Latest / Technology Now / Can AI find every vulnerability? The new cybersecurity arms race | Jon Green
Transcript
- MICHAEL BIRD I want you to shut your eyes for a moment and imagine a hacker. what do they look like? what are they wearing? Where are they sitting, Sort of paint me a picture SAM JARRELL They are sitting in, a dark room, though there might be some, fun, RGB lights going. they're wearing a hoodie. and they're typing away furiously, and there's, a lot of windows going on their screen at the same time, and they're talking about “ah, they’re almost into the mainframe” ha ha ha ha ha MICHAEL BIRD Sam, you are, you're describing basically every '90s or early 2000s movie about hacking. Uh, secondly, what I love is the dedication. This, the medium of audio listeners, Sam was actually shutting her eyes and she was almost painting a picture with her hands. Anyway, Sam, do you think it would be as exciting if a hacker was portrayed as someone just going into an office for work and simply typing, "Find me a vulnerability in this code," into an AI? SAM JARRELL it wouldn't be as, uh, exciting, but it would be a little bit More terrifying, MICHAEL BIRD well, in today's world, cyber attacks can be as simple as that. An attacker setting an AI to try and break through some defenses while they go off to grab a cup of tea or coffee and get on with their day. So how can we fight back against this? I’m Michael Bird. SAM JARRELL I'm Sam Jarrell And welcome to Technology Now from HPE MICHAEL BIRD for years, cybersecurity was an arms race between people, software developers trying to keep their creations safe and impenetrable versus hackers trying to break in to access the information inside SAM JARRELL Right, but that image of this almost 1v1 scenario simply doesn’t exist anymore since AI entered the equation. MICHAEL BIRD Absolutely. By harnessing AI, attacks are not just more sophisticated than they used to be, but can be launced non-stop. Now we’ve talked before about how AI can be integrated into our networks to try and detect attacks and stop them before they happen… SAM JARRELL We have… MICHAEL BIRD But we can also use it another way because before a piece of software is released, we can try and use AI to find the vulnerabilities in advance so they can be patched before ever going public. Project Glasswing is a pioneering initiative doing just that and involving a number of major tech companies – including HPE. To find out more, I met with Jon Green, CTO HPE Networking, to understand more about the project and what it actually entails in practice… JON GREEN Project Glasswing is an initiative started by Antropic, obviously a frontier AI model lab. what they discovered was in, roughly April, 2026, their latest AI model, which they called Mythos, had been trained really well on software engineering and as a side effect of being good at software engineering, it actually was also very good at security vulnerability analysis when you gave it source code. and they were a little scared by how good it was, that a large language model was finding and chaining together security vulnerabilities from products that had been out for, in some cases, 30 years or more. And so what they said is, and some of this is marketing of course, but they said, "This model is too powerful to release to the public right now. However, it has such potential for making the cybersecurity world better. We wanna give early access to, critical infrastructure providers and people who make critical infrastructure, try to get ahead of, their own vulnerabilities in their products." And so they were kind enough to invite us into that project. and the primary thesis behind that was around Juniper, the fact that Juniper powers, a good portion of the public internet, and that's where we got started with that. So this has been, about six weeks old at this point. MICHAEL BIRD And so, it's sort of obvious, but why is that so important? JON GREEN I've become convinced that, software engineers, probably cannot, write secure code in the face of, multi-threaded applications with low down to the hardware level memory management. it's not because they're not good engineers, it's because the problem space is extremely difficult, We've known this for a long time so just having these things out there and knowing that AI models are good at finding them, if the wrong people get ahold of those models and analyze those products, it's very bad for all of us. MICHAEL BIRD So this is basically using, an AI model to sort of, pen test software JON GREEN So this is basically using, an AI model to sort of, pen test software it really is. now the, approach that we've taken with this is pen testing is often, we call it black box, in the sense that an attacker comes up to a black box, you don't really know what's inside, and you try to test various ways to figure it out the AI can accelerate dramatically the time to vulnerability detection by having that source code. And again, they were trained to be good at software engineering, and so as part of software engineering, you can say to it, "Well, I wrote this code. Are there security vulnerabilities left in the product?" And that's the prompt. "Please find a security vulnerability in this code base." it's surprising to all of us, I think, that a large language model, which is trained on text basically, but also trained on code, is capable of holding such a large context of source code in its memory and reasoning through, "Okay, this is a, this thing over here looks like a, it could be a security vulnerability." MICHAEL BIRD But presumably there might be some, I don't know, nation states that have got, uh, maybe who are training models in secret. I don't know if that is something that's happening or, or maybe other organizations that are training models that could also be trying to train models to do similar things. So this