ChatGPT Atlas
ChatGPT Atlas was an AI-powered web browser developed by OpenAI, released on October 21, 2025 and retired on August 9, 2026, that placed ChatGPT at the center of the browsing experience: a sidebar called the Ask ChatGPT companion runs alongside any web page, an opt-in browser memory stores context from past sessions, and an Agent Mode lets ChatGPT navigate websites and complete multi-step tasks on the user's behalf [1][2]. Built on the Chromium open-source project, Atlas launched first on macOS as a free download for anyone with a ChatGPT account, with Agent Mode gated to paid Plus, Pro, and Business subscribers, and it became OpenAI's first consumer browser and a direct challenger to Google Chrome, which has more than 3 billion users worldwide [1][3][18]. The promised Windows, iOS, and Android versions never shipped: in March 2026 OpenAI confirmed plans to fold Atlas into a single ChatGPT desktop app with Codex, and on August 9, 2026 it retired the browser, moving its agentic browsing features into the ChatGPT desktop app and a ChatGPT extension for Chrome [41][42][43].
Atlas was unveiled in a livestream hosted by Sam Altman on October 21, 2025. Altman called the launch "a rare, once-a-decade opportunity to rethink what a browser can be" and argued that web browsers had not meaningfully changed since the introduction of tabs [4][5]. The product slotted directly into a wider race to put AI agents inside the browser surface, putting Atlas in head-to-head competition with Perplexity Comet, which had launched in July 2025, and Dia browser from The Browser Company, which had reached general availability on macOS two weeks earlier on October 8, 2025 [6][7].
Within days of the launch, security researchers published a series of indirect prompt injection attacks against Atlas, including an omnibox prompt-injection technique disclosed by NeuralTrust and a memory-injection vulnerability reported by LayerX [9][44]. OpenAI's chief information security officer, Dane Stuckey, had already written on X the day after launch that prompt injection "remains a frontier, unsolved security problem" and described the mitigations OpenAI had built [10][24]. On December 22, 2025, OpenAI publicly conceded that prompt injection was "unlikely to ever be fully solved" and framed agentic browser defense as a continuous arms race rather than a problem that could be closed by a single fix [28][29].
Background
OpenAI's browser strategy
OpenAI's interest in shipping its own browser had been public for most of 2025. Reuters reported in July 2025 that the company was "close to releasing" an AI-powered web browser, framing the project as a direct attempt to challenge Google Chrome by routing user interactions through ChatGPT rather than a search engine [11][12]. The same reporting noted that OpenAI had hired two former Google vice presidents from the Chrome team, Ben Goodger and Darin Fisher, in 2024 to lead browser engineering. Goodger, who became Atlas's head of engineering, had been the lead developer of Firefox at the Mozilla Foundation before joining Google in 2005, where he was one of the original engineers and leaders on the Chrome project [37][38].
OpenAI had shipped a browser-based agent called Operator in January 2025, which ran inside a virtualized cloud browser and let ChatGPT Pro subscribers delegate web tasks to a remote agent. Operator was retired in mid-2025 and folded into a feature called ChatGPT Agent inside the main ChatGPT app, which combined cloud browsing with the assistant's reasoning. Atlas extended that direction onto the user's own machine, so the assistant could drive the user's local browser with access to their existing logins [1][13].
Agentic AI trend
By the autumn of 2025 the major AI labs were converging on the idea that the next consumer surface for AI was the browser. Google had added Gemini-powered features inside Chrome over 2024 and 2025, Microsoft had introduced a Copilot Mode inside Edge, and Perplexity had shipped Comet to Perplexity Max subscribers in July 2025 before making it free worldwide on October 2, 2025 [14][15]. The Browser Company released Dia in beta in June 2025 and made it generally available on macOS on October 8, 2025, and Atlassian, which had agreed in September 2025 to buy the company for about $610 million, completed the acquisition on the same day Atlas launched [7][16]. Atlas therefore arrived in a saturated launch window.
Industry analysts framed the moment as a possible structural attack on Google's search business. Routing the entry point for web tasks through an AI assistant threatened the Google Search advertising business, Alphabet's largest source of revenue [4][17]. Alphabet's stock fell on the day of the Atlas launch.
