OpenAI's Atlas AI browser tested across 12+ real-world tasks by Nate B Jones
Solo presenter Nate B Jones reviews OpenAI's new Atlas browser.
Summary
Nate B Jones tests OpenAI's Atlas browser — an AI-enabled browser featuring a sidebar chat assistant — across more than twelve real-world tasks, including PowerPoint creation, booking a yoga class, folder automation, writing assistance, and spreadsheet calculations. He finds the browser meaningfully better than OpenAI's previous web browsing efforts but not yet a Chrome replacement, awarding it a grade of C+ to B−. He identifies a clear split between tasks where Atlas excels (low-ambiguity, linear, repetitive web work) and tasks where it struggles (aesthetically complex or high-ambiguity tasks). He also raises a significant security concern around prompt injection attacks, arguing that the industry lacks a standardized browser safety framework comparable to model safety cards.
Key Takeaways
FULL TRANSCRIPT
What OpenAI Launched
Nate B Jones: OpenAI heated up the browser wars by launching their AI-enabled browser called Atlas. I'm going to get into what they launched, what the promise was, and where I think it works and where I think it doesn't.
First off, what did they launch? It is a browser that looks a lot like Chrome or any other browser you might use, except it has a chat assistant in the sidebar. In this sense, it's very similar to other smart browsers that have already launched. The Comet browser from Perplexity comes to mind — exactly the same idea. You just launch it in the sidebar, you do your task with the chat, and then you have the main browser pane just the way it always is.
Live Demo: PowerPoint Creation
I can actually show you what that looks like. Here we are in the browser. I'm actually working on a presentation about the browser — I realize that's super meta — but you can see I have a little chat here, and you can interact with the chat. In this case, what I had the agent do is help me create a PowerPoint presentation about the agent itself.
Just to be transparent: it was able to lay out the styling effectively — it got the green and the blue highlight really well. It was able to lay out the title. It was able to lay out the dark background look, very professional. It was also able to take and expand the copy really effectively. Where it didn't work as well was it struggled with some of the details of formatting. I know this doesn't look like fancy formatting, but just to be transparent, getting it to do white text on a dark background was not something it was super good at. But that's a minor nitpick — I don't want to dwell on that.
The Larger Opportunity
I want to talk about the larger theme here. I wanted to try something that was a real-life task, just to play with it and see how it worked. And I have to say, this is closer to useful browser utility with AI than I've had in a while. That's what makes me bullish on the trajectory here. I don't know that this browser by itself today is knock-it-out-of-the-water incredible, or that it's going to dethrone Chrome. But I see the trajectory that this team is shipping on, and I'm super interested in where they go next. I do think there are some really interesting opportunities in the way this browser modality works — the way it interacts.
For example, if I say "please adjust the font size on this slide," it's going to take that as a task and actually start to do something. You can see it start to organize, the sparkles turn on, and it actually does something with my screen while I go and do other work. So if I go and look at other things — say I go check my LinkedIn — I can come back to my slide deck later and see that it has continued to work.
Ironically, what the slide deck is showing is one of the things I think is most useful here. Think about the boring work that you do on the web. Things like automating folder creation — that takes a lot of time. You can do that work much more easily if you just open a tab and ask this browser to do it.
Where Atlas Works Best
Similarly, you can think about the work you would do as a writing coach. You can see it has these issues — it actually made things worse in some cases — and this is why I don't want to overpromise. This thing does struggle with some of the aesthetics and some of the direct tasks. I also asked it to book a yoga class for me. It eventually got through, but it was about ten times more painful than booking it myself.
So I look at these situations and I ask: where are the linear tasks that enable the browser to be at its best, not at its worst? Instead of throwing it into something that has fairly high ambiguity, like a PowerPoint deck, how can I give it a task where it's set up to succeed?
Creating folders is super linear — you can't screw it up. You create the folder, you name the folder, and you're done.
Another use case I think is really effective is being a second pair of eyes on the screen. If you're writing something, can it be a Grammarly or writing coach for you on the side? I realize there are probably people at Grammarly who will hear this and sweat bullets, but I think it's fair. It can literally look at anything you're writing on the web and give you a fairly thoughtful writing critique. I think that's a great use case.
