloads, logins, autofill, password management and extensions.

Check: is the in-app browser pane allowed in your workspace?

ChatGPT Work

An agent for longer tasks that can browse and use web tools. OpenAI's release notes said it is on for members unless an admin turns it off.

Check: is ChatGPT Work and browser use on for everyone by default?

Chrome extension

The ChatGPT extension works in the user's own Chrome profile, with existing sign-ins and open tabs.

Check: is hehggadaopoacecdllhhajmbjkdcmajg allowed or blocked in Chrome policy?

Computer use

The desktop app can operate other apps, including Safari, once the user grants permission in macOS.

Check: is native computer use off where your policy requires it?

In ChatGPT Enterprise and Edu, admins can turn off the in-app browser pane, browser use in that browser, browser use through the Chrome extension and native computer use. Site rules can allow or block, and Block wins when rules overlap. Names change, so check the current admin console.

For the agent side in detail, read ChatGPT agent mode security and ChatGPT Chrome extension security.

ChatGPT Atlas risk register for IT

Copy these rows into your risk register. Priority reflects a typical company with managed Macs and no ban on ChatGPT. Adjust it to your own data and policy.

RiskWhere it lives nowPriorityControlHow a maintained list helps
Unpatched browser engineLeftover Atlas installs on MacsHighRemove the app, add a restricted software ruleBundle ID com.openai.atlas and process name for inventory and rules
Stored passwords and sessionsAtlas folders in the user's LibraryHighDelete the folders, change passwords that lived only there, end old sessions in your identity providerFinds which Macs still hold an Atlas profile
Prompt injection steering an agentChatGPT Work, browser use, the Chrome extension, other AI browsersHighTurn off agent browsing where not needed, block agent actions on sensitive sitesExtension IDs and agent domains to block for unmanaged accounts
Page content sent to an AI serviceChrome extension, in-app browserMediumAI use policy, approved tools only, data loss controls you already runLists the extensions that read pages
Agent operating other appsComputer use in the desktop appMediumTurn off native computer use, review macOS privacy permissionsIdentifies com.openai.sky.CUAService on Macs
A replacement AI browserUsers who moved to another agentic browserMediumApp control for all AI browsers, not only AtlasIdentifiers for AI browsers as they appear
Untrusted "Atlas" downloadsFiles offered as Atlas for Windows or as a new Atlas versionLowSoftware install rules, user guidanceKnown publishers and IDs make odd ones stand out

Do not stop at the browser. Atlas was one entry point for an agent with the user's sign-ins. Our AI agent risk assessment guide shows how to score the rest of them the same way.

Who is most exposed

Atlas appealed most to people who live in the browser and like to try new tools. Those groups also tend to hold the most valuable sign-ins, so start your checks with them.

Developers and admins

Signed in to code hosting, cloud consoles and admin portals. A leftover profile can hold sessions with wide rights.

Executives and assistants

Email, calendars and shared documents in one browser. Agent features were most tempting here, for booking and inbox work.

Finance and purchasing

Banking portals and supplier sites. Agent errors here mean wrong orders or payments, which is why many policies block agents on these sites.

Shared and lab Macs

Several people sign in on one machine. Atlas folders may exist in more than one user's Library, so clean every home folder.

Your response plan for this week

Five steps take most companies from "we think Atlas is gone" to a documented answer. Each step links to the detail.

  1. Day 1: inventory. Search your Mac inventory for the Atlas app and the com.openai.atlas bundle ID. Note the users.
  2. Day 2: tell users. Give a short cut-off date. Ask them to move bookmarks and to change any work password they saved only in Atlas.
  3. Day 3: remove. Run the cleanup script from the Atlas removal guide and add a restricted software rule.
  4. Day 4: successors. Set workspace controls for the in-app browser, browser use and computer use, and decide on the Chrome extension.
  5. Day 5: record. Write the outcome into your AI policy and risk register, and set a review date for new AI browsers.

What to tell your staff

Most people who used Atlas liked it, and a removal without a reason feels arbitrary. A short, plain message gets better cooperation than a silent script.

Message you can adapt

ChatGPT Atlas was retired by its maker in August 2026 and no longer gets security updates. We will remove it from company Macs on the date below.

Before then, export any bookmarks you need and change any work password you saved only in Atlas. Use our approved browser for work. Ask the help desk if you want to use ChatGPT features at work, and we will tell you which ones our policy allows.

