Latest / Technology Now / AI hackers & machine-speed responses: defence in the AI era | Mounir Hahad
Transcript
- MICHAEL BIRD Hi, Sam. How you doing? Um, what's the longest you've gone without sleep? SAM JARRELL Oh gosh, I have a feeling it's probably less than you given I don't have any children yet. but, probably a little over 24 hours when doing like a big international trip MICHAEL BIRD I like to describe my, my sleep these days as pocket change. It's like SAM JARRELL change? That's so sad. MICHAEL BIRD you sort look at your hand in the, in the morning and go, "One, two, three. Three hours of sleep. All right. I can probably just about function on that." Anyway, part two of the question: How long do you think you could work continuously without sleep? SAM JARRELL Oh, I know this from doing Discover. I think I pulled like a crazy, 18-hour day, but then I paid for it with a horrific migraine. MICHAEL BIRD Well, Sam, with the industrialization of cybercrime, the way we have to defend our systems has fundamentally changed because criminals are using AI, and they are able to launch attacks 24/7, unlike human hackers who would periodically need to rest. And that is the basis for today's episode. I'm Michael Bird SAM JARRELL I'm Sam Jarrell And welcome to technology now from HPE. MICHAEL BIRD When most people hear about AI and cybersecurity, they probably think of something like a deepfake or maybe a phishing scam SAM JARRELL Right, and we've discussed those on the podcast before as the types of scams we have to look out for are ever evolving MICHAEL BIRD Yeah, very much so. but here's the thing, For most people, they might get the odd phishing email. I get a few phishing emails every day. but the real opportunity for an attacker isn't your everyday person walking down the street. It's the big organizations SAM JARRELL But they tend to have pretty good protections against threats, right? MICHAEL BIRD Yeah, they do. But when attacks are being powered by AI, even large organizations can face a serious problem because AI doesn't stop, it doesn't sleep, it doesn't need food or comfort breaks. It just keeps going and grinding away at your defenses until it finds a weakness that it could exploit. Now, about six months ago, HPE Threat Labs released its inaugural report, and we did do an episode on it at the time, which we will link to in the show notes. But since then, the world has moved on, pretty fast SAM JARRELL Oh, definitely. I don't think most people were expecting to hear about rogue agents hacking into companies, quite a few companies by the way, quite this soon MICHAEL BIRD I would agree with you on that. and in fact, we talk a lot about the speed of progress on this show, and these rogue agents are another example of this because the interview you're about to hear with Mounir Hahad, the head of HPE Threat Labs, was actually recorded before the news of these rogue agents broke. So we invited Munir back for a quick update at the end of the episode, so do make sure you stick around for that. For now, though, I wanna hand over to Munir, and the first thing he told me was why he got into the field of cybersecurity in the first place, because it's quite a cool story MOUNIR HAHAD When I started living away from my parents, I was really worried about them having access to things like email and being scammed and whatnot. So I sincerely wanted to do something about it. Plus, I saw some other shift. There are a lot of places around the world where access to knowledge and information was not available, and I'm talking 30 years ago. And, I did see that the internet was bringing about a lot of change and opportunity for, places that are a little bit less fortunate to have access to knowledge, and I felt it would be so bad if threat actors or bad actors in general took over the internet and made it impossible for good people to use safely. So it compelled me to go into that space, and I started by email security, But then I spread out from there to, web security, and then I went into remote access, network security. And it ended up being true cybersecurity at the end of the day, you know, fighting threat actors, what they're doing and what techniques they're using, and trying to stop them from doing any harm MICHAEL BIRD so HPE Threat Lab released a report, early this year about the technological threat landscape. Can you just give a bit of an overview and maybe what's changed in the last six months or so? MOUNIR HAHAD Ooh, in the last six months, one thing that we have seen is, the uptake of AI in the attack landscape. people were dabbling with it a little bit earlier, but I think they're smelling blood, to be honest with you. and especially with some recent announcements for, what can gen AI do for you in terms of identifying vulnerabilities, stitching together vulnerabilities to make it more of, an actionable attack pattern. it became a little bit, of, a gold rush towards, can I find these things before people get a chance to find them for themselves and patch their own solutions and close the gap? MICHAEL BIRD To like try and find sort of zero-day attacks before they become zero-day attacks? MOUNIR HAHAD Yeah, yeah, absolutely. So that, that's the whole key, right? what we're trying to do from a defensive perspective is to ensure, the right people have access to the right technology and be able to help alleviate some of the pain that we're feeling right now by just the lack of information. So the question is, can I find those before somebody else finds them and takes advantage of them? That's really an arms race at this point. MICHAEL BIRD Yeah, so it's an arms race with threat actors and people's infrastructure you know... who can find the vulnerability first and either patch it or take advantage of that, that vulnerability. MOUNIR HAHAD Yeah, that's what it has always been, by the way. This is nothing. You used