Ahmed Achchak (Qevlar AI) & Réza Malekzadeh (Partech): When AI attacks, can AI defend?

0:00Andreas Munk Holm:Before we get into today's episode, take a look at your screen right now, because there is something amazing in which you're having right now. If you're an impact investor or working in corporate venture, specifically in sustainability or climate, I want to flag something for you. Happening just before the famous Morroland Festival on July 23rd in Belgium, we're curating the investor program for Love Tomorrow. And there's a separate program at the Impact Circle's Investor Lounge on Friday the 24th. It's curated, it's intimate. Conversations between impact fund managers, climate LPs, and corporate venture leads.

0:30Andreas Munk Holm:We are hosting the program. If that sounds like a room that you want to be in, I suggest you go to the show notes. Security can no longer be an afterthought. In the past, people would go build something and then say, oh, let me think about the security piece. I think right now, it's one of those components that you have to take into account from the get go, the later you try to address it, the more stuff you will have to fix. Security is now a concern to everyone. Because up until now, if you were building a startup that is not from the security space, you would typically not worry about security before probably a few years.

1:06You would worry about scaling the business, etc. But now, because the bar is actually so low from an economic perspective, it's fairly easy to attack virtually any organization that you want. But the issue is attackers never follow the exact same path again. And so that meant that you had to constantly update the playbooks or accept very little automation results, like meaningful results coming out of them. And our first idea was if we switch this logic of having a static playbook to having a dynamic AI that is able to run actions by itself. What matters the most is probably the false negative rate, because that's really the one that harms you.

1:43That means that something has happened but has gone completely undetected until it's maybe too late. And I think if you have the attitude of thinking that you're absolutely impenetrable, then you're going to suffer.

1:56Andreas Munk Holm:AI has made it trivially easy to be a bad actor. You don't need to know anything about security anymore. You don't even need to write your own malware. You don't even need to know how to pop fish. The tools are there and they scale infinitely. The question every CISO or CTO, every founder running infrastructure is asking right now is, what do you do on the other side? Ahmed Ashak and Hamsa Sayer asked that question in 2023. As machine learning engineers, not as security people. They built Kevlar, an AI that investigates and responds to security incidents for some of the largest enterprises in Europe.

2:34Andreas Munk Holm:Reza Malik-Sadeh from Partec co-led the$30 million round in March. Today's conversation is the one we set up in the prep call, how AI is changing the threat landscape and what to do about it.

2:49This show is not investment advice, and the hosts of this episode may be invested in the funds and companies featured.

2:55Andreas Munk Holm:Welcome, guys. Thank you for having us. Pleasure to be with you. Thank you. First and foremost, Ahmed, did I get some of this correct when I described your founding story and what you're doing? Yeah, you definitely have. You put it in a much shorter way than I usually do. But that's exactly it. I mean, when we created Kevlar with Hamza back in 2023, we just noticed that AI was already everywhere in security, but it was mostly used on the detection side, embedded within malware detectors like intravirus, what we call EDRs, etc. And we thought that the weak link was precisely the operational side, because you now have a lot of detection tools.

3:38You have a lot of alerts that are raised as a consequence to this. But you have a bit of an issue with the capability of security operations to deal with that. You have many signals and you just are not able to investigate all of them. And we thought that we could have a role in this. And then you're absolutely right. AI is now being widely used on the attacker side. We've all read, sometimes maybe read a bit too much about Mythos and about GPT 5.5, etc. But this is just a world we are about to start living in, one where AI has a great role both in offensive and defensive security. And yeah, we were glad to be part of that.

4:17Andreas Munk Holm:And Reza, I didn't give much background on you, but you're quite the security buff. Maybe you can talk a bit about your own background in this context, and then we can dive into the bigger conversation about the AI shift that's happening in security right now. Sure. Thank you, Andreas. Well, before turning on the dark side and becoming an investor, I actually spent some time working for a variety of tech companies, and particularly at Cisco in the security technology group. where back then I worked on an initiative called Network Admission Control, which was a new way for Cisco to kind of look at devices as they were accessing the network.

4:57We were already going through some changes in the industry, and there was a lot of innovation happening in the security field, and I was lucky enough to be part of that at Cisco.

5:07Andreas Munk Holm:Yeah, and Raza, you've spent, as you just said, 20 years in this space. You've watched the cloud rewrite how the category was shaped once. Now you're seeing AI do the same thing. Some say that this is a second opportunity for massive disruption for those that are doing harmful things. But there's also, of course, an opportunity for people like Ackman. Can you talk a bit about this shift, which you saw back then, and then what you're seeing now? I think security is a super exciting space because it all keeps changing very, very dramatically each time we have another technology shift. So many years ago, people used to have all of their applications on their own servers in their basement.