Ask staff to

  • Move bookmarks before the cut-off date
  • Change passwords saved only in Atlas
  • Use the approved browser for work sites
  • Ask before installing a new AI browser or agent extension

Ask staff not to

  • Download Atlas installers from file sharing sites
  • Sign in to work apps in an unsupported browser
  • Give AI apps screen or accessibility access without approval
  • Move work data into personal AI accounts

Signs to look for on your devices

Any of these means part of the Atlas risk is still present in your fleet. Most can be found with the inventory you already have.

  • ChatGPT Atlas.app still listed in Mac app inventory after August 2026.
  • The updater launch agent com.openai.atlas.update-helper still present in user folders.
  • The process ChatGPT Atlas appears in usage or process reports.
  • Atlas still set as the default browser on some Macs.
  • The ChatGPT extension installed in managed Chrome profiles without an approval record.
  • The ChatGPT app holding Accessibility or Screen Recording permission on Macs where computer use is not approved.
  • Sign-ins to work apps from personal ChatGPT accounts on company laptops.
  • New AI browsers appearing in inventory in the weeks after Atlas stopped working.

Hunting for agents more widely? Our guide to shadow AI agents covers the signs for agent apps, extensions and coding agents.

Security review questions for Atlas and its successors

Use these in a security review or a vendor questionnaire. The answers decide whether you block, allow with limits, or allow.

  1. How many Macs still hold an Atlas install or Atlas profile folders, and who owns them?
  2. Were any work passwords saved only in Atlas, and have they been changed?
  3. Which ChatGPT features does our workspace allow: chat, browsing, agent tasks, computer use?
  4. Can staff sign in to ChatGPT with personal accounts on company devices?
  5. On which sites may an agent act, and which sites are blocked by site rules?
  6. Is the ChatGPT Chrome extension allowed, and for which groups?
  7. Who reviews new AI browsers and agent apps as they appear?
  8. Where is the AI agent decision written down for staff?

Comparing AI browsers in the same review? Our page on Perplexity Comet security risks uses the same questions for a browser that is still supported.

Know which AI browsers and agents are on your devices

Atlas showed how fast an AI browser can arrive, spread through a company and then be retired, leaving profile data behind. The AI Agent Blocklist keeps your rules current as these tools come and go, updated every day.

See the full contents on the AI Agent Blocklist page, or compare plans for one organisation. If you need to act on Atlas today, the Atlas removal guide has the scripts.

Mac bundle IDs and process names
Chrome, Edge and Firefox extension IDs
Windows package IDs
AI agent domains
A risk grade on every entry
Files for Jamf, Intune, Chrome policy and firewalls

Ask us or order a plan

Tell us what you need to block and where your rules live. We reply within one business day.

  • Files for firewalls, DNS filters, secure web gateways, MDM and browser policies
  • Updated every day as new AI browsers, agents and extensions appear
  • Plans for one organisation, for MSPs and for product makers, or a hosted filter we run for you

We use your details only to answer you. No newsletters.

ChatGPT Atlas security FAQ

Is ChatGPT Atlas safe to use at work?

No longer. OpenAI discontinued Atlas, set 9 August 2026 as the date it stops working, and it no longer receives security updates. Remove leftover copies from company Macs and decide how to handle its successors.

Is ChatGPT Atlas dead?

Yes. OpenAI announced on 9 July 2026 that Atlas would be discontinued, alongside a new ChatGPT desktop app. Its features moved into that app and into the ChatGPT extension for Chrome.

Is ChatGPT Atlas free?

While it existed, Atlas was a free download for ChatGPT users on macOS, and agent mode needed a paid plan. It is no longer offered as a supported product.

What are the main ChatGPT Atlas security concerns?

Leftover installs that no longer get security updates, and the passwords, cookies and sessions stored inside them. Agent questions such as prompt injection now apply to the ChatGPT desktop app and the Chrome extension, as they do to any AI browser.

What about ChatGPT Atlas privacy?

Atlas could send the page in view to ChatGPT when asked, and browser memories were optional. History, cookies and sessions stay in the Atlas folders on the Mac until someone removes them.

Do I need to remove Atlas if it no longer opens?

Yes. A browser that does not start still keeps its profile on disk, with saved passwords, cookies and sessions. Remove the app and its folders, and change work passwords that were saved only in Atlas.

What replaced ChatGPT Atlas?

The ChatGPT desktop app, with a built-in browser, the ChatGPT Work agent and computer use, and the ChatGPT extension for Chrome. The older desktop app was renamed ChatGPT Classic.

OpenAI, ChatGPT and ChatGPT Atlas are trademarks of OpenAI. Perplexity and Comet are trademarks of Perplexity AI, Inc. Apple, Mac, macOS and Safari are trademarks of Apple Inc.

Google and Chrome are trademarks of Google LLC. Gartner is a trademark of Gartner, Inc. Names are used only to identify the products discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.