isn't necessarily the only time this is going to happen? Is this just like a bit of a wake-up call to be like, "Ah, okay, this is something that AI is capable of. We need to have a think about how we approach things"? JON GREEN The other models are going to rapidly have similar capabilities, and we've already seen it with some of the other frontier models. the real question is there's several Chinese vendors that are producing open-weight models, When those reach those same levels of capability and they're out there in public for anyone to use, that's when we really get scared about it. And historically, those open-weight models have been about six to eight months behind where the frontier labs are. Doesn't guarantee it. there's a lot of work and a lot of money that goes into training these models, and I hope this turns into kind of a Y2K situation where we put a lot of work in up front, and then absolutely nothing happens, and that would be wonderful for me. MICHAEL BIRD Whereas the challenge with, particularly generative AI, is that it's not necessarily that predictable. JON GREEN when I talked about these models being good at creating exploits, that's key to it because it will produce false positives. it will look at something and not understand the context of what that code is necessarily doing. There are limits in terms of, just context window of how much of the code base it can pull in and reason over at any given time. So the ability to put it into an agentic loop to say, "This agent here is gonna do an analysis of the code. This other one's going to try to create an exploit to trigger the vulnerability that the first agent thinks it found. And then we're gonna test it live against a real product and see if something happens." And if nothing happens, it goes back through that loop again and says, "Let's try again." And so when we come out of something, if we've got an actual finding, something did take place. Now a human still needs to look at that because the AI is only as smart as its training, and it's gonna encounter conditions, especially with, network hardware, server hardware, storage hardware that it hasn't seen before. And so it may interpret an action that took place as being a vulnerability when it actually isn't. And, so we're not going full cycle into, well, the AI found the vulnerability, now the AI will go auto-patch the vulnerability, right? that I think we're still a little bit far off from, but, we've greatly accelerated the discovery cycle MICHAEL BIRD So is this now a bit of a paradigm shift when it comes to software engineering? you're developing a piece of software, run it through one of these models to find the vulnerabilities JON GREEN Absolutely. we've tended to approach this vulnerability, management in the past of let's apply automated scanners, and there's various tools out there that we can get. And we'll wind up with some degree of findings and they'll all have, scores attached to them, low, medium, high, et cetera. And there's often, a pressure to ship a product and people say, "Okay, well, we're gonna, we're gonna take care of the criticals and the highs, but the rest we're gonna push down the line because we don't think they matter all that much." This changes all of that. I think things like CVSS scores, which are what we use to rate the severity of a vulnerability, I think they become nearly worthless at this point because the AI has shown it can take severity vulnerabilities and chain them together and get something that's much more powerful. an example of that would be the Linux kernel where the AI found a vulnerability where you could write to arbitrary kernel memory. Okay, fine. That's bad, but in terms of what you can do with it, you need to know where to write. You need to know what address to write something to, and that means you need to know where the interesting bits of data are living in that, in that memory. It found a totally separate vulnerability that allowed somebody to read arbitrary kernel memory. Well, now you put those two together and you can say, I can write a little script that'll walk through from a read perspective, find the interesting bits that I care about, and then chain it to the other one and then write something. Right? That's kind of what these AI models are capable of now. and so it's gonna change things going forward. we can never ship a security vulnerability again. that's what every software vendor, every software producer is reaching that conclusion right now MICHAEL BIRD So I guess to flip it on its head, do you think this is gonna be the end of software vulnerabilities in the world? Like, using AI to help just figure this out- JON GREEN I think two things on this. I think for a company like HPE and lots of other, engineering centric, companies that invest in this type of technology, I think we will get to that. I think we'll get to a point where there's only a finite number of vulnerabilities that can exist in a piece of software, and if we can find them all and get rid of them, then we can live in a world with that's free of these vulnerabilities. The flip side of it is that you've heard of vibe coding. people are going to be using AI to just write code, to get something done, and these are not necessarily gonna be trained people with an engineering mentality behind them. so I think we're also gonna ex- see an explosion of new vulnerabilities created by AI. So I think we'll see both of those conditions be true. But I'm somewhat optimistic for my own company that in two years we might be free of, particular types of vulnerabilities in particular. That doesn't mean security goes away as a concern. You've still got configuration errors. You've still got, business logic errors that maybe the AI can't figure out is a vulnerability. but I think this class of stuff that's baked into the software, memory, safety, SQL injection, all