News & Stories

Browser AI agents are less security-savvy than you, research finds

Ray Fernandez

Jul 4, 20259 min read

Browser AI agents are less security-savvy than you, research finds: Header image

Presented by big tech as the AI tech that will take productivity to the next level by doing what humans do, only better and faster, AI agents are the new hype. And within this tech, Browser AI agents are rising fast as one of the most promising areas of development. However, it is also proving to be one of the most dangerous. 

Thousands of companies deploy AI agents — SquareX Lab finds they aren’t safe 

With surveys showing that 79% of companies are already using AI agents in one way or another, the chances are high that you currently use or will soon use a Browser AI agent. But what are they, exactly?

You can think of today’s Browser AI agents as AI that has access to the internet and your accounts. You can prompt it to do research, create a presentation and send it out to colleagues, arrange your emails, download files, and do, well… basically anything else a human can do on a computer, including typing, filling in forms, and buying groceries. 

However, despite the many time-saving benefits the technology brings, a new investigation by SquareX Labs found that when it comes to protecting your data, Browser AI agents do a terrible job.

After testing open-source Browser AI agents against common, basic cyberattacks that humans are trained to avoid, SquareX concluded that the new AI Browser agent wonder-boys of the office are, in reality, the weakest link in cybersecurity. 

The fundamental “truths” of Browser AI agents

SquareX Labs explained that Browser AI agents have 2 fundamental “truths.” 

“First, they are trained to complete tasks, not to be security aware,” SquareX Labs said. Second, Browser AI agents are coded to follow the user’s instructions, “even if certain steps may expose them to security risks.”

We have seen AI apply this second fundamental truth time and time again. When at a crossroads between completing a task and breaking a rule, AI will often choose to “bend” its guardrails. 

For example, GenAI tools have been found to cite non-existent cases when used by lawyers to draft legal court documents, have lied and coerced humans to avoid being shut down, and fabricated non-existent facts to make it appear as if they had completed tasks. These are all cases of why AI is still described as “experimental” by big tech — because it makes mistakes.

A screenshot of the disclaimer on Browser Use framework on GitHub.
The warning disclaimer on the Browser Use framework on GitHub has some wild claims, including that the software will use private credentials, stored passwords, and every app and user interface in a MacBook. Image: Screenshot, Moonlock.

Tests reveal that Browser AI agents lack basic security awareness 

SquareX Labs’ investigation proved that Browser AI agents, unlike most employees, have not been trained in basic cybersecurity hygiene and practices. For example, in one test, SquareX Labs asked a Browser AI agent to search for Salesforce, log in to an account (credentials were provided), and generate a sales pitch. 

What the Browser AI Agent did next took all the time and money that companies invest in training their employees in basic cybersecurity skills and threw it out the window. 

A screenshot of the prompt given to Browser Use in the study.
A SquareX screenshot shows the prompt they gave the open-source Browser Use framework, which led to a basic cybersecurity phishing breach. Image: Screenshot, Moonlock.

First, it used Google to search for Salesforce. Then, it clicked on the first sponsored ad it found on the page. Of course, in this test, as in real life, the ad was a fake ad paid for by cybercriminals and designed to lure victims to a fake Salesforce login page.  

Not only that, but the Browser AI agent did not even notice that the login URL it clicked on was abnormal (https://login-salesforce[.]vercel.app).

Even after paying no attention to the URL (or not considering its risk — we don’t know which is worse), the Browser AI agent went ahead and entered the username and password of its user on the phishing page. 

A screenshot of the fake Salesforce login page where the Browser AI agent entered a user's credentials.
SquareX Lab shared this screenshot showing a Browser AI agent failing to recognize a suspicious URL, a red flag that human workers are trained to watch out for. Image: Screenshot, Moonlock.

The simulated digital attack didn’t stop there. The Browser AI agent was redirected to the legitimate Salesforce page, where it logged in again and completed the task it was assigned. 

Black-box agents: Loss of visibility is a serious problem 

The speed of Browser AI agents means that the example above happened in mere seconds. The problem? While companies like Open AI, with its flagship Browser AI agent Operator, claim that users can retake control of the wheel at any time when using this technology, that isn’t happening in reality.

In this way, Browser AI agents are black-box tech. They can go from A to B, apparently recklessly with no regard to privacy and security, and the user will never see what happened in the interim.

In the case above, the victim would have no idea that his Salesforce credentials had just been compromised. If they used those same credentials on other company or personal accounts, one could consider them stolen as well.

In another test, SquareX Labs asked the Browser AI agent to research and write a report on technology. It was then to search for a file-sharing app to share the report with colleagues.

Once again, when searching for a file-sharing app, it failed to recognize a malicious app that impersonates a legitimate file-sharing app. Using an OAuth attack technique, the Browser AI Agent was tricked into giving the attackers full access to the user’s Google Drive. 

A screenshot of the AI Browser agent giving away permissions and account access during from the SquareX investigation.
A screenshot from the SquareX investigation shows the Browser Use AI agent giving away excessive permissions and account access. Image: Screenshot, Moonlock.

For both demonstrations, SquareX Labs used the open-source Browser AI Agent framework Browser Use, which thousands of organizations have already deployed.

Are browser AI agents worth the risk?

SquareX Labs said that there were several clear red flags that the Browser AI agent could have identified before it was too late, such as checking suspicious sites, URLs, unfamiliar file-sharing apps, and the fact that the service was asking for unnecessarily extensive permissions. 

To do their work, Browser AI agents often need to access email, file sharing, workspaces, meetings, and collaboration apps. This presents attackers with an opportunity to run phishing and spear phishing campaigns, identity theft, account takeovers, and more.