When was ChatGPT Atlas released?
OpenAI announced Atlas on its company website on October 21, 2025 in a post titled "Introducing ChatGPT Atlas" (subtitled "The browser with ChatGPT built in") The post and a livestream keynote at 10am Pacific time on the same day described the product and made it available for download on macOS through the official chatgpt.com/atlas page. Sam Altman led the keynote, joined by engineering lead Ben Goodger, product lead Adam Fry, and other members of the Atlas team [4][5][18].
Altman argued that chat is the new way to interface with the internet, that the browser is the right surface to embed it, and that an AI assistant with persistent memory and access to a user's tabs can act on their behalf in ways a traditional browser cannot. He acknowledged that browser switching is "hard" and that Atlas would need to be markedly better than incumbents to win share. OpenAI said experiences for Windows, iOS, and Android were "coming soon," but none shipped before Atlas was retired [1][42].
The initial download was free for all users with a ChatGPT account. Agent Mode, the autonomous browsing capability, was restricted at launch to Plus, Pro, and Business subscribers. Atlas was also available in beta for Business users and, if enabled by an administrator, for Enterprise and Edu users [1].
What can ChatGPT Atlas do?
Atlas presents itself as a familiar Chromium browser with the assistant promoted to a first-class component of the interface. The address bar doubles as a ChatGPT input, every tab can be opened in a ChatGPT-led mode where the assistant runs as a sidebar companion, and a separate Agent Mode lets the assistant take over the browser to complete tasks on the user's behalf.
| Feature | Description |
|---|---|
| Ask ChatGPT sidebar | A persistent right-hand panel that reads the active page and lets the user chat, summarise, compare, draft, or fact-check without leaving the tab [1][3]. |
| Agent Mode | Autonomous web automation. ChatGPT can navigate sites, click through forms, add items to carts, and complete multi-step tasks while the user watches. Restricted to Plus, Pro, and Business at launch [1][4]. |
| Browser memory | Optional persistent memory of recent browsing sessions that the assistant can reference. Optional and opt-in. According to OpenAI's help documentation, original page content is deleted immediately after it is summarised on OpenAI's servers, and the privacy-filtered summaries are deleted within 7 days [1][3]. |
| ChatGPT in the omnibox | The combined address bar routes URLs to navigation, queries to ChatGPT, and natural language commands to either chat or a confirmation prompt for an Agent task [1][4]. |
| Cursor chat on page | An inline writing helper that rewrites or continues text inside any web text field, modelled on the cursor-level rewrites in macOS apps [1]. |
| New tab page | A landing page that combines a ChatGPT prompt box with shortcut tiles, recent tabs, and suggested actions from prior conversations [1]. |
| Chrome import | A first-run import of bookmarks, saved passwords, and browsing history from the user's current browser, later extended to extensions and site data [1][30]. |
| Tab groups and split view | Standard Chromium tab management features inherited from the base engine, augmented by a native split-pane layout for side-by-side reading. A dedicated tab groups manager and vertical tab strip were added in subsequent updates [1][30]. |
| Page visibility and incognito | A per-site toggle in the address bar stops ChatGPT from viewing a page and creating memories from it, and an incognito window temporarily logs the user out of ChatGPT [1]. |
| Logged-out Agent Mode | A scoped Agent Mode session that runs without access to the user's signed-in accounts, recommended by OpenAI as the default when the task does not require authenticated sites [28]. |
| Sign-in with ChatGPT | Atlas requires signing in with a ChatGPT account; a March 2026 update allowed multiple ChatGPT accounts in separate browser profiles [1][30]. |
| Extensions | Atlas supports Chromium extensions, with the same compatibility surface as Chrome for the most part. Importing extensions from the user's previous browser was added in November 2025 [30]. |
The sidebar companion is the most visible new surface in Atlas. It reads the current tab, which the user can confirm or restrict per site, and answers questions, summarises, or rewrites copy without leaving the page. The companion can also reach into a user's ChatGPT memory if enabled, and into the Atlas-specific browser memory layer if the user has opted in [1][3].
