How AI Broke 20 Years of Browser Security Progress

Hey everyone, it's Alan Schimmel. Working from the home studio today, but still doing interviews and talking to people, and trying to give you what you need to know for today's world. Let me introduce you to Or Shed.

Or is the co-founder and CEO of LayerX, and if you haven't heard about LayerX yet, that's okay, we're going to tell you all about them. Actually, I'm not, Or is. But first, let's hear a little bit about Or.

Or, welcome to Techstrong TV. It's great to have you here. Thank you, Alan.

Thank you for having me. My pleasure. So Or, as I mentioned, we're going to talk LayerX in a moment, but let's hear a little bit about your history and background.

Yeah. So in the last about 15 years, I'm working in network security in different hats. I actually started as a practitioner, so like many of your viewers, I was running incident response activities, security operations, et cetera.

Grew up in network security. About a couple of years ago, while working on a couple of projects, it hit me that I have a visibility gap. I was looking for a new sensor, and I'll explain the logic.

Before that, the world was very much aligned on files, file systems, devices, network low-level security. But where are users? Users are causing most of the problems.

Users are doing stupidware on SaaS applications, identities online where they process data. Now with AI, it's tenfolds more significant. And we were looking for a way to create a way to find where users act.

So we wanted security where work is happening, and it was a very fundamental shift in paradigm from trying to deploy something very heavy on the network, and that's our vision. That's our vision, to get security to where work happens, and eventually being able to detonate and reduce the blast radius of what users are doing online, in SaaS applications, in AI tools, from the user standpoint. Accordingly, we started LayerX a couple of years ago.

We started with one product, which is a browser security platform. We're now expanding to other desktop applications as well. And our purpose is not to provide security on the network level.

We deploy security on the very dangerous 10 inches between the user and the keyboard. What do they type? What kind of prompts do they send?

What kind of activities and which other tools are existing within their workspace? Excellent. We didn't mention websites and what's the on-ramp here to work with LayerX?

So, employees spend about 90% of their time online. They work on the operating system on their Mac or the Windows, but many things are happening in the browser. So the browser is a full-fledged operating system.

It has extensions, which is files working in the context of the browser, processes that have access to data like DLLs or EXE on operating systems. There are assets, cookies, passwords, credentials that are hidden inside the browser memory files that are accessible to those extensions. And now we also got agents.

So the first step into autonomous agents by the AI vendors actually came in the form of something in the browser. Those are AI browsers- Yeah ... and AI extensions such as Cloud for Chrome.

Comes the question why. Why did all these great AI companies started their user journey in the browser? That's because of the fact that this is where work happens.

So within this environment, which is basically a full-fledged operating system, so many things happen. Your most important assets are there, your cookies, your credentials, your identities. Now you have agents running inside a browser.

Can you tell for sure, is it a user now browsing Salesforce or maybe an agent they've assigned a task while they went to make themself a cup of coffee? So we don't really know what's going on, and everything happening inside is very abstract, very complex environment that doesn't have visibility from the outside. So we just treat this as a brave new world.

We want to get inside that browser and inside the application layer and really understand what's happening not only on the network, but everywhere, including the device. The way the users are acting, the way the agents are acting, and actually provide full visibility. I love it.

You mentioned the AI browser, this new Google Cloud. I think something Google came out with a day or two ago, some rubber duck or somewhere, who's using multiple AIs now for stuff as well. And they're great.

I use it. I'm a big fan of all these things. The problem is, I'm also a security person, and I sit and think, are we really thinking through the implications of what we're giving this access to, of what it has a window into.

And of course, Mythos, Glasswing, and everything that came out the week or two before, and its ability to find vulnerabilities. And the browser, I read an interesting article a couple, maybe it was a week or two ago. Stolen credentials were the primary way in for a long time.

But recently, we've seen the rise again of browser... Of vulnerabilities. Stupid things, cross-site scripting, buffer overflow, stuff that should've gone away 15 years ago, is now on the rise because AI is...

Well, a lot of the code's being generated by AI, it's not being tested well, and AI is better at finding these kinds of vulnerabilities. Some of them have been sitting there in plain sight for years. So once again, the browser is the focal point of the way in for these malware folks.

It's a good time, I guess, to do a survey and see what people are thinking about and what are they doing about their browser extension security. And I know you guys recently did a report on this, yeah? Yep.

So reverting what you said, I agree about everything. And coming from offensive security, every new technology, first of all, gets used and then gets hacked. You don't put the carriage before the horse.

Mm-hmm. That's the way things happen in life. You first have a new workspace.

If it goes well, you add security later on. AI in the browser is making a couple of changes. First of all is the fact that now we have multiple agents running in user space.

I recommend for the technical folks to read about WebMCP API. Now every extension will be able to navigate the web independently and mimicking user activity. Distinguishing between a user and an agent will become very, very complex, and in a way, we're going towards a world in which malware applications and users are all acting on the same manner inside a device.

