Is ChatGPT Atlas Safe for Client Work in 2026?

Assess ChatGPT Atlas before using it with client work. Distinguish browser and remote-agent data, check permissions and set the practice’s access boundary.

Published
Updated

No, don’t approve ChatGPT Atlas for new client work. OpenAI has retired it and stopped updates. Review any retained data separately from ChatGPT’s supported browser features.

For an accounting practice, the useful question is which information an assistant can read and which accounts it can act through. A privacy setting alone cannot answer both.

What the Browser Can See and Do

Atlas combined a browser on your computer with ChatGPT assistance and an agent that could operate websites. Its retirement changes the decision from approving a browser to removing it from practice use.

OpenAI’s original Atlas description explains the earlier distinctions. The sidebar used page context to answer questions. Optional browser memories retained useful browsing context. Agent mode could take actions inside the browser.

The address-bar visibility setting controlled whether ChatGPT could see a site’s content and create browser memories from it. Browsing-content training and training for chats were separate choices. Attached website content could become part of a chat.

Those historical settings explain possible exposure. They are not instructions to restart a retired browser or proof that its saved data has disappeared.

Separate the Browser Environments

An ordinary page open on a staff member’s computer, text sent to an assistant and a remote signed-in session have different access boundaries.

EnvironmentWhat it containsWhat your practice must assess
Local browser profileLocal history, saved logins and the pages staff openWhich profile holds practice access and whether an assistant can read its pages
Assistant conversationText, files and page context supplied to the assistantWhether each item is permitted to leave the original location and remain in chat history
ChatGPT cloud browserIts own remote browser data and website sessionsWhich sites are allowed and which account the assistant can use
Retained Atlas dataAtlas-specific browser information and separate ChatGPT conversationsWhich records need recovery and which access must be removed

The supported cloud browser in ChatGPT runs on a separate computer. It does not automatically use your device’s open tabs or existing browser sign-ins. A website login creates a separate remote session that can persist for future tasks.

If you are asking what remote browser data in ChatGPT means, focus on that remote profile’s cookies and signed-in sessions. Clearing your laptop’s browser history does not clear this separate profile.

The ChatGPT desktop browser also has its own profile. Its settings are under Browser, while remote website permissions are under Cloud browser. Record which experience staff actually use before applying a settings checklist.

Assess Client Data and Untrusted Pages

Assess the entire task’s access, including files and connected accounts. A request to summarise one document can expose more information if the assistant can also browse an inbox or document library.

Consider a fictional practice assessing a client’s draft financial statements. The manager first creates a dummy document with no client identifiers. A staff member asks for a summary using only that document.

If the staff member uploads the document, its contents enter the assistant conversation. If they instead sign the cloud browser into a document library, the remote session can access the account’s permitted material. These are different approvals.

For this assessment, exclude real financial statements and leave the document library disconnected. The manager checks whether the summary stays within the supplied dummy document. An unexpected request to open another account ends the exercise.

Treat Page Instructions as Untrusted

A page can contain an instruction aimed at the assistant, such as a request to send information somewhere else. OpenAI calls this prompt injection: outside content tries to redirect the assistant from the user’s task.

OpenAI’s Atlas security research describes its own adversarial testing and model hardening. Its work shows why a browser agent needs protection against instructions embedded in material it reads.

In the fictional document task, a line asking the assistant to open an unrelated upload site is outside the manager’s request. Stop the task, preserve the task record and inspect what the assistant actually accessed or sent.

Turning off model training addresses one use of data. It does not remove a signed-in account, block every misleading instruction or reverse an action already taken. Review exposure and actions separately from training preferences.

Set and Verify the Practice Boundary

Exclude retired Atlas from approved practice browsers. For supported ChatGPT browser use, start with public research and dummy examples under a named manager’s approval.

Write the permitted task in concrete terms: summarise public guidance, cite the source and leave every practice account signed out. Exclude client records, identity documents and credentials from that approval. A later client-data proposal needs its own review of account terms and access.

Give staff only the access their approved task needs. Keep permission to read a source separate from permission to change a record or send a message. Name the person who reviews the result before staff rely on it.

Confirm Settings and Remove Access

Use the following checks for an approved cloud-browser assessment:

  1. Open Settings, then Cloud browser. Select Always ask for website access and inspect individual website exceptions. Block sites outside the agreed task.
  2. Inspect connected apps separately. Remove unnecessary connections so a browser restriction does not leave another route into practice information.
  3. Use dummy data to perform the approved task. Review the visited sites and the output against the original request before approving any consequential action.
  4. Under Cloud browser, open Browser data and clear the relevant site’s cookies. Confirm that the next account-dependent task requires a new sign-in.

Website permission and action confirmation are separate in OpenAI’s documented design. Allowing a site does not approve every action there. A manager must inspect the recipient and information involved before approving a send.

For a former Atlas user, keep necessary records before removing local data. OpenAI says deleting chats does not delete Atlas-specific data, and uninstalling Atlas does not establish deletion. Use its privacy-request route for retained Atlas data questions.

Finish the practice record with the approved task, permitted accounts, reviewer and result of the sign-out check. Reopen approval if staff need another account or a different category of information.

One client form. One finished return.

Your client completes one guided form. Tax Genie checks it, chases mistakes and sends the details to your engagement letter and tax software.

See how Tax Genie works

Made for Australian accounting firms

Tax GenieSee how Tax Genie works