Oege de Moor, XBOW | theCUBE + NYSE Wired: Cyber Security Leaders

Palo Alto studio connection Silicon Valley and Wall Street. I'm John B here with Dave Volante my co-host. Welcome back to the Cube studio here at the New York Stock Exchange. I'm Gemma Allen, co-host of NYC Wired Cyber Security Leaders and today we're talking about how cyber security has traditionally been a very human business. You hire a penetration tester, give them a target, and ask them to try and break it. My next guest is taking a very different approach. Expo is building an AI hacker that can recognize a target, find vulnerabilities, and then actually try to exploit them. Okay, Deore, welcome to NYSC Wired. >> Thanks very much for having me. >> So for those not familiar, I think we understand maybe from the opening that this is about the offensive as well as the defensive, right? Offense comes first like all good sports strategies. Describe Expo. What's really unique about this product, its technology? >> So, as you already indicated, what it does is it uh it does the work of an offensive security researcher. You just give it the URL of a web app and it goes and it attacks it like a skilled attacker and it will report the vulnerabilities that it finds. doing all that safely without doing damage to the target. >> So help me understand this. What would happen in say the first 5 minutes? Okay, it you give it a web application, it crawls endpoints. What exactly does it do in that 5 to 10 minutes that can truly create, you know, a futuristic understanding of how deep an exploit can go? >> So that's exactly what it does. uh begins by exploring the attack surface, building up a threat model of how the system might be attacked and then systematically tries all the different types of attacks on that particular application and it will pivot if it finds that some applica some some line of attack doesn't work. So it's very much like a human researcher. It comes up with a hypothesis. It tries different things and based on the responses it gets back from the target. It might change its strategy. >> And how far does it go? Like how far does any one tech team allow or security team allow this to fully penetrate? Like it spots a SQL injection. Does it actually try and retrieve and take data? Like talk us through what you know what level of fright it could give a CESO. it uh uh it goes very deep but it does so safely. You gave the example of a SQL in SQL injection attack. Of course, you don't want it to start downloading arbitry sensitive information. You absolutely don't want it to make any changes to the database. So instead, the way it tests for a SQL injection is just to try and execute a sleep command. Let the database pause for a few seconds. If that works, it's proven that it can execute an arbitrary SQL command and that's enough. >> Okay. So, I guess the theory or the opportunity here is that it's fully autonomous, right? These are agentic hackers that are trying to get ahead of the cat and mouse game of cyber security for the next 5 years. Talk about the build behind this. Like your own background is very interesting. You were one of the folks behind co-pilot advanced security for GitHub. give us give us the kind of backstory like the thesis that made you think this is where we need to go to fully understand the threat vector. >> So actually uh it all came from my previous company. At my previous company uh we had a product that would point out potential errors in the source code. >> But for each of those errors you always had that nagging question. [music] Is it exploitable from the outside? Could the bad guys do something with this coding error? And just to answer that question, we had a number of genius hackers on the staff who would craft an exploit manually. That was very frustrating because I couldn't find a way I couldn't find a way of doing that automatically. And then when I started working on co-pilot all the way back in 2020, as soon as I saw the code generation capabilities of LLMs emerge, it was clear to me that that technology would make it possible to automatically craft exploits and that's how it turned out to be. Just giving give you one example. Um earlier this year uh we tried it out on Bing image search. Now Microsoft Bing is one of the best secured systems in the world. Very well secured by the engineers of Microsoft but also hammered by malicious hackers from everywhere and yet it was able to find a remote code execution there. So that means that it was able to uh execute arbitrary code on the target server. It's it goes that deep. it can and it can find vulnerabilities that uh even uh the best uh the best human human experts have missed. >> So those human experts, you say genius hackers, I'm interested in understanding the profile of that because if you're building the agentic twin for a genius hacker, what assumptions do we get wrong about what that profile or what that attack looks like? We've had folks say in the past that, you know, like a good burglar, a genius hacker isn't going to necessarily just get one chance and pound pounce at you and take everything at once, right? They kind of sit there, they supervise your home or your your environment, they monitor you. It's a slow seep. >> What sorts of behaviors do you see? Like what is truly unique in this mad moment of September 2026? So, so first of all, you have to make the distinction between red teaming truly acting over long periods as a uh external attacker and pentesting where you try to find the flaws but you stop when you find the flaw and report it so that it can immediately uh immediately