That's one thing. Second thing, the security by design of browsers. Maybe some of you remember that 20 years ago, you can buy an exploit kit on the web, pay 50 bucks a month, although then it's equal to 200 bucks a month of today's terms, but it's still not a lot of money.

And you have free access to exploits on Explorer. And it was really, really easy to exploit browsers and get to full on-device code execution, sandbox escape, et cetera. Since then, the different browser vendors, that by the way, provide us an amazing nation state-grade security for free, have added sandboxing mechanism.

The sandbox is at the foundations of security of each browsers, the fact that everything is running in a container. In the last half a year, the different browsers, most of them actually added a new mechanism that's overriding the sandbox, and as an AI mechanism, can now hop between different tabs. It's working in its own environment.

It has visibility to everything. It doesn't come with its own security, so there are a lot of vulnerabilities. But more than anything, 20 years of security progress were pretty much moved aside at the benefit of pushing this AI that's now overriding all the sandboxing.

So forget about anything you knew before. Does it take us to the same attacks as before? No, but it's the same impact coming from a new vector.

So if before you would run XSS from one tab to another from the device memory, now you just hide a prompt in one tab and find yourself injecting or exploiting another site, stealing cookies, stealing data, injecting memory into the local MD file of the agent on the device. All these things are creating some sort of a comeback to traditional attack flows, but being accomplished in a new manner. I would say that in every technological cycle, you have the hype, and then gradually the security is catching up and getting to a steady state.

I think we're only at the beginning because the AI is moving way faster than anything else. And then you mentioned Glasswing, the ability to fuzz vulnerabilities in real time. Chromium is an open-source project.

Every browser that's going to the web is saying, "Hey, this is my user agent. " And browsers do that just because. They just say what version they have.

Just imagine if it takes you five minutes to find a vulnerability and get it to exploitation. Even traditional attacks are about to make a comeback. So different browsers that are driven by Chromium are not patching at the same sequence.

I don't want to name drop or put a finger at any specific vendor, but there are some companies that claim to be enterprise secure browsers that are forking from Chromium. On average, they have one or two weeks delay from Chromium. That one or two weeks traditionally has been in no time.

It's really hard to find a vulnerability, a zero-day in Chromium and build an RC, but now it takes five minutes. If an attacker can get access, and they will get access to those models, having a browser not updated for one week actually will become a big problem. So we moved from a world in which security by design was perfect, near perfect, to a world in which old attacks are coming back in new forms, in which we have new kinds of attacks because we have in-browser agents, and the fast-paced vulnerability scanning will allow even old attacks making a comeback as well.

So are getting back into the device as a control plane. The browser is the main attack surface, and everything is about to go back to the way it was 20 years ago. Crazy.

Crazy. I just want to put sort of a tail on the donkey for when we talk about browser extensions and browser extension security. 99%, maybe more than 99%, because I don't know how many decimal points you went into in the report, but 99% of users have at least one browser extension installed.

So for all intents and purposes, everybody has this. Everybody has extensions. 75%, three-quarters, a critical mass, have high or critical permission levels.

Right? This ain't zero. That's not zero trust, right?

This is giving me the keys to the car. That's a scary statistic, right? It's a scary finding in the survey.

Right. So first of all, for the audience, Lerix has a technology alliance with Google. So we are providing sandboxing capabilities and engines to the Chrome Store.

The Chrome Web Store, I'd say 70% to 80% of all the extensions in the world. And we basically get ability to understand the dimensions of this ecosystem. So what it is, an extension in the browser is like a file on the operating system.

Can a file be malicious? Yes, it defines it as malware. So a malicious browser extension is the malware of the browser.

The more assets you have inside a browser, the better it gets to be an attacker. So we moved from a world in which hacking the browser was just excellent because this is where we had cookies and credentials. " Let's pairing from signing in and using likeFive different MFAs at once is based on those secrets stored in the browser.

So the browser was an amazing place to hack. Today, it's an outstanding, the best in the world, because this is not only where you have identities, but also where you have AI and where your agents are also running. So browser extensions, to the extent of their capabilities and permissions, can steal all the data.

What kind of permissions do they have? Access to cookies. One line of code, they can dump all your cookies.

They can use your browser to route traffic, so they can actually streamline their flow into the application from your browser. They can create an overlay on top of every app to steal your data, and now they've access to AI. There are two new permissions that are coming in in the last couple of months.

One of them is Prompt API. Now, browser extensions can run AI models in the browser to do more sophisticated things. For example, personalizing an attack just for you.

The second one is WebMCP API. That will allow extensions to navigate the web as if they had eyes and ears. Basically mimicking user activity.

Many organizations nowadays understand the risk, and that came after the fact that it was a neglected attack vector that people did not pay attention to. In Christmas '24, there was a big breach. A couple of cybersecurity companies got hacked, and attackers actually took over their account in the Google Chrome Store, replacing the browser extension with malware.

That's a very interesting TTP. So attackers take a benign browser extension that's already approved or allowed. It's put in some sort of an allow list in your organization.