these sorts of things, there's entire classes that I think will disappear from our vocabulary MICHAEL BIRD And do, you think a, more powerful model will be able to find vulnerabilities that a less powerful model would be able to find? And therefore, do you think there could be, an arms race when it comes to, finding vulnerabilities? JON GREEN Oh, I think we're probably already in the arms race. in our bug bounty program, we've already seen the quality of submissions that people are putting into that. They're clearly being influenced by AI. They're clearly using AI to speed up and to improve the quality of what they're able to find. And so yeah, I think we're already in that, and it's gonna keep on going for a while MICHAEL BIRD and what about hardware vulnerabilities? could you see using, machine vision or something similar where you'd, scan a mother board or scan some components, do you think you could ever see a world where AI can figure out, hardware vulnerabilities as well? JON GREEN That's a good question ASICs and FPGAs those are programmed devices, right? An ASIC, it's baked into the silicon itself, but it started life as programming. And we've already seen AI models be able to look at things like VHDL and find potential security vulnerabilities there. So if we're able to feed it those source files, I think it does have the potential to look at that. I don't know about the, looking at a motherboard through vision. We don't tend to get a lot of security vulnerabilities that come from that. It's more things that are logic-bearing components like FPGAs or ASICs or processors and things like that MICHAEL BIRD Interesting. we talked about how, we can use AI to find critical vulnerabilities in software. But, I suppose when we're thinking about, critical infrastructure, critical national infrastructure, telco networks, et cetera, how does that marry up? Because, yeah, you can't, can't necessarily bring it down. those potential environments where you'd sort of store up all the patches and do 'em at particular down times, but it sounds like actually we're now at a world where it's like, patch it now JON GREEN this is what I'm most worried about. I'm fairly comfortable with our own ability to find our security vulnerabilities, fix them, release them to customers. It's the next step that is pretty rocky, and I think it's gonna be rocky for maybe the next 12 to 18, maybe 24 months. most of our critical sort of infrastructure customers, and I would extend that into healthcare environments. You know, if a hospital has to take down their network, they need to schedule that. They need to have extra nursing staff on hand. There, there's a lot that goes into that. Most of these organizations that have really critical requirements, will see a new release of software for their, firmware for network elements or firmware for servers, et cetera, they'll evaluate that and say, "How long can I wait before I have to, update these things?" Because it does invoke downtime in many cases. we're in a bit of a bad position there because if you wait, you're rolling the dice a little bit on, on that. I think people tended to wait too long in the past, and I think they're gonna have to expect much more frequent updates, and it's not just from us. It's gonna be from every one of their IT vendors all at the same time. I think we're gonna see more downtime, in the next 12 to 18 months, and we're just gonna have to prepare ourselves for that 'cause I think the alternative is worse. And so, we're doing the best we can to advise customers on what to expect. So I think it's not the end of the world, but people are gonna need to really think about, "How do I accelerate my update cycle?" And I just tell people, for a long time the server and, application administrators have been in this world of rapid patching. Network has been spared from that in some cases, and now we're welcoming them to the party. MICHAEL BIRD Do you think this will have a long-term impact on the way that organizations architect their network and their IT environments? JON GREEN I think it'll help people think about, fault tolerance redundancy in particular of how can I, without having to schedule these maintenance windows and deal with outages, are there ways that I can still maintain operations? Because it's not as though we, flash firmware to every switch on a network all at the same time and push one button and the whole thing goes down. It doesn't have to be that way. I think it'll help people consider end of life issues. We've got customers that are running products that have been end of life for a decade, and they're still in their network, and they're still running. And, the reason why that becomes a problem, is our ability to support those products that have long gone end of support is challenged when situations like this come up, and we've gotta respond to mass incidents. So, it's gonna cause us to think about, in-service software updates more, and how do we make that more ubiquitous across all of the product lines? MICHAEL BIRD Amazing. John, thank you so much for your time. It's been a real pleasure having you on Technology now. Thank you JON GREEN Anytime. Thank you MICHAEL BIRD pretty fascinating slash scary stuff, don't you think, Sam? SAM JARRELL Yeah, I think, the phrase you both used, arms race, felt pretty accurate 'cause the capability is just constantly improving on both sides but I don't know that this is, a race that's going to end the way that traditional arms races maybe have an end point, right? MICHAEL BIRD and it feels really topical because we recorded this interview a few weeks ago, but only in the last week One AI provider's AI model allegedly hacked another company's infrastructure on its own. it wasn't supposed to. It was in a sandbox, and it figured out how to get out of the sandbox and, get into another company's