The Genius Browser That Eats Your CPU for Breakfast

A review of Atlas—OpenAI’s AI browser

On October 21 OpenAI unveiled the Atlas browser with a built-in chatbot and an AI assistant capable of interacting with websites on the user’s behalf.

ChatGPT Plus subscribers can use all browser tools at no extra cost. For now, it is available only on macOS.

ForkLog tested Atlas and reviewed early feedback.

Interface and usability

You can download the browser from the company’s official site. It will prompt you to sign in and import settings from Chrome or Safari. You can skip the latter.

Because Atlas is built on Chromium, migration is painless—history, bookmarks, saved passwords and even open tabs can be imported. The interface is minimalist.

Navigation and past chats live on the left; settings sit in the top-right. A nice touch: a slider lets you change the browser’s colour.

The home screen largely mirrors ChatGPT’s page design. A large prompt field sits in the centre alongside a list of AI models.

By default, search routes through the chatbot. You can switch to Google if needed: open a new tab, enter a query and Atlas will offer the alternative.

The lack of an explicit default-search setting irks at first, but after a couple of hours you get used to choosing manually each time.

OpenAI’s browser supports any extensions from the Chrome Web Store, which will delight many crypto users.

That partly offsets a shortcoming: Atlas lacks a built-in ad blocker, so you will need AdBlock. It also has no integration with 1Password or other password managers—an unwelcome surprise for some.

Atlas matches other browsers for speed and smoothness. Page loads are standard. One quirk, however, is high power draw. Although it uses no more RAM than Chrome, users report it “gobbles” about 1% of battery for every two minutes of activity.

One user also saw the browser crash twice without obvious cause.

A more serious flaw is the lack of profiles. Browsing and settings tie to a single ChatGPT account, mixing work and personal queries. Separate work and personal profiles would fix this.

Chat assistant

Atlas’s signature feature—the one that earns it the “AI browser” moniker—is deep ChatGPT integration throughout web surfing. The assistant is always at hand: click “Ask ChatGPT” on the right sidebar. It immediately understands the page context and answers accordingly.

All of ChatGPT’s capabilities are available in-browser: ask about the current tab, get a summary of a long text or a plain-English explanation of a knotty concept, translate a passage—without third-party tools.

Before long your search style shifts: instead of typing keywords, you write full questions.

“Previously, Chrome made web surfing fast, and Atlas makes it smart,” — rightly noted one user.

Using Atlas underscores how much time ordinary browsers waste on copying links or hunting through long articles for the right passage.

Memory

A memory feature lets the browser remember which sites and topics you have viewed, and it factors these into replies. Atlas can retain details from previously visited pages and tailor results to personal interests.

For instance, if you accidentally close several tabs with film lists and later ask for movie recommendations, the AI will consider the pages you visited and suggest options that fit your tastes.

Another use: open several product pages and ask ChatGPT to compare them.

Memory is undeniably a meaningful upgrade in personalisation. Yet pushing this too far could mark the end of classic open search, with results increasingly filtered.

Some users also found that enabling memory “dumbs down” the model: it starts making odd assumptions about the interlocutor and slips into an overly personalised tone.

“Atlas gets about two to three times smarter the second you turn off memories,” — noted the user.

Others disable it over perceived bias—without it, answers are supposedly more candid and precise.

The effect likely depends on how you use the browser. One safe conclusion: treat it with caution.

AI agent

Another pillar of Atlas is its AI agent. A special mode lets it execute user commands autonomously—reserve tickets, place orders, fill forms—while you watch it click through sites.

As the video shows, it thinks for a long time: a human would check out far faster. Atlas completed the task up to the point of entering personal data. From there you must proceed manually—and do so every time.

Some bloggers shared their tests. Atlas rarely sees tasks through. The much-touted automatic purchasing works only for Walmart in the United States; elsewhere the AI trips up. Booking a table is tricky, too. In some cases it confirmed a reservation that turned out to be a hallucination.

Such tales spawned jokes that the agent behaves like an intern who requires constant supervision. The potential, though, is considerable.

Privacy and security

Atlas strives for transparency: settings let you toggle memory and control what the browser knows about you.

OpenAI says it does not use browsing history to train its models. A privacy panel shows which data and sites ChatGPT is accessing.

Even so, experts remain sceptical about security. The risks look uncomfortably high.

Developer Simon Willison underlined that it is unclear how AI browsers are protected against an attack called “prompt injection”, so it is too early to trust such tools.

“At the moment the defence appears to amount to the user having to watch the agent mode’s actions closely the entire time. […] I certainly will not trust any of these products until a group of security researchers gives them a good shakedown,” Willison wrote.

Subsequently, OpenAI’s chief information security officer, Dane Stuckey, outlined measures against “prompt injection”. He stressed that the problem is important and not fully solved. The startup ran extensive tests and trained models to ignore malicious prompts. It implemented multilayered guardrails and attack detectors, plus a rapid response system.

Protection modes:

• Logged-out Mode — the agent acts on the user’s behalf but without access to accounts or passwords;
• Logged-in Mode — has access only for “very trusted sites and narrow tasks”;
• Watch Mode — if the assistant operates on an important site, such as a bank, the user must watch its actions. Leaving the tab pauses the process.

Willison argued that requiring constant oversight is too flimsy a guarantee of reliability.

On data privacy, scepticism stems from the browser’s tight coupling to cloud AI. In essence, a significant share of user activity is sent to OpenAI’s servers for analysis. Even if data are stored locally and not used to train models, the mere fact that the browser remembers the user’s reasoning is unsettling.

For ordinary web surfing Atlas is fine, but logging into a bank, conducting DeFi transactions via a primary crypto wallet or importing private keys into MetaMask looks questionable.

There is also a question of attention steering. If news feeds can shape opinions by curating content, could Atlas do something similar—especially as it personalises over time?

Worth a try

Has Atlas’s debut been a revolution like ChatGPT’s was? Unlikely.

AI browsers predate it—Dia and Opera’s offerings, for instance. Atlas brings nothing fundamentally new. The core ideas—chatbot in lieu of search, an on-page helper, an automation agent—are already implemented by rivals.

Even so, native ChatGPT integration does set OpenAI’s effort apart. If the chatbot is already woven into a user’s routine, moving to Atlas is seamless. Right after installation the browser gets access to chat history, personal settings and any paid subscription privileges (if any).

Atlas’s target audience comprises heavy AI users, tinkerers and professionals. For the mass market it may feel raw and unfamiliar—for example, returning an opaque response to a query like “youtube” instead of jumping straight to YouTube. And search via a chatbot takes several seconds longer than traditional Googling, where responses are near-instant.