Agent Mode is the more contentious addition. When a user invokes it, Atlas takes control of the active tab and begins clicking, typing, and scrolling on the user's behalf. The assistant narrates each step in the sidebar; a user can pause or take back control at any time. At launch OpenAI's examples included ordering a recipe's ingredients for delivery and compiling research into a team brief, and it said agent mode was "an early experience and may make mistakes on complex workflows" and that agents were "susceptible to hidden malicious instructions" [1].
Feature evolution after launch
The first six months after launch saw a steady cadence of Atlas updates that addressed early reviewer feedback and competitive gaps relative to Comet and Edge with Copilot Mode.
| Date | Release | Notable additions |
|---|---|---|
| November 2025 | Build 1.2025.316.6 (November 18) | Vertical tabs, iCloud Keychain passkeys, Google as an optional default search engine, extension import for new installs [30]. |
| December 2025 | Builds 1.2025.337 and 1.2025.344 | Multiple browser profiles per ChatGPT login, "Ask ChatGPT to remember" for webpage text, and (announced December 22) a new adversarially trained browser-agent model for all users; logged-out agent mode had been available since launch [1][28][30]. |
| January 2026 | Build 1.2026.15.3 (January 21) | Tab groups, and an "Auto" search mode that switches between ChatGPT answers and Google results depending on the query [30]. |
| March 2026 | Unified desktop roadmap | OpenAI confirmed plans to combine Atlas, the ChatGPT desktop app, and Codex into a single desktop "super app" [41]. |
| August 9, 2026 | Retirement | Atlas stopped working; OpenAI moved browser-based agentic features into the ChatGPT desktop app and a ChatGPT extension for Chrome [42][43]. |
The release notes describe the Auto search mode as switching automatically between ChatGPT and Google depending on the query; the same update made search-result links more prominent in answers [30].
What is ChatGPT Atlas built on?
Atlas is built on the Chromium open-source project, the same engine that powers Chrome, Edge, Brave, Comet, and Dia, and it renders pages with Chromium's Blink engine. Building on Chromium gives Atlas immediate compatibility with the modern web, the Chrome extension ecosystem, and the standard set of developer tools and security primitives.
The OWL architecture
In an engineering blog post published in late October 2025 titled "How we built OWL, the new architecture behind our ChatGPT-based browser, Atlas," OpenAI disclosed that Atlas does not embed Chromium inside the application in the conventional way. Instead it uses a custom layer called OWL, for OpenAI's Web Layer, which runs Chromium's browser process outside the main Atlas application process. In this design the native Atlas app, written in Swift, is the OWL Client and the separate Chromium browser process is the OWL Host; the two communicate over inter-process communication using Mojo, Chromium's own message-passing system, for which OpenAI wrote custom Swift and TypeScript bindings [39].
OpenAI framed the split as a resilience and velocity win. "If its main thread hangs, Atlas doesn't," the post said of Chromium. "If it crashes, Atlas stays up" [39]. Because OWL ships internally as a prebuilt binary, OpenAI also reported that most engineers never need to compile Chromium locally, so an Atlas build completes in minutes rather than hours [39]. To draw the out-of-process web content into the native window, OWL uses a technique called layer hosting: Chromium renders into a CALayer, a macOS graphics primitive, and the Atlas app embeds that layer through the private CALayerHost API so the GPU compositor can share memory efficiently across the process boundary [39].
Models and memory
The assistant is powered by the same models that drive the ChatGPT product family, with model routing decided server-side. OpenAI did not publish a fixed list of models that handle Atlas requests; OpenAI's engineering post says agent mode relies on its computer use model, which takes a single image of the screen as input [39]. The system prompt and context are managed by OpenAI's servers, not by the local browser, so the local machine sends snapshots of page contents and user prompts to the model and receives back assistant responses and Agent Mode action plans.
