Friend or Foe? AI's Double-Edged Impact on Cybersecurity

Sharif Abuadbba

What I'm going to try to do in the next couple of slides, is basically to try to (sometimes you have to have a catchy title “friend or foe”) zoom into the cyber space and see the impact of AI from both angles, so to speak. I like numbers, so you're going to have ten. You can follow me ten. The slides I cheated somewhere, so one slide is split into two. So you have 11 there. Who likes reading books here? All right. Great. The harder question. Who read the book Strange Case of Doctor Jekyll and Mr. Hyde.

I'm sure you are not as old as the book. Definitely. It's 1886. So, the book is basically written by a Scottish writer, Robert Louis in 1886. And it goes, it's a novel, right? And the discussion these days about AI and cyber, I don't know if that reminds me of that book. The it's a well-known story for dualities, which is the same entity, human invention, holds both thoughts of good and evil. So in that book, that writer wrote that the Doctor Jekyll is the main character who’s fascinated about the human nature and created a potion medicine, so to speak, to transform part of him to trigger the darker side of him that brought him to Mr. Hyde.

Right? And he thought, as a primary person, Doctor Jekyll, that he would always have that control onto the other character. And it didn't end well. I'll have a different answer later. But that duality is very important in today's discussion. Bringing it back to cyber security, as you know, the friend part is basically we think that we could do red teaming, identify vulnerabilities, all of that. Whereas the foe side of it is very clear: the cyber criminals and all of that will leverage it to speed up

and all of that. So we will try to talk about both angles in the discussion today. All right. So I realised in such talk we have a diverse audience between people who might be very good in cyber space all the way to people who might not be familiar with the cyber space. So for the expert, forgive me, I'm doing 101 of cyber here to carry on the discussion. Basically in cyber space, to bring you to the ground, we usually follow at high level two frameworks. An the attacker side.

It's a MITRE ATT&CK framework. How adversaries move across the attack chain of computer systems. There are 15 categories. I just picked six to make it simple, between reconnaissance and scanning systems to initial access to privilege escalation within the system to credential access. Again, sometimes the attackers need to do something we call “lateral movement” between computers to do stuff or systems all the way to command and control. That's what the attacker want. Once they get that, it's time to do damage and do exfiltration of the system.

Each one of those steps, historically, needs competence, needs skills, needs time and all of that. I'll have some examples where we are at now with the AI side. On the defender side. NIST Cybersecurity Framework 2.0 is the one. Usually it's the 101. And we go a couple of steps there. Govern, which is basically to set the appetite risk of the organisation to identify what are your digital assets, dependencies, all of that. We’ve got to protect now. How to make sure the system is hard? Configurations and all of that.

You need to have detection in place. People use AI and other tools to detect and you need a response in case you’re compromised and you need to recover. Restore the system. So this is just, basics on the, both the defender on the attacker. I will go further with an example. So at least when I talk later about vulnerability, exploitation, exfiltration, you have an example. Hopefully you don't go and try that example. And in most cases, systems will not have that kind of silly example. But the example is very simple.

There is a website that offers PDFs as a download. Very straightforward. What we do usually behind the scene for you: you just request the website link and behind the scene as a developer we would have a command like GET https://portal.example/download?file={filename}. Straightforward. The attacker knows something through scanning is basically that input is not checked (that simple “report.pdf””. So I can’t change it as an attacker. Right. So that's the vulnerability. What can the attacker do? They will just change the “report.pdf” into something else. Guess what? Secret.txt? Could be the configuration of the system?

Whatever. So that's where the attackers' variations will happen. And if the attackers are lucky that you don't have the checks in a place they already got your, whether it's secrets.txt or .md or configuration of the system, now they gain some information from the system, they will look at it and they will use it for next steps of attack. I'm just giving four quick examples here to move on. All right. So the theme number three I have is basically that AI compresses attack economics. That's the key where the AI leveraged by the foe in the cyber space.