Another great use case is just letting the LLM do the planning and thinking for you where you have complexity. A great example: look at this spreadsheet, perform some simple calculations off the spreadsheet, and then come back. As long as it's in the browser and it can see it, it can do that math for you. That can simplify budgeting, simplify financial tasks, simplify a lot of the basic math and thinking that we do around the web — because a lot of it happens in the browser anyway.
A creative one I've come up with — you tell me if you think this is effective — is time tracking. In theory, this should work really well for tracking time spent per site or task by interacting with the browser around what you were actually doing.
Atlas Memory Feature
And that brings up one of the most powerful features of the Atlas browser: it remembers more about you the more you use it. It has an Atlas-specific memory set that is private to you. They say that if you are interacting with the browser more, you get more value — the browser knows you better. It knows your previous chats, the previous places you visited, and it understands what you are trying to do because it has already seen your work. So those are some of the positives.
Where Atlas Falls Short
If we turn around and look at some of the difficult things — you saw some of them in the demo. There are challenges around ambiguous tasks like PowerPoint deck creation. There's also an unclear use case around the value of the utility they're going for.
To unpack that: if it books your yoga session for you in twenty minutes instead of you taking two minutes to do it yourself, is that really adding value even if it is doing all the work? I have questions. Another example: if it is going to shop for you, do you lose the pleasure of shopping? Do you lose the pleasure of planning the trip? Planning a trip is actually one of the things right on this browser's suggested use case list. So we have to think about what we want to delegate to an AI browser — and that's going to become a very real question, because this is not the last update we'll see on this browser, and this is not the last AI browser that's going to launch. AI browsers are a big thing.
Security: The Prompt Injection Problem
And that brings me to my final question: what about security? Right now, these browsers are taking in the text from a website. That's actually one of the great use cases — they can summarize text on a website, they can look at YouTube videos and tell you what's in the video. But if you use them to do that and you are on the wrong kind of page, a prompt injection attack is possible. If someone has put text on that page that instructs an LLM to do something malicious, it's not clear that the LLM in the browser can distinguish that from legitimate content.
In fact, there are known vulnerabilities in other AI browsers that I expect would persist here, where the browser will treat every piece of text it gets on the page — even malicious text — as part of the prompt. And then where are you? At that point, it just follows the prompt and the prompt injection attack succeeds.
It seems like the LLM browser builders expect us to just watch these things browse around the web very slowly, and that's how we protect ourselves from prompt injection attacks. But to actually have value, we need to not have to watch them. They need to get faster, and we need to not have to watch them. It's not clear to me yet how we demonstrate safety.
I know the teams at Perplexity building the Comet browser and the teams at OpenAI building Atlas care deeply about security — I'm not suggesting they don't. But I'd like to see the kind of browser safety card development that we've had with model safety cards, where we start to say: we know this browser is safe in these ways because we've tested it for these vulnerabilities, this is the known risk profile, and this is what you should use it for. Otherwise, there's an assumption that the browser is default safe — which is what Chrome has taught us, and that's dangerous here — or there's an assumption that it should never be used, which I think is also an overreaction.
Final Grade and Comparison to Comet
I do think there's real value here for what I'm going to call boring web work that is low-ambiguity. If I want to set it to triage my email for a long time in the background while I go off and do something else, and it's fairly low risk because maybe it's just making folders for email — fine. It can go do that. Anything like that, where you can't screw it up and you just have to logically follow the task on the web for a long time, it's going to be great at. And that's fine with me, because we humans don't love doing that work. So if it wants to pick that up, great.
Overall, my grade for this browser is C+ to B−. It's not really at a point where I think it's going to overtake Chrome, but it is much, much better than the web browsing value I've seen from OpenAI previously. Agent mode — I didn't get a lot of value from that. This is definitely better than that. I can see the trajectory of this team, and in six months, this could be a really interesting browser.
If you want my quick take versus Comet: I still prefer Comet a little bit, because it has some data inputs and outputs on key sites that make it useful. I use it for LinkedIn a lot because I can see pending invitations through the Comet browser, and it's super helpful. It also has great plugins to calendar. I expect that will get fixed with Atlas too, but it's not there yet. The speed I get from that direct data input/output is useful.
I think we're going to get to a two-speed web where we'll start to see those data inputs and outputs becoming very useful for agentic browsers where they're available, and we're going to see slower, off-road service where the browser has to use the UI. That's my first impression. What did you think of Atlas? I'd be curious to hear.