There’s a new type of browser making the rounds that doesn’t just show you websites but also attempts to be a digital assistant. These new AI browsers are called Perplexity Comet, ChatGPT Atlas, and Microsoft Edge with Copilot. They claim to be able to summarize web pages, perform tasks, and even do things online for you.

It sounds like a great idea, especially if you’re in a hurry. But there are some dangers involved that most business owners probably never thought of.

There are real security concerns with AI browsers

These browsers are not simply search engines. They are capable of reading what is on your screen, interpreting documents, and even accessing information from open browser tabs, email inboxes, and internal applications.

Much of this happens in the cloud, outside your organization and out of your control.

If a member of your team opens a financial dashboard or customer information with one of these browsers, it can be analyzed and stored by the AI, unbeknownst to you.

Prompt injection makes it worse

An AI browser processes language like a human would. This makes it susceptible to a different type of attack, which is known as prompt injection. Prompt injection is basically a form of attack whereby a hacker embeds a command on a webpage, and the AI browser follows it blindly.

It could be as simple as loading a webpage with invisible code that tells the AI browser to send information elsewhere or click on a malicious link.

A real-world attack, known as the “reprompt attack,” demonstrated how these AI browsers could be tricked using only layers of instructions written in plain language. The browser was unaware that it was being tricked; it was simply following the command.

Traditional security tools won’t catch this

Firewalls, antivirus software, and endpoint security solutions do not search for natural language commands. These solutions are designed to identify known threats, which are code-based, not conversation-based. AI solutions muddy this distinction. Once you allow them to control browser activity, the game changes.

You may not even realize they are operating on your network

Most of these solutions can be downloaded without needing admin approval. If someone on your staff is interested or has a good understanding of technology, they can download an AI browser and begin using it.

This is how shadow IT gets introduced into the organization: untested, unlogged, and unmonitored software. And if that browser is pulling client or employee data into the cloud, it’s already too late.

If you’re going to allow AI browsers, put guardrails in place

You don’t have to ban the idea completely. But don’t leave the door wide open, either. A few clear steps can keep your data and team safer:

• Set a policy. Establish what browsers are allowed and what browsers are not. Include AI browsers in the list of browsers that are not allowed, unless it has been vetted and approved.
• Lock it down. If your IT vendor is able to lock down the installation of browsers or the sidebar assistants, ask them to please do so. Also, lock down the permissions so the AI browsers cannot integrate into your internal systems.
• Teach your team. They need to know that the AI browsers can see everything on the screen, not just what is typed into it. If they’re currently logged into QuickBooks, payroll, or some other system, the AI browsers might be able to access it too.
• Stick to low-risk use cases. If you’re currently using an AI browser, only use it to view public information. Do not use it to read emails, check finances, or anything that requires login access.
• Look for shadow use. Audit your devices regularly. If the AI browser is found on someone’s computer, there should be a process in place to review it and remove it.

The tech isn’t mature enough to trust with sensitive work

These tools were not created with business-grade security in mind. They were created with the idea of curiosity and speed, not security.

If security, compliance, and trustworthiness are important to your industry, then treat AI-powered browsers like any other unknown application: with caution and boundaries.

Use the technology that makes you feel secure, not the technology that makes you wonder if you still own it.