5:47And it was quite easy to protect at that point. You would put a big Cisco firewall in front of your network, and you were able to filter pretty much everything coming in your network, and everything was running locally. And then the cloud came, and you started putting your applications elsewhere on other people's infrastructure. And then you actually started using components or libraries or pieces of technology from third parties to build your own app. And then you relied on other people to provide services and applications to you. And so that was the first major shift where that traditional approach to security could no longer work.

6:25You could no longer put a big piece of protection in your basement and be all done with it. today we're seeing a second major shift where AI is just dramatically changing the way people can approach technology and you know good guys like bad guys and typically the bad guys are always on the lookout for disruption because they can start and come hurt you in a way that you had not predicted in the past and so AI is helping bad guys automate attacks find new threats find new holes in your environment. And so it's just making it much, much easier for the bad guys to try and hurt you. And if on the defense side, you're not using it, then you're going to be at a disadvantage.

7:11So I think it's a great opportunity for people like Ahmed and his team to come and deliver something disruptive and something truly useful for the market out there.

7:23Andreas Munk Holm:Let me ask you one really quick question before we go over to Ahmed. Is this an opportunity or threat for only the enterprise level? Or is it something that even us smaller business owners need to start thinking about? I think it's an opportunity for everyone to think about, because even a small business owner is going to be attacked today. One of the things that AI makes easy is attack everyone, or as many people as you want. And then it's a math game of, you know, the law of great numbers. If I attack a gazillion people and I can actually, you know, hurt a few, I'm going to make what I want as an attacker.

8:03So why would I limit myself? You know, it used to be complicated to attack because you had to use tools by hand. So you would only go after the big fish. But today you can go after everyone. I mean, even my kids get so many phishing emails every day. is just scary. I actually cannot agree more with Reza. If anything, I think the increase of risk is probably higher on the small and medium businesses than it is on the enterprise accounts, because enterprises have already been under attacks, heavy attacks for quite some time, precisely because of what Reza said. These are the large fishers, the large whales that everyone has been trying to harm.

8:40But now you have a technology that makes it economically much easier to do harm at scale. And so obviously, I think that the increase of risk, as I said, is probably stronger on the S &P side.

8:52Andreas Munk Holm:Ahmed, I want to pull on a thread here because you came from ML, not security. So machine learning and AI was very much close to your heart, but then we saw this shift happen. I'd love to ask you, what did you see that the security industry hadn't seen because of your AI background? That's a good question. I think that I don't know whether it's the AI background that helped us on that specific answer that I'm about to give but I think that not coming from security just helped us dare to do things differently I'll explain what I mean by this if you look at 2023, that's the year where Kevlar got started, at the time the term AI SOC didn't even exist, like no one knew what that was, and the way security operation centers used to deal with attacks with incidents was either through completely manual operations, like humans looking at that, running investigations by themselves, pivoting from a tool to another until they have enough data to conclude that something is happening or that this is a case we can dismiss.

9:56Or they had what we call playbooks in tools called source that are responsible precisely for automation. But the issue with these playbooks is that at least back then, they were completely static, which means that you had to build a bunch of if-then-else rules nested together. And then you had a robot that would basically do whatever you told it to. But the issue is attackers never follow the exact same path again. And so that meant that you had to constantly update the playbooks or accept very little automation results, like meaningful results coming out of them. And when we started, we thought that that just sounded too old-fashioned.

10:34You said something, Andres, earlier, which is trying to balance the field a little bit between attackers and defenders, and it just seemed that attackers were favored a lot by this way of doing things, and we thought that we could dramatically change it. And our first idea was, okay, what if we switch this logic of having a static playbook to having a dynamic AI that is able to run actions by itself? And again, at the time, this word agentic wasn't even a thing yet. We needed a few marketers to work on it, to create it. But we had this idea of an AI that is able to sequentially draw actions that mimic human behavior and that makes it much faster, much more scalable to deal with incidents.

11:19And so coming back to your original question, I think that not coming from security gave us maybe the trust, maybe beginner's mistakes, if you want, but sometimes these mistakes turn out to be good ideas. And that's really what gave us the idea for Kevlar. Yeah.