They hack it, they buy it, they compromise someone's account, and then they find a way to get it in. In some cases, they even put a very simple extension on the Chrome Store called Dark Mode for ChatGPT. Over time, they add more and more code.

So they build up. So I call that, it's a Shawshank Redemption kind of attack flow. Day by day, they add a little bit of malicious code.

After 15 years, they are breaking free- They're rolling ... and they're able to steal your data. So that's how attackers are working.

Traditionally, what organizations were doing is running allow lists or block lists, but attackers already understand that. And for them, it's actually an opportunity, because if they can get an extension which is in your allow list, they have the keys to the kingdom. Are browsers doing anything about that?

It's not that easy because the different browser vendors are also using extensions. What happened in the last couple of years, since AI came to the market, before AI came to the market, what kind of extensions? Let's imagine a conversation between a CISO and the user population.

" It's not for their business. It's not helping us. So which users would eventually get extensions?

Mostly developers, historically, mostly developers. And then a small set of extensions, like under 10. Now with AI, there are 3,300 extensions calling themselves ChatGPT something on the Chrome Web Store.

I doubt that 3,300 of them, or even 300 of them, are actually tied to OpenAI. The main change has been that moving from a world in which developers were causing most of the overhead, but they have a lot of political power, but they're a small population, to a world in which everyone says, "I need this extension. " What?

Don't you want me to do my job? Don't you want the company to be more successful? So that changed the conversation between the user population and IT to the security team.

The security team is being left without any choice. That takes us to the report. Eventually, when you look at what kind of extensions you have in your environment, everyone got one, so it's not surprising.

But most of your users have many of them. " And by the way, the malicious ones, they never call themselves malicious extension. They have some sort of a name.

It can be- Yeah ... Amazon coupon code, something- No ... ChatGPT.

Name alike, sound-alike names. This is something- Exactly ... in repos and everything for a long time.

But- Putting aside that all kinds of flaws by design, for example, if you use an allow list mechanism and you have an on-device malware, it can push an extension with the same ID as something in your allow list. It is a crazy thing that now we see attacks in which an on-device malware is pushing a malicious browser extension. Going back 20 years, 20 years ago, the browser was just a checkpoint.

It was the device TSA. It was just checking who you are, and you want to get to the device. Nowadays, actually, the data in the browser is more important than the data on the device.

So you find organizations finding a malware on the device, pushing a browser extension, that's a privilege escalation. So you have plenty of extensions. Most of them have, I say, knockout, game over permissions, which is ability to run scripts, ability to open the debugger, or ability to access cookies.

If you have something that has a ability to access cookies and it's also an agent in the browser- You're done ... you can't really tell you're done. You have no idea who is inside, and it's impossible- Yeah ...

to prevent account takeovers. Now, probably some of the viewers are asking themselves, "Yeah, Vali, how big is it? Is there a chance I have a malicious extension?

" And in reality, first of all, every opportunity that we ever had in LayerX, not in deployment, but during the POC stage, we found a malicious browser extension. So when they stole LayerX on 20 users, you'll find a malware in the browser. That's one.

Second thing, you never have a smoking gun, but you get the signals. A user that has multiple account takeovers in a row, and you never really know, pardon my French, is that person so dumb he falls victim into every phishing site, or is it something else? No, there's something there.

In all those cases in which you have a user that just gets hacked again and again and again, and you don't know why, there is a very good chance it's a malicious browser extension. Excellent. Or unfortunately, we went through our time here.

We didn't get a chance to dive into this survey as much as I wanted, but I think this discussion was more important. But I do want people to go get this survey. How can they get the survey and find out, see for themselves some of the things we're talking about here and what people like them are saying and doing?

Where do we get this survey? Yeah. So first of all, LayerX is a gold standard in browser extension analysis.

We have the largest database in the world. We have the best partnerships. We work with all the browser vendors on that.

So if someone will find this, it's us. We are releasing information to the ecosystem. We have a publicly available sandbox you can use.

You can run extensions. com site, and this is also where you can download our different reports, the browser extension report. We have other reports as well, and you can also access the Extensionpedia, which is our free resource you can use to understand which extensions you have, what kind of risk do they have, understanding the risk score, and you can even consume the LayerX risk scoring on Google Workspace, and hopefully soon also on other browser management tools.

So we are giving away to the community. We're very happy to do that. , And you can find all of that on our site.

I love it. Or thank you. We packed a lot in here.

I appreciate it. Keep up the great work with LayerX. I got a feeling in the next couple of months, we're going to see a lot of action in this arena, and we're going to count on people like you and LayerX to help us.

We'll have you back soon. Thank you very much, Kevin. Thank you.

Or Shadrira, CEO, co-founder of LayerX. Go look at your extensions. Go download the report.

Go see for yourself what you got on yours, because we all, as I said, 99% of us have extensions. They're a potential backdoor, front door, window into us. We're going to take a break on Textrunk TV.

We'll be back in a minute.