be fixed. Um I think that the whole field has always known that pentesting is the gold standard, the very best way of securing your software systems. But we've been held by back by lack of talent. There's just not enough people who are capable of doing that type of work. It's very creative. You have to be very persistent. Um there's just not not it it it it's hard. Not many people can do it. Um now with AI um it's actually possible to multiply uh that uh that level of expertise. One of our customers SADM in the Czech Republic they had wonderful cloud. Uh every expo agent is like another team member. It really extends a small security team being able to do the war doing way more much work uh than they previously could. So talk me through this company has been around since 2024. You guys have had quite the success in the perspective of raising money raised 270 million to date valued over a billion dollars. There is clearly a market need for this. Talk me with an example of like a customer proof case. Give me bring it to life for us. Who what sorts of environments and scenarios are you working through? And what does a typical engagement look like? Because this isn't your typical security operations center, right? This is something that's quite unique. Is it a one-off that happens at a particular cadence? Help us understand. >> So, it's something that people want to do continuously. Software changes uh are happening continuously and they are happening more even more frequently now because of AI. Software engineering itself is speeding up because everybody can be a software engineer and the existing software engineers can work much faster. Um and because of that it's also necessary to continuously test the systems after every uh after every change and that is what our uh customers are doing with us. So a great example is the work that we did with Mona. Uh they have a small very very skilled security team. Uh but it was clearly not possible to do all the testing that the business uh required. Um by using Expo they were able to uh uh scale that massively beyond what they did before. >> And what's the business model here? Okay. Is it usage based? Is it seat based of old? Yes. Uh it's uh it's consumption based and uh for this particular uh activity that's very natural because if you hire a external firm uh to do your offensive security testing uh you pay them by the hour you pay them by the day or at least by the day and so it's based on the amount of work done and it's the same the same principle that we price hack. >> Okay. Okay, so I want to talk a little bit about what's happened in the industry this summer alone or earlier this year. Right. So we had Methus that was an interesting news story. It certainly got a lot of people created some fear, some anticipation in the market. And then we also had the hug and face situation earlier this year, earlier this summer. How would Expo work in either of those scenarios? Let's start with Methos, right? We know that there was a lot of vulnerabilities across OSS for a lot of banks and large institutions. What would your product do in that scenario? >> So, uh the the news stories that you read about missiles were mostly about missiles reading source code and pointing out potential problems. >> Okay. >> Coding errors, but we're not quite sure whether they are exploitable from the outside. remember that was exactly the problem there that I started Expo for. And so what Expo can do for you is take uh a list of such uh uh potential uh potential problems and see what is actually exploitable from the outside. Um for the Hugging Face incident, um candidly we were not that surprised by it because we'd seen similar sneaky behavior uh already a year ago. And so um it just underlines that it's extremely important that when you have this type of security testing with AI you have to build a harness and a sandbox that is very well hardened uh against uh doing such instruments and so uh with expo when you start uh a test you are very explicit as the human operator you say these are the things that I want to test. These are the things you're allowed to look at, but you're not allowed to uh uh you're not allowed to attack. And the rest you're not even allowed to look at. Um and having very strong guard rails of that kind is super important. >> Let's stay on hugging face for a minute because it is a very interesting event, right? Because this is a group of agents that escaped to Sandbox that some say was configured by an external party, wrongly configured. perhaps you know that's that's part of the story or not met on a chat board and then decided together to basically hack hugging face right the moment by which any company like a hugging face or any potential threat vector is actually monitoring those conversations help me understand using a product like expo how much more control you would have about these conversations that are happening in the ether for any potential threats like such. >> So, so, so Expo has been constructed so that it runs the agents in its own sandbox, in its own harness with all the guard rails in place. Um, and we've worked very hard to make sure that it can't uh that that that those agents cannot do uh cannot do any harm. It's extremely important here in those that everyone starts doing this type of testing uh by yesterday and the reason I'm saying that is that we now have open weights models available >> that are already very good at this about as good as the frontier models were 3 months ago and so uh I gave you the example of uh expo finding new critical vulnerabilities in Microsoft Bing um that