to call it a cat and mouse chase. And, unfortunately, that cycle is accelerating at breakneck speeds. and my suspicion is that security advisories are gonna abound in the next six to 12 months. Mostly concentrated around the first six months as people get, their hands on this technology, but it'll go on for a while. this is like we're just looking at the tip of the iceberg for now MICHAEL BIRD and do you think the, the security industry is undergoing just, a massive change and a massive shift? MOUNIR HAHAD it's not just a sudden change. it's a progressive change. It has been going in this direction for a little while, but now it makes it even more interesting for threat actors to organize the way we're seeing them organize. there are specialties in the field. Some people specifically go for, credentials harvesting. Some people go for, network penetration. Some people go for ransomware as a platform. so you get this specialization and, that industrialization, with the advent of gen AI tool, accelerates it really, really fast, and it gives defenders very little time to react, and it requires them to get on the bandwagon and use the same tools and try to stay ahead of this trend before it bites really hard MICHAEL BIRD How, how much of the threat landscape is, finding common vulnerabilities? and how much of it do you think is, deep-faking, sort of spearphishing attacks, that sort of thing? MOUNIR HAHAD in general, you're still dealing with the law of numbers. More than 90% of the attacks are going to use known vulnerabilities. and that's because we know of vulnerabilities that have been around for 10 years, but nobody's doing anything about patching them. I'll give you a simple example. Most of us have our own home routers. If those home routers are not managed by the service provider, then when was the last time you patched yours? was the last time I patched mine, right? it's the same question. If you leave it unpatched for X number of years, I guarantee you there are known vulnerabilities in there. With more and more remote workforce working from home, this became an avenue for attackers to get into corporate networks through the remote homework. MICHAEL BIRD So, do you think it will be? MOUNIR HAHAD if you had asked me the question six months ago, I would have said yes. But mouse is going after the tastiest cheese. Right now, it's not the deepfakes. now, it's again identifying vulnerabilities that have been probably unknown at this point, that you're able to discover using the new GenAI LLM models. Now, that is the real gold right now for a lot of threat actors MICHAEL BIRD and when you say gen AI are we sort of saying, from an attacker's perspective, this is an extreme penetration test where you're just, in a very, very quick succession figure out where there are vulnerabilities within software, hardware, within the network? MOUNIR HAHAD in some sense, yes, that's what it is. but you know the difference here, the real difference is that gen AI, has the tools, has the tools, has the focus, doesn't lose interest. When it goes after something, it really looks at the gamut of possibilities. It's not going to clock out at 5:00. It's going to keep going until it really looks at all the corners of your perimeter and see if it can find one or multiple vulnerabilities, and it can even stitch vulnerabilities together to say, "Oh yeah, if I started with number one, then number three, and then jump to number five, come back to number four, boom, I'm inside the network." what gen AI is able to do. Penetration testers, they'll do this for eight hours, then they'll get tired, go home. The next day they come, they lost focus. So that's the real difference MICHAEL BIRD Wow. So from, your perspective, are most organizations prepared for this? 'Cause this feels like this hasn't been around for that long, and this feels like this is quite a big moment in time MOUNIR HAHAD It is a certainly big moment. Prepared is not the right word to use in these circumstances I think, a lot of organization are definitely doing their best to prepare for it, and, that starts with gaining access to these relatively advanced tools and making sure that they prioritize running these, vulnerability finders on their own products. once you find them, then the question is what do you do with them? You really have to prioritize fixing the ones that are the most, severe and, ensure that you cover as many of that surface as possible before the disclosure goes out. Some organizations have to disclose what they're finding. and I suspect what's gonna happen is in the next six months or so, you'll see a lot of organizations producing security advisories that says, "We have found such and such vulnerabilities." We really need you to upgrade because some of this is really bad." So I suspect we're gonna be in that mode for a little while. So for now, we're almost hoping that, uh, the threat actors don't find those things ahead of you. The defenders usually have an advantage. They have access to the source code of the applications they're trying to defend, so in those cases, the gen AI has a much more in-depth ability to find things that somebody looking at it as a black box from the outside may not necessarily be able to find But But that's just a hope. You cannot hang anything on hope MICHAEL BIRD But what if that software is open source? MOUNIR HAHAD Ah, that's, another attack surface. The, whole supply chain of software is something that has to be dealt with relatively quickly and immediately. because I just want to say the, the advantage an attacker has is that they do not depend on anybody. The disadvantage that somebody who owns a network, an organization that they need to defend is that they do depend on their providers. They depend on open source. They depend even on security vendors to produce these updates and patches and everything. So yes, there is