11:35Andreas Munk Holm:Right. I want to ask you about this because of course, underdogs is something that we in the industry tend to really like underdogs that come from another industry and try and disrupt an established one. It of course, is what we've seen multiple times happen. We've seen it with Mark Zuckerberg, with Facebook that ended up completely disrupting media. Media was completely oblivious to that. And we've seen it as well with Airbnb, where the hotel chain said this is never going to happen, but then it did. And we've seen it again and again. Can you share a bit about how you as a VC think when you see a team that's coming from outside of an established industry and saying, we can disrupt it, whether you like it, whether you don't like it, what it makes you think and hesitate or double down on?

12:20look I actually think it's it's it's pretty uh pretty interesting because when you've been used to doing something then your ability to go out of it is actually limited time and again there are all these examples of why you're doing it like this well because we've always done it like this and I think that if you come from a different environment and of course you're super smart like these guys are and you're super hard working then you can actually think of it and do it a different way. You know, historically, we had another example in a completely different industry of Alain, who is a healthcare provider in France.

12:54And the two founders came from a different, totally different environment. And they thought, hey, let's go disrupt this thing. And they started by building a mobile app. And of course, all the traditional healthcare providers in France laughed at them, but they don't anymore because of, you know, the fact that their new approach brought such an ease of use that they're now killing it and really winning massive markets. And I think that's the interesting thing is if you come from somewhere else, you're going to think about things differently. You're going to be able to not be stuck in the same thought processes that others have been.

13:32And you can then, if you're really good, bring something innovative enough and something disruptive enough that it becomes super interesting. In a way, it's easier to think out of the box when you don't have a box at all to start from. So that's basically what summarizes the state we were in back then. So you just learn from first principles. We learned from speaking with SOC analysts. We literally reached out to people on LinkedIn and we spoke with whoever wanted to speak to us. And that included some of the largest organizations on earth. We were ourselves surprised with this. But it also gave us the trust that whatever we were building was interesting to these folks and that it was worth our times, worth a bit of sleepless nights.

14:18So, yeah, that's, I think, the advantage that we had back then.

14:23Andreas Munk Holm:Aghman, I want to ask you because you described Kavla as essentially putting an AI on the defensive side to fight the AI on the offensive side. I'd love to just walk through what that really looks like, not at the product level, but at the threat level. So we have on the one side an AI attacker that's doing what they need to do to get into a machine. But what does Kavla do in response? That's a good question. I mean, again, when you think about AI on the attacker side, you should think about it as running multiple, trying to create multiple holes in your organization, right? Like whoever the organization, you have attackers that want to get in for multiple reasons.

15:01And so you have different doors that you can use. You have the email security one. Reza mentioned phishing. that's a classic one, it's still very effective. And if anything, generative AI is making it easier to build very effective, scalable phishing campaigns. But you also have the classic malware route, you have the network, et cetera, et cetera. So basically what happens is you have security operation centers, which are basically teams of people that are responsible for triaging these signals. Because you basically have tons of signals, but you have no idea on which one are really malicious, which ones are not.

15:40And so what AI on the offensive side does is that it makes it, A, easier to do this. B, it means that you really have to react extremely fast. So even the time to react is something that a lot of organizations now want to bring down. That's where AI on the defensive side comes to play because it allows you to deal with all the signals. It allows you to do that much faster than relying on humans. humans still play a key role in our product in the sense that they are the ones that know the context about the organizations they are the ones that knows for example that i don't know something true for bank of america is less true for bnp paribas still these are two similar organizations in a sense the same industry but completely different realities completely different histories so you still need humans and humans are do bring a lot of value to the table but thanks to our product they're not dealing with incidents anymore they're dealing with facts we come to them saying we know that this is not a thesis.

16:36This is a real attack. And here is the pattern that we have seen up to you now. Do you want me to run remediation? Do you want me here's the suggested remediation that Kevlar would have run up to you? Do you want me to do it? Do you want to do it by yourself? Do you want to reconfirm everything? Here is the audit trail of the investigation. So here are all the steps that we have run, all the tools that we have connected to all the raw data, plus the interpretation from Kevlar that led us to believe that this is a malicious case. And so it's a complete game changer in the sense that you're not doing it by yourself the heavy load is taken away by the ai and it comes back to you with the results of the investigation and a good thing we are noticing more and more is our customers now entrust the product so much that they're asking for if not autonomous at least automated remediation as part of that meaning that they know that our accuracy level whenever kevlar says something is malicious or is not are probably higher than their own SOC.

