OpenAI and Perplexity have both shipped browsers this year, Atlas and Comet, opening what’s being framed as an AI browser war, and both companies are pitching them in the same grandiose register. Sam Altman called the browser moment “a rare once-a-decade opportunity to rethink what a browser can be about.” OpenAI’s Fidji Simo has described Atlas and ChatGPT becoming “the operating system for your life.” Perplexity’s Aravind Srinivas calls Comet a “cognitive operating system,” and a colleague of his has gone further, branding it the “operating system of your mind.”
None of that holds up. A browser is not an operating system, any more than a chat interface is one. Chrome, Edge, and Safari matter because each is welded to a much larger stack of proprietary tools, accounts, and workflows tied to a specific operating system. That deep integration, not the browser chrome itself, is what actually locks people in. Atlas and Comet don’t have it, and copying the interface doesn’t create it.
A textbook case of survivorship bias
There’s a useful analogy from World War II. The U.S. military asked statisticians to figure out where to add armor to bombers, based on the pattern of bullet holes on planes that returned from missions. The obvious answer was to reinforce the areas most heavily hit. Mathematician Abraham Wald pointed out the flaw: those planes survived their damage. The areas with no bullet holes on the survivors were the areas that, when hit, brought the plane down. The military was studying the wrong dataset entirely.
It’s the same error behind every “dropouts become billionaires” argument and every self-help book that reverse-engineers success from a handful of morning routines. OpenAI and Perplexity appear to have looked at Google, Microsoft, and Apple, noticed that each one owns a browser, and concluded that owning a browser is a meaningful ingredient of platform dominance. It’s a plausible-sounding correlation built on a tiny, self-selected sample.
Plan B: if you can’t buy Chrome, clone it
This isn’t a coincidence of timing either. While antitrust proceedings against Google were underway, both companies floated interest in acquiring Chrome outright, Perplexity went as far as submitting an unsolicited $34.5 billion bid. The court declined to force a sale. Atlas and Comet, both built on Chromium, both replicating Chrome’s tab layout, address bar, and extension model, look like the fallback plan that had been in development the whole time.
That’s not unusual on its own. Microsoft built Edge the same way, and even non-Chromium browsers like Safari and Firefox have absorbed features Chrome popularized first. What’s different here is the layer bolted on top: agentic browsing, where the browser clicks, reads, and completes tasks on your behalf, from booking travel to filling a shopping cart, while you do something else.
Agentic browsing mostly makes something already invisible, visible
The underlying capability isn’t new. Every time you prompt ChatGPT or Perplexity and it goes and finds information to answer you, a headless browser is already doing that work behind the scenes. What Atlas and Comet add is a window onto that process, you can now watch the browser move between pages, restructure a spreadsheet, or add items to a cart, instead of just receiving the finished answer.
Whether that visibility is a feature most people want is a separate question. For anyone who still enjoys reading the actual article instead of a summary of it, or genuinely likes browsing and stumbling across things, watching an agent execute clicks on your behalf isn’t obviously more useful. Realistically, most users will tab away and wait for a completion notification rather than watch the process unfold.
The measurement and security problems these browsers create
Agentic browsing creates real headaches for anyone running a website. When a headless bot crawls your site directly, it typically identifies itself, PerplexityBot, for example. When Comet visits a site on a user’s behalf, it shows up in your analytics looking like an ordinary Chromium session with a normal IP address. You often can’t tell whether a human or an agent is on the other end, which quietly undermines the traffic and click metrics teams have relied on for years. That same ambiguity opens the door to ad fraud: if agents can be served ads indistinguishable from human traffic, and can generate thousands of requests per second, the incentive to exploit that gap is obvious.
The security exposure is worse. Analysts at Gartner have advised organizations to block AI browsers outright over cybersecurity concerns, prompt injection chief among them. Researchers at LayerX Security demonstrated an attack they named CometJacking, where clicking a single malicious link triggers hidden instructions that direct Comet’s AI to exfiltrate sensitive data the browser already has authorized access to, no phishing or stolen credentials required. Weeks later, the same researchers found a comparable vulnerability in Atlas, capable of injecting malicious instructions directly into ChatGPT’s memory. Handing an agentic browser a task like booking travel or editing a document currently feels less like delegating to an assistant and more like handing a stranger your keys. Separately, Amazon has already gone to court over Comet’s behavior on its platform, and the outcome of the related CFAA case over whether agents even have a legal right to visit a site the way a human browser does will shape how much latitude these tools get going forward.
Where these browsers might actually earn a place
None of this means Atlas and Comet are dead ends. Developers and QA teams could use them to simulate real user journeys at scale, testing how a site behaves across scenarios far faster than manual testing allows. SEO teams have a more specific use case: watching how an agentic browser interprets a page’s structure is a reasonable proxy for how the same company’s hidden headless crawler “reads” your content, which can surface gaps worth fixing before they show up as missed citations.
Technical users comfortable with the risk profile will also find genuine automation value, pulling data from scattered sources into a spreadsheet, or monitoring a set of pages for changes, without writing a script. But there’s an irony in exactly who ends up finding these browsers useful. OpenAI and Perplexity need mainstream consumer behavior data to train and monetize their models. The users most likely to actually adopt these tools right now are developers and power users running structured, repetitive workflows, precisely the audience least representative of typical consumer browsing.
Why the incumbents are still favored to win
For Atlas or Comet to pull share from Chrome, Safari, Edge, or Firefox, they need a value proposition compelling enough to outweigh the security risk and the total absence of the ecosystem lock-in that makes the incumbents sticky. As Chrome and Edge keep folding their own AI features directly into the existing browser, that gap keeps closing rather than widening. If AI meaningfully changes how people browse, it’s more likely to happen inside Chrome or Edge than inside a standalone challenger, because Google and Microsoft already have both the distribution and their own frontier models. Firefox is worth watching too: its newer AI controls, which let users choose which AI features to enable rather than having them switched on by default, may end up revealing what people actually want from AI in a browser more clearly than either Atlas or Comet has so far.
The AI browser war has barely started, and it isn’t going to resolve on the strength of a launch demo. The eventual winners will be decided by user experience and a clear value proposition, not by borrowed language about operating systems for your mind. If agentic browsing succeeds at all, it will be because it disappears so completely into existing workflows that nobody talks about it as a separate category anymore.
Frequently asked questions
What is agentic browsing?
It’s a browser mode where the AI navigates, clicks, and completes multi-step tasks, like booking travel or shopping, on a user’s behalf, instead of the user doing that navigation manually.
Are Atlas and Comet built on Chrome?
Both are built on Chromium, the open-source engine behind Chrome, and closely replicate Chrome’s interface: tabs, address bar, and extension support.
Can website analytics tell the difference between a human and an agentic browser visit?
Usually not. Unlike a headless crawler that identifies itself by user agent, an agentic browser visiting a site on a user’s behalf typically appears as an ordinary Chromium session with a normal IP address, which is part of why the agentic web is complicating traditional traffic measurement.
Are AI browsers a security risk?
Security researchers have documented real attacks, including a prompt-injection technique against Comet dubbed CometJacking and a related vulnerability found in Atlas, and Gartner has advised organizations to consider blocking AI browsers over these concerns.