A screenshot showing an example of a browser than implements Browser AI agents.
Get ready for a wave of Browser AI agents. An abundance of Browser AI agent offerings could be seized by cybercriminals as an opportunity to run a variety of cyberattacks. Image: Screenshot, Moonlock.

SquareX found vulnerabilities across the following Browser AI agent tasks: 

  • Purchasing groceries
  • Collecting coupons
  • Sharing documents.
  • Researching and collaborating
  • Filling out forms

Of course, this list is not exhaustive. 

Unchecked innovation: Who pays for the AI developers’ “move fast, break things” strategy?

Sadly, the fact that Browser AI agents fail so miserably in detecting common cybercriminal tricks and have less cybersecurity awareness than humans comes as no surprise.

The race for AI innovation is moving so fast that competitors are rolling out new models and new AI technologies with little to no security or compliance considerations. Users, and even companies, agree to these risks when they download the app. The risks are often found in the fine print of the Terms and Conditions sections.

Technology that is developed at this incredible speed and pushed out into the market is rarely private or safe.

It’s painful to come to terms with the basic failures demonstrated by “experimental” technologies that are widely deployed in companies around the world. We would all like AI that works flawlessly. However, that is not a reality. Much of the responsibility lies on the user and how much risk they are willing to take. 

At the time of this writing, there aren’t many Browser AI agents operating, but that will soon change.

Some of the most well-respected Browser AI agents include OpenAI’s Operator. It is only available to Pro users as an early research preview in the United States for a price of $200 as part of the Pro account. 

Other Browser AI agents, like Browser Use, are open-source and designed for the more tech-savvy users. AI open-source projects are always a risk; for example, Browser Use has 447 issues or bugs reported. 

A screenshot of the GitHub page of Browser Use.

There are other Browser AI agents, such as Harpa AI, used by about half a million professionals, Monica AI, AIPRM for ChatGPT, and Merlin AI. Most Browser AI agents are currently keeping a low profile in “join-the-waiting-list” mode, or they are not free or public yet.  

Some companies are already developing or have developed Browser AI agents for macOS as well, such as Simular and the macOS open-source version of Browser Use. 

Browser AI agents available on official app stores

Checking official app stores, we found that Browser AI agents for Android devices are showing up in abundance on the store. In contrast, the Apple App Store has a shorter list of available Browser AI agents due to the stronger guardrails in place in the Apple App Store. If you wish to use a Browser AI agent on Apple devices, stick to the App Store.

A screenshot of the Apple App Store showing results for Browser AI agent apps.
The Apple App Store has tighter security guardrails. A search shows that Browser AI agent apps are slowly but surely on their way. Image: Screenshot, Moonlock.

Given the potential, hype, and appeal that Browser AI agents have for work and personal technology users, we expect the global market to be flooded in less than a couple of months.

A screenshot of Android search results for Browser AI agent apps.
On Andriod, results for Browser AI agent apps are already abundant. Image: Screenshot, Moonlock.

This in itself will create a new problem for users and a new opportunity for criminals. They will surely impersonate Browser AI agents that go mainstream and popular, as well as develop fake but convincing Browser AI Agent sites to trick users into giving away credentials, personal data, or downloading malware. 

To be clear, current public AI models available online can also be used to automate tasks that AI agents do. This includes scrapping email data, running scripts on Excel or Google Sheets, and much more. The problem is that Browser AI agents significantly lower the bar for automation. Call it automation with fancy buttons and a nice user interface. It’s like no-code automation.

What you should know if you want to use a Browser Agent

Users should be aware that installing a Browser AI agent extension may grant it access to a wide array of files, resources, and accounts, including: 

  • Files 
  • Web pages
  • PDFs
  • Downloadable files
  • YouTube
  • Selected text
  • Uploaded files
  • Online Accounts
  • Gmail
  • LinkedIn
  • X
  • Google Docs
  • Google Drive/Dropbox
  • Personal accounts (such as Mail or iCloud if given permission)
  • Browsing Data
  • Current tab and URL
  • Full page content
  • Search results
  • Clipboard and Notes data
  • Integrations and automations
  • ChatGPT/Claude/Gemini/Llama
  • Zapier/Make/n8n
  • Notion/Slack/Calendar/Teams/Zoom

To stay safe, users should limit agent permissions, granting access only to trusted sites and accounts.

Browser AI agents should not be able to access: 

  • Passwords
  • Tokens
  • Full cloud storage
  • Other extensions

You can also add explicit guardrails to a Browser AI agent by including instructions in prompts that prevent agents from granting OAuth or entering credentials.

Ideally, users should be able to configure Browser AI agents so that they can verify URLs and permissions manually. However, we do not expect this setting to be available by default on most Browser AI agents.

Anti-malware solutions are poised to play a key role in the near future if they can detect, flag, and shut down live threats linked to Browser AI agent activity. As such, staying aware of anti-malware and browser cybersecurity solutions is a good idea.  

We advise that users stay away from Browser AI agents until they have a clear understanding of the risks involved in their use. 

Final thoughts

AI agents and Browser AI agents are an inevitable technology that will soon be everywhere.

Will humans soon lose their appetite for typing, navigating the web, and doing their own work entirely? It seems unlikely, but this new AI tech does have positive uses, from acceleration to augmentation and collaboration.

Keep in mind that when it comes to cybersecurity awareness, Browser AI agents have not been properly trained. In many cases, you have better cybersecurity awareness than they do. As SquareX found, Browser AI agents have replaced us as “the weakest link in any organization,” and that says a lot.

MoonLock Banner
Ray Fernandez Ray Fernandez
Ray has been covering tech and cybersecurity for over 15 years. His work has appeared on TechRepublic, VentureBeat, Forbes, Entrepreneur, and the Microsoft Blog, among others.