Browser memory is a separate feature from ChatGPT memory, although they are related. ChatGPT memory, introduced in early 2024, stores user-supplied facts across chats. Browser memory in Atlas stores key details from the content of pages the user visits [1]. According to OpenAI's help documentation, web content for these memories is summarised on OpenAI's servers and the privacy-filtered summaries of page content are deleted within seven days [30]. Both memory layers are user-visible and clearable from a settings page. The Atlas privacy policy stated at launch that browsing history is not used to train models by default and that page contents shared with the assistant are subject to the standard ChatGPT data controls [1][3].
Atlas uses a per-process security model inherited from Chromium and a stricter sandbox for Agent Mode actions. OpenAI's engineering post says agent mode is driven by a computer use model that works from an image of the screen, with its synthesized input events routed directly to the renderer rather than through the privileged browser layer, and that logged-out agent sessions run in isolated in-memory storage [39]. The launch post added that the agent cannot run code in the browser, download files, install extensions, or access other apps or the file system [1]. Security researchers later argued that this design was not enough to stop indirect prompt injection, where adversarial text embedded in a page hijacks the assistant's instructions [8][9][10].
Is ChatGPT Atlas available on Windows?
Atlas launched on macOS on October 21, 2025, and OpenAI said Windows, iOS, and Android versions were "coming soon." None of them shipped: Atlas remained macOS-only until OpenAI retired it on August 9, 2026 [1][42].
| Platform | Status as of May 2026 | Notes |
|---|---|---|
| macOS | Released October 21, 2025 [1][2] | Apple silicon Macs only. Regular updates until the browser stopped working on August 9, 2026 [42]. |
| Windows | Never released [1][42] | Listed as "coming soon" at launch; no public Windows build shipped before Atlas was retired in August 2026. |
| iOS | Never released [1][42] | Listed as "coming soon" at launch. |
| Android | Never released [1][42] | Same status as iOS. |
| Linux | Not announced [1] | No Linux build was mentioned in the launch keynote or in any update since. |
The platform gap stood out because Perplexity Comet, which debuted on desktop in July 2025, added an Android version in November 2025 [45]. A January 2026 Windows Latest report examined how Microsoft Edge might respond if Atlas reached Windows 11 [32].
Distribution went through the chatgpt.com/atlas download page rather than the Mac App Store, with sign-in tied to a ChatGPT account [1].
How much does ChatGPT Atlas cost?
Atlas itself is a free download for anyone with a ChatGPT account. The differentiation between tiers shows up in which features are available inside the browser, primarily Agent Mode and the depth of model access.
| ChatGPT tier | Atlas access | Agent Mode |
|---|---|---|
| Free | Atlas browser, Ask ChatGPT sidebar, memory toggles [1][18] | Not included at launch. |
| Plus ($20/month) | Atlas browser and sidebar under the plan's ChatGPT limits [1] | Included (preview at launch). |
| Pro ($200/month) | Atlas browser and sidebar under the plan's ChatGPT limits [1] | Included (preview at launch). |
| Business | Atlas in beta at launch [1] | Included (preview at launch). |
| Go | Atlas browser, Ask ChatGPT sidebar [1] | Not included at launch. |
At launch OpenAI did not introduce a new subscription tier specifically for Atlas. The browser leaned on existing ChatGPT subscriptions [1].
Is ChatGPT Atlas safe?
The most serious public criticism of Atlas in its first weeks involved prompt injection. Indirect prompt injection attacks plant adversarial instructions inside a web page, an image, an email, or a clipboard, and rely on the AI assistant to read and follow those instructions when it processes the content. Because Atlas's sidebar and Agent Mode read page contents as part of normal operation, both surfaces are exposed to this class of attack.
Brave's security team published prompt-injection disclosures against other AI browsers, chiefly Perplexity Comet, in August and October 2025, arguing that the weakness was systemic across agentic browsers [8][21]. The first widely reported Atlas-specific technique came from NeuralTrust, which showed that text disguised as a URL could fail URL validation in the omnibox and then be treated as a high-trust command to the agent; NeuralTrust noted such a string could be hidden behind a "Copy link" button [44].