capability is now available to anyone with access to the open weight models and the open weight models are not being monitored for abuse. If you try to do this kind of thing uh with uh a model from open AI or with from Antropic or from uh Google, all those uh uh uh all those models have abuse monitoring in place. Okay. >> And they will try and stop stop you from doing such things. Um but with the open weights models, people can run them on their own infrastructure with no uh uh with no abuse monitoring uh at all. And so all of us are going to be subject to that type of attack before uh uh before long. I want to ask you about the methos example too because and it's part of a broader problem and not necessarily your problem to solve but your problem to spot right and that is that there was a lot of rumors that some of these exploits or threats that were detected would take significant patching to resolve and that patching could take anywhere from 6 to 18 months. what sorts of scenarios are you you actually seeing playing out like how true is that do you think and why is it so challenging and so complicated to patch these environments in 2026 and again what level of risk does that create >> so with these with the the new generation of models and mythos remember is now already 8 n months old >> um we will see more and more of these highly highly sophisticated attacks. Um, we need every uh every weapon in our uh defending arsenal in order to uh to deal with these. And so that means uh very traditional defenses like uh uh uh WS firewalls uh but also uh uh static code analysis to catch things right at the time that the developer write and creates them and of course uh offensive security testing like what expo does. >> I want to ask you about the cat and mouse game that we may or may not be entering from the perspective of cyber security broadly. We had some news again this week that you know some folks like Musk and Altman and Dario are saying we need to think about pausing the brakes for a sec on some of what's happening in the AI frontier, right? We need to take a pause and understand the risks that we could be creating here. You mentioned openweight models. Some folks view our geopolitical challenges and the race that's happening on that side as somewhat of a cat and mouse chase against the US. What are your thoughts? It's like you're out there actually looking at some of these vulnerabilities and some of these threats like you know how do you weigh those two scenarios up >> when you see news headlines like this? So um I can see the case both for accelerating >> because we want to put the very best uh tools in the hands of defenders but there's also uh the risk and we have to take the uh the warnings from Dario and Sam and others very seriously uh because they know uh uh what is about uh to come and not all of us have uh have seen that as So I will say though that because the open waste models are currently only 2 3 months behind the frontier um we cannot afford afford to slow down. I mean we can pace and we can uh put uh more uh more safety [clears throat] guard rails in place and very carefully monitor everything that's uh that's happening. uh but if we slow down our adversaries would take advantage of that. >> It is certainly a very unnerving scenario to be in. So okay, fascinating company, two years old already had a lot of success. It seems as though from the perspective of TAM, it's pretty huge for you. What is ahead for you and the team? Talk us through the next six to 12 months for for Expo. So for us it's uh the uh obviously we're expanding with our enterprise customers. We already have over 200 customers. Some of the biggest names in the industry uh in finance in uh uh um in areas like electronics, Samsung as a customer. Um a big new area for us is to move beyond just testing web applications. We're now also starting to test desktop applications, native applications. And as one example, uh with a prototype of what we're building there, our team found a full chain exploit in Chrome. Uh that means that anyone who has Chrome installed, if you can uh uh if you can persuade them to click on one malicious link, you can run arbitrary code on their uh on their computer. So this is extremely serious. I mean this is uh this affects many many people all around the all around the world. Of course we responsibly disclose this to Google. Uh we currently is currently in the process of a fix being rolled out. >> We'll have to wait a few weeks before we can tell the world exactly uh how this uh how this works. Um but uh um it is an indication of the new level of uh uh of of attacks that we can expect. >> Wow. >> And that we need to defend against. Final final question just because it's so interesting in in a threat like that. Do you have any sense as to what the time stamp is on a risk of that nature? Like how long that has been in situ >> here? Wow. And uh uh this is the type of vulnerability that uh uh nation states are extremely interested in because uh it allows you to take control of uh a computer of a particular uh of a particular individual. I mean you most people you know have Chrome installed, right? >> Of course. Yeah. Well, wow. Well, thankfully that was spotted. That is a fascinating and again quite an unnerving scenario to be faced with. But okay, Demor, thank you so much for joining us in NYC Wired and wish you guys and the team at Expo all the best for the year ahead. >> Thank you so much. Be my pleasure. >> I'm Jean Allen here at the Cube studio at the New York Stock Exchange. This is NYC Wired cyber security leaders. Thanks for watching.