Securin Researcher Maps AI’s Zero-Day Surge

Hey guys, thanks for the intro. We're here with Avril Verma, who's head of research for Securin, and we're talking about zero day vulnerabilities. There's a lot more of them, and well, Avril's been looking into this, so let's just jump right in.

Avril, welcome to the show. Thank you, Mike. Thank you so much for having me here, and excited to talk about this topic.

Yeah. So you guys did some significant research in this area, and I think we at least all feel that there's more zero day vulnerabilities being discovered in the age of AI, but maybe there's just more vulnerabilities being discovered, but not disclosed. But either way, what's your assessment of what's going on here?

What did the research show? I think we are at a very interesting point, because it used to be that vulnerability research required hours and hours of human effort. Now, you can point your Claude or you can point your GPT to any open source software and discover bugs on your own.

There's definitely been a significant increase in the number of bugs that we have discovered using AI, and that definitely is going to be a problem for the cybersecurity community in general, because we know that as defenders are disclosing everything that we discover, I am sure that the attackers are not. So you have a lot more bugs, but you only have a lot more bugs that you can actually see. You don't know the ones which people are not just disclosing out to the world and keeping it in their vault for use later.

Well, to your point, it seems like nation states are hoarding vulnerabilities that they discover and trying to weaponize those, and then they move on once they're discovered by everybody else. But, do you have any sense, like is the number of vulnerabilities we don't know about greater than the ones we do know about, or what's the sense of just how many of these things are there? I think when we started using AI to discover bugs on our own, our numbers stands now at about 400 plus, and this is when we took a dedicated effort to discover bugs and validate them.

Right? There's not a lot of cybersecurity companies who are doing that. Yes, there's Anthropic Zero Day research, there is us, and there's a few others out there.

But the attackers have a lot more motivation to discover these bugs, and we have seen the number of exploited in the wild vulnerabilities grow. So you have to connect the dots here that they are using AI to discover more bugs. There are more bugs which are being exploited by these threat actors, and you can definitely see that in the data that we have.

So the average time that used to be taken from disclosure to exploitation in the wild has come down from maybe two or three days last year to just a few hours. This means that the attackers are not even waiting for the vulnerability to be disclosed. They have already written exploit for it before people have a patch for it.

So that's why it's down to hours now. Right? It's not even that I got a notification today, and I can fix it seven days later.

That seven days period means there's probably a thousand IPs out there which are being targeted by your nation state actors, because they already wrote an exploit for it. It was already there in their vault. They just happen to use it now.

Do we need to revisit the way we think about patches? Because historically, somebody would create a patch, but we wouldn't apply it because we wanted to test it. We were afraid that something would break.

We would sit and watch everybody else for a while to see if there were reports that this patch was solid or not. But it feels like the bad guys are now exploiting things faster than ever. So do we need a different way of thinking about how we apply our patches?

100%. If the attackers are using AI to find bugs faster, we have to use AI to fix the bugs faster as well. I have seen it as a two-pronged problem.

So the first one is that there are more bugs being discovered and the rate of remediation is not catching up. We ourselves have disclosed over 400 bugs, but there's about 100 of them which we don't have responses from the vendor, because they're taking time to fix it. So until they release the fixes, we cannot publicly disclose those CVs.

The second, and one which is very less talked about, is what about all the bugs that we have discovered in end of life software? So it's not just that the vendor will release a patch for it, they actually won't, because for them, this is end of life and it is end of support. And when you go to critical infrastructure, like you said, people don't just want to apply patches straightforward.

If I have a router which is end of life two or three years ago, I don't want to patch immediately because I would need to take down all of my network, and maybe there's a hospital depending on it. Maybe there is power supply depending on it. So I have known vulnerabilities in it.

I don't have a patch from the vendor, and I can't take it down as well. So this is where we have to start thinking about how do we use AI to not just find bugs, but write the fixes for these bugs faster as well, and that is where most of our cybersecurity researches has to be focused, moving on from here. How severe are these vulnerabilities that are being discovered?

Because some people are dismissive. They're basically saying all the severe ones have already been discovered and these are low level ones. But other folks are saying, even if they are low level ones, the bad guys are getting smarter about how to chain these things together to create an attack that's even more lethal.

So what's your assessment? 100%. One thing is that the severity reported by AI for all of these bugs definitely does not match the human assessment We are seeing these bugs reported by Claude.

We are seeing the bugs reported by ours as well. The severity that AI assigns to it is definitely overstated than when a human validates it. In fact, from our own statistics, only 73% of the bugs that we discovered using AI had the right severity assigned.

At the same time, however, even if they are medium-rated bugs or low-rated bugs, they can be chained with higher and critical bugs, and immediately, attackers also have AI to chain these bugs together. So it used to be you write an exploit for one vulnerability, you try it, you either succeed or you fail. But with AI capability at my hand, I can write exploits for five of these bugs together.

Maybe two or three of them are mediums, the other two are criticals, but I have an exploit chain for me that works for me. So one thing for sure, the AI-reported severity can be overstated, but that does not mean that you can just discard it. Those are still bugs.

You have no idea how, when attacker chains these bugs together, what sort of impact it can cause in the targeted environment. Relationship between the security folks and the application development teams need to change. I don't think it's a secret that there's often not a lot of love lost.

One group blames the other one for the root of all their evil, and the other one just trying to figure out how to get around various restrictions because all they really want to do is write code and ship it as fast as possible. Can we have a different conversation now? I think we are gone past the day where security was out of your software development life cycle.