17:32So now they really start to entrust us with decision making beyond the investigation side. And so that's where it gets really interesting because now we can close the loop much faster. Again, think that AI on the offensive side, if you were to reduce it to one major impact, that's the need for a much faster reaction. And so now what we are noticing is customers trust us on the investigation side. And so they're asking for, as I said, automated response to make sure that in a couple of minutes, three minutes, four minutes, not only is the investigation done, not only is the culprit detected, but also we have taken remediation actions, we have changed something in the system to harden it and make it even better for the next time.

18:12And so that's really what we're heading for. We do not want ourselves to be triage agents. We want ourselves to contribute to making the security posture of the organizations we help protect, making it much better and that's the idea of a self-healing security posture if you want. We get breached or at least we get an attack, we automatically adjust something and the AI can definitely help on this.

18:35Andreas Munk Holm:And to those that thought they heard a baby just before, you did. We have a security breach on Ahmed's side but that's being taken care of in the background. Now I want to ask both of you because what you're describing is a bit that the whole fortress model of security is that you have to accept that you cannot keep attackers out anymore. You have to assume that there is a breach or it's going to happen no matter what. What you now need to do is detect it and react as fast as possible. I want to ask both of you, is this already industry standard belief or is that something that you're still working with clients and other people you raise on your side with other founders to understand and see that this is where security is moving?

19:17I don't think it's universally accepted yet because you always have this issue of, oh, not to me, you know, others maybe. But the reality is that it's true. Like you really have to assume that nothing can be absolutely perfect. You can lock all the doors, they can get in through the window or the ventilation system or some other place. And I think if you have the attitude of thinking that you're absolutely impenetrable, then you're going to suffer. But if you do accept the fact that it's not a matter of if, it's a matter of when, then you're going to put in place the right things knowing that at some point, you know, someone's going to click on that phishing email.

19:56At some point, something's going to go through that other, you know, defense system. And therefore, you can have prepared for minimal, basically, cost to you by being prepared for detecting it, reacting to it, having trained people in advance. So I think that should be the ideal mindset for people. And the more attacks are actually put in the press, the more people are starting to take note of that and understanding that that's the new reality. But it's not perfectly universal yet. We have a slight bias here because we only work with enterprise accounts, right? Like, or large MSSPs. And when I speak with CISOs there, they clearly have understood it to be the case like the vast majority of them now know that again as you said andres this fortress model does not work anymore does not stand and that you should basically assume breach and if anything the new waves of ai publications on the security side and by this i mean clearly claude mythos and gpt 5.5 that's more on the vulnerability side of things but if anything it's showing that even on the vulnerability side you just cannot catch up You have millions.

21:08If you're an organization like, I don't know, a car manufacturer or a bank, you probably have millions of vulnerabilities everywhere. So you just need to know, you need to assume that these are all tools that an attacker can use. So the question becomes, okay, I cannot patch all of them. What do I do now? And that's where this role of AI being a sentinel, someone who constantly controls your shores, making sure that as soon as someone gets in, we're able to take them out. That's a model that probably summarizes better what the future of security would look like.

21:40Andreas Munk Holm:One thing with AI is, of course, that once you allow it to make decisions and become agentic, you're also assuming a bit of risk there. And I'd love to ask, how do you govern that risk and how much is yours versus the client's? Is it you who's the responsible party or is it the CISO who bought the product? That's a complex one. The way we deal with it is, first of all, we obviously try to increase as much as possible the accuracy rate of the tool, right? Like making sure that whenever it says something is not harmful, is benign, we'd better be absolutely certain about that. I'm citing this example in particular because everyone talks about false positives in security.

22:22but actually what matters the most is probably the false negative rate because that's really the one that harms you. That means that something has happened but has gone completely undetected until it's maybe too late. And so clearly we have a bias towards making our AI as conservative as possible meaning that it never says this is not dangerous unless it has checked every single possible case unless it has checked every possible tool it has at its disposal. It even knows when to raise its hand and to say, here, I don't think anything is happening, but because I didn't have access to the logs from system A and system B that I think would have been helpful for me, I will not conclude.

23:02Up to you to review it, up to you to add context, etc. And so the reason is, we really want to take this false negative rate as close to zero as possible, because that's the one that destroys trust with our customers. So essentially what I'm saying here is, the way we have designed the product, and we have a lot of graph AI, for example, that goes into it, It's not just an LLM that runs everything. We have designed it in a way to make sure that the false negative rate in particular is as close to zero as visible. But actually, we go beyond this. Basically, when it comes to remediation, giving the customer the ability to choose if they want to go with our remediation or not.