A couple of things here are the key ones. But there are many others. The skill barriers which I mentioned earlier before in cybersecurity. The guys who are in the cyber space: we map the attackers, state actors, criminals, all the way to we call them script kiddies (who are just, you know, university students – we have bright ones – but they're just trying to learn). And we can learn that through the attack on the system, we can tell actually this is actually a script kiddy, or this is a state actor.

The problem with the AI is that it's blurring those boundaries. That's the skill barrier. A young kid, 16 years old, with the capabilities of the AI model – they could ask all these questions and get the right answers if they have the motives to go for it. Second is basically the time to action. It's very clear the switch between the stages I was talking about. Before the human would be between them doing reconnaissance, exploitation, and exfiltration. Now the AI can do that faster, especially if they understand the context.

Cost of retries? The same. Now it's a cost of tokens. Million tokens will get cheaper and cheaper. They can do all these variations and learn through that. And finally, what you all observe is basically the scale: personalisation and contextualisation. So as an attacker I can go after ten organisations at the same time and can contextualise the attack across all these ten organisations. There are no issues from the attacker's perspective to do that. Now, I'll go back to give you a little bit more, once we are doing the analysis now, where the LLMs or the agentic AI are good at across MITRE attack framework.

So where the attackers now? I'm sure they are using it in various ways. Very quickly. In reconnaissance, summarising targets is basically “I'm trying to gather as much as information about my target”, right. So the LLMs / agentic AI is doing that reasonably well now. Right. Exposing assets, collecting information from open intelligence, you go to MISP and others. There is a lot of information out there and you try to synthesise that very quickly. You go to the initial access, which happened, attackers need to generate phishing emails,

convincing pretext and all of that. The LLMs are good at that now. The privilege escalation. Basically, privilege escalation is when they get access to the system they want to have a little bit of admin access. If there is a misconfiguration, the LLMs now understand that a little bit. Sometimes the attackers will give it access to some tools (like Metasploit and some of those tools), and the attackers could potentially fill in that gap. Credential access as well. The example I gave in the last slide is where you’ve got the secrets.

So from the secrets the LLMs can analyse what I have got, what is the value I have in the stuff that I extracted. And it could identify quickly (if you are not doing a good job as a defender) how to extract that lateral movement with the access to tools, you could move between devices and command and control. It's the same command and control. Once you are there, you need to generate scripts (the LLMs and agentic AI are good at that). All the way to exfiltration. In exfiltration, you need to collect documents (it could be large scale documents, with the AI nowadays you could analyse all of that).

In my talk so far it sounds like I'm doing a favour for the attacker. I'm giving you all the “tick, tick, tick”, and that's intentional by the way. Why? Because what I learned in cyber space is you don't assume that the attacker is not capable and does not have those capabilities. Let's go to the end. Right. They have all these capabilities. Then what? How could we react as a defender? You are expecting me to mention Mythos? I can mention Mythos here, and I will move to the next slide because everybody talking about it so very quickly.

Our colleagues at the UK AISI. So Mythos is not available now to the community – you probably heard it's only available for 40 organisations – but our colleagues at the UK AISI they have a “cyber range”. A cyber range is basically a network simulation, a couple of computers connected to each other. And they tried many of the models that are coming up. The latest is Mythos. I'm showing you some high level numbers just to have the feeling. Within that cyber range that is around 32 steps.

So Mythos was the first model that gets all the way to 32 steps. Before that, we didn't get anywhere close to that number. We usually do multiple tries (in case it's by luck), Mythos got there three out of ten, which is the highest among other models as well. If you want to dig deeper a little bit into those steps, the average Mythos got is around 22 steps out of 32. To give you a sense of where other models are at: the closest one was 16. So I'm trying to give you an indication that we have a slope.

So they are getting better and better and the success rate is 73%. But I also want to not have fear mongering here. I want to ground the results a little bit. One one side there is a continuous improvement. It’s clear the models are getting better at plan-act in complex systems, but at the same time in that cyber range there is no defender in place, so no detection. If you want to categorise organisations, it's in the lower quartile of organisations usually where they don't have the capital to have defence and detection.