It used to be that you write code, you finish your product, then you do your static and dynamic testing, and then the security engineers are asking the coders to fix things. I don't think you can afford that time anymore. By the time you have shipped a product and the security engine's reviewing your code, by that time, there is already an exploit out for it.

So security has to shift left. You have to start thinking about code security while you're writing code. So I don't think that there is only application developers out there.

There have to be application developers who have to know what security means, and they have to write secure code all the way from the start. And the great thing is that you can do it now because all these AI models are pretty good at finding security bugs. So why not make that part of your software development life cycle itself?

" That's exactly what OpenAI, with their Hardwork project, are doing. That's exactly what Anthropic with their code security are aiming to do as well. So gone are the days where you are waiting for a security engineer to find the bugs for you.

You have to know these bugs using AI while you're writing code. You have to fix it before you even push it to your production application or even your testing application environment. Is it your sense that maybe at some point we might have more secure applications?

Because right now I feel like all the old stuff we already deployed is being outed with new vulnerabilities. There was probably a wave of new applications that were built using old AI coding tools, and I'm using the word old as in six months ago. And those things generated a fair amount of vulnerabilities because the underlying reasoning capability of the models used wasn't as great as they are today.

But if I look at where we're going, are we maybe in a period here where we're going to suffer, but longer term, maybe application security is going to improve? Is there hope? Yeah, I think there is definitely hope because I think cybersecurity in general, the attackers have always been faster than defenders.

We have always been playing catch-up, but at the same time, those same tools are available to the defenders as they are to the attackers. So it just takes us defenders a little bit more time to understand the breadth and depth of the problem. But given the right tools, we develop the processes that make sure that security is in the mindset right from the get-go.

So right now you will see a lot of wave of bugs coming up, but at the same time, defenders like us are analyzing, all right, what's the AI behavior that is able to discover these bugs? What are the kind of weaknesses that AI discovers which are actually critical versus what is maybe not so important or maybe is a medium class? And therefore, we can build our own models that while my code is being written, while I am developing, I know the kind of bugs that any attacker can discover using Claude.

So therefore, I also know that these are the kind of bugs I have to fix using Claude before the code is even shipped. It's just that for us, a little bit more analytics has to happen because we are not attackers who can just build things, throw stuff on the wall, and all the attackers need is one thing to stick to the wall, whereas we have to protect the whole surface. So it's going to take a while before things get better, but things are definitely going to get better because we know the right path forward.

There's a lot of hype around the latest generation of AI models, whether it's Mythos or whatever number of ChatGPT we're on, and there's an assumption somewhere along the line that says if we restrict access to these things, maybe we won't see as many vulnerabilities, and we'll have some responsible thing, and the defenders will get an advantage. 5 or some of these open source models that are coming up are going to be just as smart. So, is the cat out of the bag, as it were, and there's no going back at this point.

So how long before we have something that feels like a daily catastrophe? You're absolutely on point there. 6 Sol, the way they are being marketed is that these models are going to break cybersecurity altogether.

But we have been working with the lower models, your Opus and your Sonnet, and I can assure you they are equally good at finding bugs. The attackers do not need a Fable or Mythos to find critical bugs in the application. They can easily do so with the available models to them as well.

And the Chinese models, the Chinese open source models seem to be catching up very fast. 6, they experimented with it. 2.

6. So, at the same time, again, I keep coming back to the same point, however, is that I have seen these models be equally good at defending as well. Right?

So it's all about who can act with these models faster, who can build the right pipelines, add the right harnesses to these models to make code secure. Because there may be only 10 bugs that the attacker can discover, but if you are fixing those 10 bugs before your code is shipped out there, you leave nothing for the attacker to find, right? You have used the same AI models to fix things, the same bugs that another person would have used to find, right?

So, it's not a catastrophe as everyone is calling it out to me. " So that's a change in perspective that we need from the industry in general. So do you think culturally we're prepared?

It doesn't sound to me... There's technical issues, no doubt, but it also seems to me that this is all now happening at machine speed, and it's not clear to me that we humans are kind of prepared to operate at that speed. And if we have the cognitive load to keep up with all the AI that's discovering these vulnerabilities and the number of patches that need to be generated.

And, so what's your best advice to folks about how to cope with all this stuff? Yeah. I think culturally, we definitely have a problem because we grab to all of this news very fast.

Attackers are doing this, and attackers are doing that. We get into a panic mode. I think everybody just needs to take a step back and, be patient, be a little bit more analytical in their approach, right?

So it's not just that, "Oh, everything is broken now. " There are definitely steps that you can take. You just have to take a mindset of not panicking, do not believe the hype, believe only the real numbers, right?

There is a lot of hype around these models, trust me, but the real numbers state otherwise. We ourselves, during our zero-day process, we have a human validation gate as well. So everything that AI discovers, a human has to be able to reproduce and validate, and we have a 90% true positive rate.

So yes, AI is doing a very good job, but there is still 10% out there, which is noise. And I think as a defender, if you're wasting your time on the 10% noise without understanding what the 90% true positives are, that's where you go wrong. Because you have taken this 2,000 plus Anthropic discovered bugs as your problem statement.

Whereas what you really need to understand is what's the 30 or 40 of them which are actually critical, and that's the one I need to go after. So don't panic with the numbers. Try and analyze the data to find what's the real issue here and focus there.

So that's how you divert your efforts into fixing the real problems rather than fixing the volume of AI-generated bugs that are coming out there. All right, folks. You heard it here.

Take a deep breath, and the worst thing you could possibly do is run around like a chicken with your head cut off. Hey, Avil, thanks for being on the show. Thank you so much, Mike.

Always a pleasure to be here. All right. And back to you guys in the studio.