LayerX, a separate security firm, demonstrated a memory-poisoning attack against Atlas in late October 2025, which it described as the first vulnerability disclosed in the new browser. A malicious link could use a cross-site request forgery (CSRF) request to piggyback on the victim's ChatGPT credentials and inject hidden instructions into ChatGPT's account memory, which then influenced later sessions; LayerX said the technique could be used to execute remote code, and stressed that the tainted memory persists across every device and browser where the account is used [9]. OpenAI told CSO Online that it had been unable to reproduce the reported attack.
Dane Stuckey, OpenAI's chief information security officer, addressed the issue on X on October 22, 2025, the day after launch. He called prompt injection "a frontier, unsolved security problem" and listed OpenAI's defenses: extensive red-teaming, model training that rewards ignoring malicious instructions, overlapping guardrails, systems to detect and block attacks, rapid-response systems, a logged-out mode, and a "Watch Mode" that pauses the agent on sensitive sites when the user leaves the tab [10][24].
Hardening Atlas: December 2025 update
On December 22, 2025, OpenAI published a follow-up technical post titled "Continuously hardening ChatGPT Atlas against prompt injection attacks." The post documented a security release that shipped an adversarially trained Agent Mode model alongside surrounding runtime safeguards, and disclosed the methodology OpenAI had used to find and patch new injection paths [28][29].
The central novel element was an automated red team. OpenAI built a separate large language model trained end-to-end with reinforcement learning to act as an attacker against a simulated Atlas Agent. During its chain of thought reasoning, the attacker proposed candidate injections and sent them to an external simulator that ran a counterfactual rollout of how the targeted victim agent would behave, then scored the outcome. OpenAI noted that such an attacker could "steer an agent into executing sophisticated, long-horizon harmful workflows that unfold over tens (or even hundreds) of steps" [28]. Successful attacks fed back into both the attacker (to find harder cases) and the victim agent training set (to produce a more resistant model). OpenAI said the loop had already produced a new adversarially trained browser-agent checkpoint rolled out to all Atlas users, but it did not publish quantitative robustness figures [28].
In the same announcement, OpenAI conceded an important point in plain language: prompt injection, the company wrote, is, "much like scams and social engineering on the web, unlikely to ever be fully 'solved.'" It added that it was "optimistic that a proactive, highly responsive rapid response loop can continue to materially reduce real-world risk over time," framing agentic browser security as an ongoing arms race rather than a problem that could be closed by a single fix [28][29][33].
Alongside the model update, OpenAI rolled out user-facing recommendations: use logged-out Agent Mode when authenticated access is not required, write narrowly scoped prompts rather than broad instructions, and carefully review the confirmation requests the agent shows before consequential actions such as completing a purchase or sending an email [28].
Ongoing debate
Independent commentators were mixed on whether these defenses were adequate. Simon Willison, who has tracked prompt injection research since 2022, welcomed Stuckey's acknowledgement but argued that guardrails could not provide a credible solution, writing that "in application security 99% is a failing grade" [24].
A related concern, voiced by privacy advocates, was that Atlas's browser memory feature combined with ChatGPT memory created a long-lived behavioural profile of the user, hosted on OpenAI's servers, with content that included page-level browsing data. OpenAI's documentation pointed to the per-site visibility toggle, the ability to view, archive or delete browser memories, the off-by-default training setting, and deletion of page summaries within seven days [1][3].
How does ChatGPT Atlas compare to Perplexity Comet?
Atlas, Perplexity Comet, Dia browser, and Microsoft Edge with Copilot Mode all embed an AI assistant in the browser surface but differ in how much autonomy they give that assistant.