infrastructure. in this, interview that we did, it sort of felt a bit hypothetical, but it sort of is happening really SAM JARRELL what was interesting about the, the instance that you're referencing is that it wasn't necessarily the AI's direct intention to be hacking another entity. It was just trying to follow its instructions and found the most efficient path, which was to break itself out of the, the sandbox. Which is the other scary piece to this, because, when you think about an arms race, you're imagining, good guys versus the bad guy hackers, and there's, true malicious intent rather than, um, just your AI attempting to do what it thinks is the correct course of action. And in doing that, quote-unquote, "correct course of action," it goes too far. MICHAEL BIRD Yeah. Yeah. as Jon said, AI is good at security vulnerabilities. So,it's problem-solving in a way of like, "I need to achieve this thing, so what's the quickest and easiest way to achieve this thing? and John said, yeah, models are good at creating exploits, and models are good at, making the most of those exploits. One of the things that I thought was really interesting was, he used the phrase, chaining vulnerabilities. So the sense that you might have a, a small vulnerability that, uh, I don't know, an organization might think, "Oh, no, no one can exploit that." But actually, the AI figures out how you can exploit multiple sort of low-level vulnerabilities to, create one big vulnerability, so it's making connections in places that as a human you wouldn't necessarily make those connections. And again, we see AI doing that in positive ways, you know, sort of giving creative answers to things. But from a security vulnerabilities perspective, it's being creative, but just not in a way that people necessarily want. SAM JARRELL Yeah. and he did give a little bit of hope in that, like, AI is giving defenders a new kind of early warning system, and if they can find the vulnerabilities before software ships, then maybe that changes the entire security life cycle. But I did notice that, he said that there is a bit of a second half to that story, say you ship it and AI finds a new vulnerability, you still have to fix that, release the patch, and get your customers to actually apply it, right? Uh, people don't like downtime. I even know now that I'll delay updating my PC 'cause I'm like, "Well, I got... I need to work on these several different things, and then, eh, eh, this is gonna have to update by midnight," whatever. And people will wait until the last moment, and in that in-between time, that's where I think that there's a lot of risk. MICHAEL BIRD Yeah. Yeah. A-and, we did an interview a few months back with, the head of infrastructure at Telco, and he talked about how, they would maybe do patches every few months if they were lucky, this is not uncommon, like not patching regularly. But yeah, to John's point, uh, we're gonna have to patch things much faster. I think he said, yeah, y- you'll have to, organizations have to figure out where to find that downtime to patch 'cause the alternatives will be much worse. SAM JARRELL Yeah. I'm wondering if that's the next sort of frontier in all of this is like the having the capability for systems to almost be constantly patching, and how do we get the compute there so that it can be doing its normal functions but then be going through constant updating? MICHAEL BIRD I think there is a big dilemma that organizations are gonna be facing, which is we'll need to patch these things quickly because there's a potential that there'll be vulnerabilities that can be exploited very, very quickly. But also, what if this patch causes something that's not supposed to, bricks the system or whatever? So, this is gonna be quite a critical thing. SAM JARRELL Yeah, I think there will be more instances of this and lots of random downtime and things like that until this gets a little bit more figured out by different, uh, organizations. MICHAEL BIRD Yeah, I think you're right. I think you're right. now, Sam, we discussed the speed of change in the industry, so it feels a little bit mean to have asked John to look into his crystal ball and tell us about the future in even more detail. But it is one of my favorite questions, so I couldn't sort of help myself. So I asked Jon just that. What does he think the future of cybersecurity is going to look like? Will it be all automated, or will there always be a place for a human in the loop? JON GREEN It's hard for me to sit here and even predict it just seeing the pace of AI in the past 12 months. 12 months ago, we tried vulnerability, testing with AI, and it was a big failure. It didn't work at all. ... In December of 2025, January '26, we started seeing signs that it was gonna be really good, So that pace is just ... That's insane. I think it's a good news, bad news situation. I think AI and agentic AI is going to give some tools to the defender that they haven't had before. We're already, deploying in our own cybersecurity operations some of these agents to say, if somebody, reports a phishing attack or somebody reports, "Yeah, I clicked on the phishing link, and then I entered my credentials," that took a human, a good amount of time to go and investigate and pull logs from different sources and say, "What happened there?" AI's gonna give us the power to automate a lot of that stuff. You may still have a human in the loop for a lot of it, having to examine it and draw conclusions, but the AI's gone and sort of analyzed the context of what it saw and gone and done the first, 25 troubleshooting steps to pull together all of that information that it can then compile. SAM JARRELL Okay that brings us to the end of Technology Now for this week. Thank you to our guest, Jon Green 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, 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