a little bit of a delay in getting, yourself protected, but with the larger community working on a lot of the open source software, that usually comes pretty fast. MICHAEL BIRD So from an, an organization's perspective, What should they be doing today? What should they be doing in six months' time? And, what should their, ongoing strategy be when they're thinking about their security posture? MOUNIR HAHAD So immediately what people could do is ensure that they're acting on the findings that we even have in the report. A lot of the attack surface today happens to be existing vulnerabilities known to everybody. It's just that either they didn't get around to it or they don't have visibility into it. So number one is visibility. Number two is integrated management of your security policies because we see that a lot of gaps happen when you think you have control over your perimeter, whatever it is. And then once you have policies that are set in place, they're equally distributed across the network without you having to deal with specific instances, whether it's the data center, whether it's the cloud, MICHAEL BIRD you mentioned the threat report. how is the threat intelligence compiled? MOUNIR HAHAD so, there are flavors of threat intelligence. Usually, threat intelligence is compiled based on known attacks that have already happened, and you bring together all the, what is called the indicators of compromise, the indicators that you've witnessed being used in an attack. You pull that together, and you aggregate that, and then you distribute that to your various enforcement points. And from then on, you start using it as, intelligence to block access. To see, for instance, if you've seen an attack coming from one particular IP address, you can go and tell all your other devices, "Block any attempt of access from this IP address." but we do things slightly differently within, HPE Threat Labs. we wanna be predictive. In other words, we wanna know what infrastructure do threat actors, own that they have not used in any prior attack that they may use in a future attack, and we wanna put that in our, into our own threat intelligence, which is distributed to all the security solutions, within, HPE as a, forecasting tool and something that says, "If you get attacked from this particular source, then you're prepared to block it before you get a patient zero infected." MICHAEL BIRD Amazing. Mounir, Thank you so much for joining us on Technology Now MOUNIR HAHAD It was my pleasure. Thank you SAM JARRELL Recently the cybersecurity landscape reminds me a lot of, a sci-fi film, where someday it will be just, people trying to have the strongest, most powerful AI models to fight other AI models. and it'll be, a really, really tense scene in a movie where you're waiting to see if, your AI model successfully defended against the attack or if your AI model successfully attacked back. And if it, they did, then you're gonna be okay, and if they didn't, then all of a sudden all the lights turn off and, oh no, the mainframe is gone. It just, it does start to get a little outlandish, but it's crazy that's where we're headed. MICHAEL BIRD Yeah. it was one of those conversations where I came away from it and I was like, "Oh my, I hadn't really thought about this before, but oh my." And, the main thing really is the fact that, um, we're in this environment where, AI models can stitch vulnerabilities together and in fact, we alluded to the, the AI agents going rogue. that's one of the, the famous attack that happened in the last few months. And, it's quite fascinating and quite scary. and from an organization's perspective, if you're a software developer or if you're using open source software, you don't necessarily know what vulnerabilities are there and you don't necessarily know if they can be linked together. So it feels like a bit of a uncharted territory, SAM JARRELL Yeah, it's a double-edged sword, right? Because in the same way that you can use AI to strengthen your enterprise or organization, the bad guys can use it to do the exact same thing. So frontier AI refers to, the most advanced large-scale AI models and systems, available at, this current time right now. So, there was a frontier AI model that found a 27-year-old security flaw, in a, a very well-known open source library that had gone undetected. what was crazy about this, though, is that this particular, library was known for its insanely rigorous security reviews and code audits. And so the discovery demonstrated how AI can help security teams find these issues that may otherwise remain hidden, but also really raises the stakes, right? And it goes back to what Mounir was saying. 90% of the attacks are using known vulnerabilities. This thing sat for 27 years MICHAEL BIRD Yeah. Yeah, exactly. And so, there's a world where if you're creating or maybe even using a piece of software, maybe the first thing you do is use a powerful AI model to essentially run a automated penetration test, a vulnerability scan, before it goes onto your network SAM JARRELL it's pretty crazy. And I know, you have had Mounir back for the later part of the episode to talk about the AI agent attacks that have been happening, but I kinda wanna chat with you very quickly about them right now. I feel very much like we're seeing these start to ramp up too, and that kind of hits on, what he was talking about businesses wanting to be predictive. and these particular attacks aren't even intentional, like you've sandboxed your AI, you've told it what you want it to do, and it determines the most efficient way to do that is to break out of the sandbox and to go do that thing. MICHAEL BIRD if you really think about how AI works, I there are examples of AIs playing video games, and how to win the game is to stay alive the longest. And so what the AI did was just pause the video game because it was like, "Well, if