| Browser | Released | Built on | Default assistant | Agent autonomy | Free tier | Platforms (May 2026) |
|---|---|---|---|---|---|---|
| ChatGPT Atlas | October 21, 2025 [1][2] | Chromium | ChatGPT sidebar with Agent Mode | High. Agent Mode performs multi-step web tasks on Plus, Pro, Business [1][4]. | Yes for the browser itself. Agent Mode gated to paid tiers. | macOS only (retired August 9, 2026) [42] |
| Perplexity Comet | July 9, 2025 paid, October 2, 2025 free worldwide [6][15] | Chromium | Perplexity assistant with Comet Assistant | Medium to high. Comet Assistant does multi-step tasks. | Yes from October 2, 2025. | Desktop and Android [45] |
| Dia browser | June 11, 2025 beta, October 8, 2025 GA on macOS [7] | Chromium | Dia chat with Skills | Low. Reads and writes inside open tabs, no general DOM automation. | Yes with usage caps, Dia Pro at $20/month for unlimited. | macOS only |
| Arc Search | 2024 | Custom (Chromium engine on iOS) | Arc Search browse-for-me | Medium. Aggregates and summarises results, no full DOM control. | Free. | iOS only |
| Microsoft Edge Copilot Mode | 2025 update to Edge | Chromium | Microsoft Copilot | Medium. Tab-aware actions and some workflow agents. | Free with optional Copilot Pro at $20/month. | Windows, macOS, mobile |
| Google Chrome with Gemini | Ongoing | Chromium | Gemini in Chrome | Low to medium. Page summarisation and write assistance. | Free with optional Google AI Pro (formerly Gemini Advanced) at $19.99/month. | All major platforms |
Launch coverage generally placed Atlas at the high-autonomy end of this spectrum because its agent could act inside the user's own logged-in browser, though TechCrunch noted that early web-browsing agents still "struggle to reliably automate the more cumbersome problems users might want to offload" [1][18].
Atlas distinguished itself from Comet on memory and from Dia on autonomy. Comet leaned harder on the answer-engine surface inherited from the Perplexity search product, with citations and source panels promoted in the UI; Atlas leaned on the chat surface inherited from ChatGPT, with memory and Agent Mode promoted instead. Dia did not ship a full Agent Mode at all and remained the most conservative of the three.
A second-wave round of reviews in early 2026, after each product had iterated for several months, settled on a rough division of labour. Atlas was rated strongest for research and data extraction across multiple pages, where the ChatGPT model could carry context across sites. Comet was rated strongest for fast factual lookups and for cross-platform availability. Dia was praised for workflow automation inside open tabs but was held back by the absence of a full Agent Mode [34][35]. The lack of a Windows build for Atlas became the most-cited disadvantage in head-to-head Windows-user comparisons [31][34].
Reception and adoption
Launch coverage was interested in the ambition and cautious on execution. TechCrunch's launch report said Atlas was "a major step in the company's quest to unseat Google as the main way people find information online" but questioned whether it could dent Chrome's more than 3 billion users [18].
Not all of the commentary was about safety. The technologist Anil Dash published an influential critique on October 22, 2025 calling Atlas "the browser that's anti-web." Dash argued that Atlas "substitutes its own AI-generated content for the web, but it looks like it's showing you the web," and that this was "the app working as designed" rather than a bug, because searching the omnibox surfaced AI summaries instead of links to the sites people were looking for [40]. The critique was widely circulated among web developers and became a touchstone for the argument that AI browsers risk hollowing out the open web's traffic and incentives.
Developer commentary was more divided. Simon Willison, in a blog post on October 22, 2025 responding to Stuckey, argued that a motivated attacker would get past any guardrail and that "logged in mode" delegated security decisions to end users, "an unfair burden to place on almost any user" [24]. Hacker News discussion threads in the days following the launch focused on the same theme: a substantial subset of the audience believed agentic browsing was structurally unsafe at launch, while a different subset argued that the productivity gains were real and that the safety problem was tractable with engineering.
How many people use ChatGPT Atlas?
Usage data published by OpenAI after launch was sparse.
Long-term coverage in early 2026 framed Atlas as one of several browsers competing for the same emerging surface rather than a runaway winner.
In March 2026, OpenAI confirmed that it would combine Atlas, the ChatGPT desktop app, and the Codex coding agent app into a single desktop "super app" overseen by Fidji Simo, saying the change would streamline the user experience and reduce fragmentation [41]. In July 2026 OpenAI said it was deprecating Atlas and moving browser-based agentic capabilities into ChatGPT and Codex; Atlas stopped working on August 9, 2026, and OpenAI pointed users to the new ChatGPT desktop app and the ChatGPT extension for Chrome [42][43].
