ChatGPT Atlas

ChatGPT Atlas was OpenAI's AI browser, a Chromium-based browser for macOS with ChatGPT built in. It launched on Oct 21, 2025, and its agent mode, a preview for Plus, Pro and Business users, could navigate, click and finish multi-step tasks like booking an appointment or filling a grocery cart inside the user's own browser. OpenAI said in March 2026 that it would fold Atlas, the ChatGPT desktop app and Codex into one desktop app, announced the retirement on Jul 9, 2026 and shut Atlas down on Aug 9, 2026, less than ten months after launch. The Windows version never shipped.

People still search for how to detect Atlas, partly because older logs and fraud reviews mention it and partly because its agent features did not go away. They moved into a browser inside the ChatGPT desktop app, a ChatGPT Chrome extension and a cloud browser that OpenAI runs for agents, and that cloud browser is the part site owners can actually verify. Across sites that use Verisoul, the signed OpenAI cloud browser (ChatGPT agent) made up 13% of the personal-agent sessions we named between Oct 29 and Nov 1, 2026, and it was the most widespread agent after Meta Muse, turning up on about 40% as many customer sites as Muse.

How ChatGPT Atlas showed up on your site

Atlas agent mode ran on the user's own Mac, inside the user's own browser, with their cookies, signed-in sessions and home or office IP. OpenAI never documented a user agent token, request signature or IP list for Atlas traffic, and page requests carried a stock Chrome on macOS user agent with Client Hints that matched Google Chrome, so to a website a visit looked like a person on Chrome for macOS. The one side channel was tab previews, because Atlas fetched /favicon.ico for its sidebar with a CFNetwork user agent naming ChatGPT Atlas, which could tag the visit at the network layer even though the page requests were clean. OpenAI's published chatgpt-user.json list covers only its separate ChatGPT-User fetcher. The successor cloud browser is a different story, because it runs on OpenAI's side and signs its requests with Web Bot Auth, sending a Signature-Agent header for chatgpt.com with keys published in OpenAI's key directory, and Cloudflare recognizes it in its signed agents program under the tag chatgpt-agent.

Why ChatGPT Atlas was hard to detect

Atlas was a real browser on a real person's computer, which removed every signal that usually separates a bot from a customer. There was no datacenter IP, no headless browser and no declared identity, the session cookies belonged to an account the site already trusted, and the agent logic sat outside Chromium instead of being injected into pages as JavaScript, so a page script had nothing to find. The only evidence left was behavior. Researchers at UC Davis studied Atlas in their FP-Agent paper and saw its agent paste text into fields instead of typing, scroll in bursts and move the pointer straight to click targets without the wandering path a hand makes.

The successors split the problem in two. OpenAI's cloud browser is the easy half, since its signature can be verified and it runs on a cloud machine. The browser in the ChatGPT desktop app and the ChatGPT Chrome extension run on the user's own device like Atlas did, and we have not seen OpenAI document a label for either, so they need the same behavioral approach.

How Verisoul detects ChatGPT Atlas and its successors

Verisoul names OpenAI's cloud browser by verifying its Web Bot Auth signature against OpenAI's published chatgpt.com key, then checking that the device and the behavior match a cloud agent. A signature can't be copied the way a user agent can, so a chatgpt_agent result means the session really came from OpenAI. In-browser sessions, which is what Atlas produced and what the desktop app browser and Chrome extension produce now, carry no signature, so Verisoul flags them through bot score when the behavior is automated and does not name them.

"agent": {
  "id": "chatgpt_agent",
  "name": "ChatGPT agent",
  "provider": "OpenAI",
  "type": "ai_agent",
  "detection_method": "signed",
  "detection_signals": [
    "cryptographic",
    "behavioral"
  ]
}

• id is chatgpt_agent, the stable name for OpenAI's signed cloud browser
• type is ai_agent, meaning software acting for a ChatGPT user
• detection_method is signed, meaning the identity was proven with a cryptographic signature
• detection_signals lists the verified signature, a cloud-hosted browser with a virtual machine device profile, and automated clicks, typing and page flow

The Agent Atlas entry for ChatGPT agent describes the same signals, and the AI agent detection docs explain how agent and bot score fit together in the response.

How to detect ChatGPT Atlas successors step by step

1. Add the Verisoul SDK to sign-up, login, checkout and payout pages, following the AI agent detection docs
2. Read the agent field and tag chatgpt_agent sessions as OpenAI's verified cloud browser
3. Treat a high bot score with no agent name on a known account as a possible in-browser agent, such as the ChatGPT desktop app browser or Chrome extension
4. Write separate rules for signed and unsigned agent sessions on risky steps and review them weekly

Should you block ChatGPT Atlas?

Atlas itself no longer sends traffic, so there is nothing to block. For its successors, the signed cloud browser is working for a real ChatGPT user, which is why most teams let it browse and check out and send it to review on sign-up and payouts, where an agent creating accounts or moving money deserves a closer look.

Unsigned in-browser sessions are usually the account holder getting help with an errand. A hard block would turn away real customers, so a high bot score on those sessions works better as a reason to review card changes, payouts and repeated payment attempts than as a reason to stop the session.