The part of AI browsers that makes me pause is not the chatbot window or the search results. It is the moment an AI agent can start doing things inside a browser where I am already logged in.
If an AI browser can navigate Amazon, use saved credentials, and take actions for me, then the obvious question is simple: should it ever be allowed to buy something without me confirming it first?
Quick Answer
My answer is no. For anything that spends money, changes account settings, submits personal information, or takes a real action under my name, there needs to be a final human confirmation step.
That confirmation should happen at the end of the task, right before the action is completed. The AI can prepare the cart, fill the form, or line up the next step, but I want a clear prompt that says what is about to happen and requires me to approve it.
Why AI Browser Autonomy Feels Different
Regular browser security is already a challenge. We are used to being careful with email links, fake websites, phishing pages, and malware. AI browsers add a new layer because the browser is no longer just waiting for me to click around.
With something like Perplexity Comet, the concern is that the AI can operate inside the same environment where I am already signed in. That means it may have access to active sessions, saved passwords, shopping accounts, and services that trust my browser.
That changes the risk. It is not just about whether a website can trick me. It is about whether a website, link, or prompt could cause the AI browser to do something I did not intend.
The Amazon Example
The example that stuck with me is Amazon. If an AI browser can go to Amazon and purchase things for me, that sounds convenient on the surface. But convenience gets uncomfortable fast when there is no clear stopping point.
I do not want an AI agent to be able to complete a purchase without interaction from me. It should be able to help search, compare, add items to a cart, or explain options. But the final purchase should require a human click, password, biometric check, or some other deliberate confirmation.
The important part is that the approval needs to be specific. Not just a vague permission like "let AI browse for me," but a clear final confirmation of the exact action: what is being bought, where it is being sent, and how much it costs.
Why I Quit Comet When I Am Done
Because of that concern, I am careful about leaving Comet running when I am not at my computer. I make a conscious effort to quit it and make sure it is not sitting in the background.
That may sound cautious, but it is a practical habit. If an AI browser has access to logged-in sessions and can take actions inside websites, I do not want it active when I am not there to see what is happening.
This is less about assuming something bad will happen every time and more about reducing the surface area. If I am not using the AI browser, I would rather it be closed.
The Link Problem
The bigger security question is what happens when links, prompts, malware, or malicious websites are designed specifically for AI browsers.
We already worry about clicking links in emails. Now imagine clicking a link that opens in an AI browser, and that browser starts moving through a website while you are already logged in. If the AI can interpret page instructions and act on them, the old phishing problem becomes more complicated.
That is where this gets scary. If something happens through your browser, from your IP address, using your account, and while your session is logged in, it can be harder to prove that you did not intentionally perform the action.
- The action could come from your own device.
- It could use your existing logged-in account.
- It could rely on saved passwords or active sessions.
- It could look normal to the service receiving the request.
What I Want From AI Browser Security
The fix is not to avoid AI browsers completely. The useful parts are real. But the safety model needs to match the level of access these tools have.
At minimum, AI browsers should separate browsing help from account-level actions. Looking something up is different from buying something. Summarizing a page is different from submitting a form. Adding an item to a cart is different from checking out.
For sensitive actions, I want the browser to stop and ask. That confirmation should be visible, understandable, and impossible to miss.
- Require confirmation before purchases.
- Require confirmation before submitting personal information.
- Require confirmation before changing account settings.
- Make it obvious when the AI is acting inside a logged-in account.
- Give users a simple way to shut down background activity.
Key Takeaways
- AI browsers create new security concerns because they can act inside websites where you may already be logged in.
- Purchases, form submissions, account changes, and other sensitive actions should require final human confirmation.
- Saved passwords and active sessions make AI browser autonomy more powerful, but also riskier.
- I am cautious about leaving Comet running when I am away from my computer.
- The biggest concern is not normal use, but malicious links or websites designed to trigger unwanted AI-driven actions.
Watch the Video
The video above for the full discussion from the DIA vs. Comet browser comparison, including why AI browser autonomy raises real security questions in everyday use.