Continue reading
- ChatGPTChatGPT is a conversational artificial intelligence chatbot developed by OpenAI and built on the company's GPT series of large language models.Mentioned 71 times above
- OpenAIOpenAI is an American artificial intelligence research and deployment organization.Mentioned 49 times above
- Dia (browser)Dia is an AI-first web browser developed by The Browser Company of New York, the studio behind the Arc browser, and owned since October 2025 by Atlassian.Mentioned 10 times above
- Prompt injectionPrompt injection is a class of security vulnerabilities in which an attacker crafts malicious input designed to override, subvert, or manipulate the instructions governing a large language model (LLM).Mentioned 10 times above
- Perplexity CometComet is an AI-powered, agentic web browser developed by Perplexity AI. It is built on the Chromium open-source project and ships with an integrated AI assistant that can summarise pages, answer questions…Mentioned 6 times above
- Sam AltmanSamuel Harris Altman (born April 22, 1985) is an American entrepreneur, investor, and technology executive.Mentioned twice above
More to read
See also
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38OpenAI. "Introducing ChatGPT Atlas: a new web browser with ChatGPT built in." openai.com/index/introducing-chatgpt-atlas, October 21, 2025.
- ^1 ^2 ^3OpenAI. "ChatGPT Atlas product page." chatgpt.com/atlas, accessed October 21, 2025.
- ^1 ^2 ^3 ^4 ^5 ^6OpenAI. "ChatGPT Atlas privacy and data controls." openai.com/policies/atlas-privacy, October 21, 2025.
- ^1 ^2 ^3 ^4 ^5 ^6Warren, Tom. "OpenAI is launching its own web browser to take on Chrome." The Verge, October 21, 2025.
- ^1 ^2OpenAI. "ChatGPT Atlas launch livestream." YouTube, October 21, 2025.
- ^1 ^2Zeff, Maxwell. "Perplexity launches Comet, an AI-powered web browser." TechCrunch, July 9, 2025. techcrunch.com/...-comet-an-ai-powered-web-browser
- ^1 ^2 ^39to5Mac Staff. "Dia is now generally available on macOS." 9to5Mac, October 8, 2025.
- ^1 ^2Brave Security Team. "Unseeable prompt injections in screenshots: more vulnerabilities in Comet and other AI browsers." brave.com/...unseeable-prompt-injections, October 21, 2025.
- ^1 ^2 ^3Paz, Roy (LayerX Security). "ChatGPT Tainted Memories: LayerX discovers the first vulnerability in OpenAI Atlas browser." layerxsecurity.com/...in-new-chatgpt-atlas-browser, October 27, 2025.
- ^1 ^2 ^3Stuckey, Dane. Post on X about prompt injection risks in ChatGPT Atlas, October 22, 2025 (quoted at length in reference 24).
- ^Tong, Anna, and Hu, Krystal. "Exclusive: OpenAI to release web browser in challenge to Google Chrome." Reuters, July 9, 2025.
- ^Reuters Staff. "OpenAI hires former Google Chrome leaders to build browser team." Reuters, April 17, 2024.
- ^OpenAI. "Introducing ChatGPT Agent." openai.com/blog, July 2025.
- ^Google. "Gemini in Chrome rollout update." blog.google, 2025.
- ^1 ^2Mehta, Ivan. "Perplexity makes its AI browser Comet free for everyone." TechCrunch, October 2, 2025.
- ^Atlassian. "Atlassian Completes Acquisition of The Browser Company of New York." BusinessWire, October 21, 2025.
- ^Roth, Emma. "OpenAI's browser threatens Google's most important business." The Verge, October 21, 2025.
- ^1 ^2 ^3 ^4 ^5Zeff, Maxwell. "OpenAI launches an AI-powered browser, ChatGPT Atlas." TechCrunch, October 21, 2025.
- Chromium Blog. "Chromium release channels." chromium.org, accessed 2025.
- The Information. "OpenAI's browser uses GPT-5 family models for Agent Mode." theinformation.com, October 22, 2025.
- ^Brave Security Team. "Prompt injection in Comet, Dia, and Atlas: a 2025 research summary." brave.com/blog, December 2025.