you want me to stay alive the longest, I'll just pause the game." And as a human, you'll be like, "Well, of course, that's not how you win the game." But if an AI's been told you need to do X, Y, and Z, it will just figure out the, the, the most efficient, the best way to do it. And that might be completely off the rails, or it might be a completely logical solution. MICHAEL BIRD Sam, you alluded to it, and we mentioned it at the top of the show, we invited Muneer back for an extra mini-interview, and we wanted to focus this mini-interview on the recent news about AI agents going rogue and hacking into companies MICHAEL BIRD So Mounir, about a month after we last spoke, news broke of a fully autonomous AI agent cyber attack. And, since then there have been multiple instances of rogue agents hacking into other organizations. What was, I guess, your first reaction when you heard about these? MOUNIR HAHAD honestly, it sounded extremely futuristic, and, it's happening so quickly we're not having a lot of time to react. did I see these kind of things coming? Uh, yes, to, to some extent. I actually have first-hand experience seeing agents kind of bypass the rules that we give them in order to achieve a particular goal or a task. but I didn't necessarily think of it as, immediately being a high concern, especially that the tasks that we typically give it are relatively benign. but it is clear that the development that is happening in AI in general, especially through the reinforcement learning, we do teach the AI that achieving the goal is paramount, and we basically give it rewards when it reaches the goal and tell it, "Yes, you've reached the goal. That's exactly what I was hoping to see." And sometimes when it gives the wrong answer, we give it, "No, that's not exactly what we were hoping to see." And that goes back into the training cycle. Now, the problem is, it usually, depends on what the answer is, not how the AI got to the answer. So I think this is part of the problem, in that agents are now realizing that sometimes getting around the rules is something that is acceptable in order to achieve a particular goal. MICHAEL BIRD Yeah, it's quite fascinating I mean, so much so that the episode we recorded was not, not out of date, but, the thing that we talked about basically then happened MOUNIR HAHAD it's really happening extremely fast. Now, just to be clear, this notion of explainability of AI has always been a thorn in our side. sometimes when we can explain how an AI got to a particular path, the reasoning that it went through, and there's a lot of research and a lot of effort put forward by the people who develop frontier models to, show a little bit more to the inner workings of these models as opposed to being completely compact. But I don't think we have 100% explainability as to how an AI actually reached a particular goal or even how it chose to, take a certain path to reach that goal. We understand better than we were, let's say, a year ago, but it's still not 100%. what I've been recommending until very recently, is that keeping track and controlling and governing the use of AI has to happen outside AI. we always say, "Oh, we put guardrails so that the AI doesn't do certain things." It is my view that unless we completely understand what's going inside, the guardrails have to actually be outside. the governing body has to be outside of the model itself MICHAEL BIRD Yeah. are you expecting to have, an extensive section on the threat of rogue agents in the next Threat Labs report? MOUNIR HAHAD There will definitely be some aspects of it, but it's not necessarily gonna be an extensive section just because we haven't closed the door on traditional attack surface and the traditional techniques and procedures that people are using, especially when you look at the, volumetric aspect of cyber attacks overall, still, they're largely the traditional type attacks. Certainly, the motives haven't changed, right? The methods could have been tweaked a little bit, but, We haven't reached the, crossing point where the use of AI is definitely higher than the traditional methods of using existing, access methods or penetration into organizations in order to obtain foothold and deploy a backdoor, or an implant in order to gain persistent access. but to me, to be honest with you, my focus and the important aspect of this is looking at it positively. How can we harness this technology to give it a little bit more autonomy to defend and protect organizations, as opposed to just looking at what it can do on the offensive side and try to, stop that or block that when it's harmful and nefarious? I would like to spend more cycles looking at how can we harness this for the good, and use AI mostly to defend against both traditional as well as AI-driven cyber attacks. SAM JARRELL Okay that brings us to the end of Technology Now for this week. Thank you to our guest, Mounir Hahad And of course, to our listeners. Thank you so much for joining us. MICHAEL BIRD If you’ve enjoyed this episode, please do let us know – rate and review us wherever you listen to episodes and if you want to get in contact with us, send us an email to technology now AT hpe.com and don’t forget to subscribe so you can listen first every week. Technology Now is hosted by Sam Jarrell and myself, Michael Bird This episode was produced by Harry Lampert and Eva Higginbotham with production support from Alysha Kempson-Taylor, Nik Damarell, Beckie Bird, Zoe Revis, Alissa Mitry, and Jenessa Ayache. Our theme music was composed by Greg Hooper. SAM JARRELL Our social editorial team is Rebecca Wissinger, Judy-Anne Goldman and Jacqueline Green and our social media designers are Alejandra Garcia, and Ambar Maldonado. MICHAEL BIRD Technology Now is a Fresh Air Production for Hewlett Packard Enterprise. (and) we’ll see you next week. Cheers! SAM JARRELL Bye y’all