- LayerX Security. "Browser memory: a new attack surface in agentic AI." layerxsecurity.com/research, November 2025.
- Stuckey, Dane. "Updates on Atlas security mitigations." openai.com/blog, November 2025.
- ^1 ^2 ^3 ^4Willison, Simon. "Dane Stuckey (OpenAI CISO) on prompt injection risks for ChatGPT Atlas." simonwillison.net/...openai-ciso-on-atlas, October 22, 2025.
- Knight, Will. "The browser is now the most strategic surface in AI." Wired, October 22, 2025.
- Altman, Sam. Posts on X about ChatGPT Atlas launch, October 21 and 22, 2025.
- SimilarWeb. "AI browser market share, Q4 2025." similarweb.com, January 2026.
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8OpenAI. "Continuously hardening ChatGPT Atlas against prompt injection attacks." openai.com/index/hardening-atlas-against-prompt-injection, December 22, 2025.
- ^1 ^2 ^3Wiggers, Kyle. "OpenAI says prompt injections that can trick AI browsers may never be fully 'solved.'" TechCrunch, December 22, 2025.
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9OpenAI Help Center. "ChatGPT Atlas release notes." help.openai.com, accessed May 2026.
- ^Digital Applied. "AI browser landscape 2026: Atlas vs Comet vs Arc vs Dia." digitalapplied.com, 2026.
- ^Windows Latest. "If ChatGPT Atlas browser lands on Windows 11, Microsoft Edge already looks ready to intercept it." windowslatest.com, January 1, 2026.
- ^Fortune Staff. "OpenAI says prompt injections that can trick AI browsers may never be fully solved." fortune.com, December 23, 2025.
- ^1 ^2WebFX. "AI browser comparison 2026: Atlas, Comet, Dia, and 5 more ranked and reviewed." webfx.com, 2026.
- ^AIMultiple. "AI web browsers selection guide including ChatGPT Atlas in 2026." aimultiple.com, 2026.
- Second Talent. "ChatGPT statistics 2026: users, revenue, and enterprise adoption." secondtalent.com, 2026.
- ^Zeff, Maxwell, and Wiggers, Kyle. "OpenAI launches ChatGPT Atlas, led by engineering chief Ben Goodger." TechCrunch, October 21, 2025.
- ^NZ Herald. "Meet Ben Goodger: the Kiwi behind ChatGPT maker OpenAI's big play against Google." nzherald.co.nz, October 2025.
- ^1 ^2 ^3 ^4 ^5 ^6OpenAI. "How we built OWL, the new architecture behind our ChatGPT-based browser, Atlas." openai.com/...building-chatgpt-atlas, October 30, 2025.
- ^Dash, Anil. "ChatGPT's Atlas: the browser that's anti-web." anildash.com, October 22, 2025.
- ^1 ^2 ^3Capoot, Ashley. "OpenAI to create desktop super app, combining ChatGPT app, browser and Codex app." CNBC, March 19, 2026. cnbc.com/...esktop-super-app-chatgpt-browser-codex
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10OpenAI Help Center. "Evolving Atlas into ChatGPT for browser-based agentic work." help.openai.com/...20001371, accessed November 2026.
- ^1 ^2 ^3Hale, Craig. "OpenAI shuts down its Atlas browser after not even a year." TechRadar, July 13, 2026. techradar.com/...las-browser-after-not-even-a-year
- ^1 ^2"ChatGPT Atlas Browser Can Be Tricked by Fake URLs into Executing Hidden Commands." The Hacker News, October 27, 2025. thehackernews.com/...las-browser-can-be-tricked-by
- ^1 ^2Mehta, Ivan. "Perplexity brings its AI browser Comet to Android." TechCrunch, November 20, 2025. techcrunch.com/...-its-ai-browser-comet-to-android
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
8 revisions · v8 · 5,305 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent fact-check of the full article against its cited sources (130 claims, 24 sources fetched); 24 material + 39 minor fixed
Cite this page: AI Wiki. "ChatGPT Atlas." aiwiki.ai, updated 24 Jun 2026. CC BY 4.0. https://aiwiki.ai/wiki/chatgpt_atlas