We do have those organisations, so those organisations are exposed. All right. My theme number six is basically – and it could be a little bit of marketing play – a lot of conversations are around the vulnerability detection. Basically the discovery vulnerabilities. “Oh, the system has vulnerabilities that humans never detect and we managed to detect”. Now I'll talk quickly about the detection side and then what is next. So on the detection side the play is between the the friend and foe at the same time. On the friend side, it’s very clear.

AI can find the vulnerabilities very quickly. We could use it as a defender advantage. I could prioritise and triage very quickly. That's the assumption. It could help me support patching of those systems. Prerelease security checks. So moving forward, while I release some of those systems, when they have those vulnerabilities all the way to hardening some of the legacy code base. But then we can't ignore the foe side of us. The capability may diffuse beyond trusted defenders organisations, so it will be accessible by others. And arguably now even some latest open, open source models like GLM 5.2. If you heard about it, they are a few months behind the Frontier Labs, so they are catching up.

Outpace verification, getting cheaper access to vulnerabilities and so on and so forth. But the most important point I want you to take from here is basically: vulnerability detection is one major step, but what that means is how are we going to go after the detection? I need patching and fast paced execution. So if I'm detecting vulnerabilities on the scale of thousands per day, do we have the capabilities now to create those patches (timely), execute those patches without ruining our systems? We want our systems to function and one thing we learned through the last many years is when you touch that system, the system will be unstable through all these patches.

So those are the challenges that we are not talking about now, because the focus is more on vulnerability detection. Okay, let's go a little bit to the defender side now. I think on the from the defender side the structural advantage is… I'll go three by three here. So on the attacker side we talked about cheap retries and fast iterations. We talked about asymmetric access. Because one win is enough for the attacker. It's like a goalkeeper and an attacker. As the goalkeeper, I need to catch everything.

If I lost two, I lost the game. Whereas as the attacker, I could shoot many, but then I get one, I'm the winner and have the day. But then as a defender we shouldn't lose our balance. We still have full asset knowledge, the context of the context of the system is still on our side as a defender.,Unless you have an inside that attacker, which is different story, but in most of the cases that's our strength. We have logs, telemetry, identity, data access and all of those, we have the authority to isolate, patch, and so on.

And in most of the cases, when I go to events like this and I see people face-to-face, it reminds me this is also our strength. As a defender, we can come into big rooms across countries. I have our colleagues from the AISI. We have the network with other AISIs. We can do a lot of things in the open where the attackers can't. So that is all strength that we shouldn't forget. But I think, more importantly: assume that we start with a lot of legacy systems now.

And the discussion is: of course, we need to repair those legacy systems but the ideal situation I think it's a lost game for us to keep chasing the attackers and try to follow up with the vulnerabilities. If a system now running has 500 vulnerabilities, we need to deal with that. But the ideal situation is we should move as a defender to the secure by design mentality rather than just chasing vulnerabilities. And there are many ideas there. I'm just mentioning with the high level design time security, so we could actually design the system where it has at least potential vulnerabilities, least dependencies on other things, and so on, so forth.

The other thing is policy and configuration is like a proactive way rather than reactive way. And most importantly, of course, as a legacy system, we do have a lot of systems now as a legacy, we need to have code hardening. We deal with some of those problems. But I think short term, in my view, it's more about how to fast detect and solve those parts in terms of patching and so on. But then the medium to longer term is about “secure by design”. All right. Theme nine: the horizon is basically between continuous assurance, resilience and rapid recovery.

So I'll just say a few words about those three. I think the continuous assurance I touched on that in the previous slide – which is basically rather than wait for the vulnerability to happen (certificate expires) we are talking to big organisations where they are all on their back foot just trying to catch up. It has to move into more proactive assurance before those certificates expire. I need to know a month ahead or something like that. All that stuff AI could help with. But more importantly, I think it's like, we are in the time of the World

Cup now, and when you watch teams, you can tell when the team is resilient: when they get the goal and then they could come back. That's the resilient spot. No matter what we are going to do on the defence side: resilience and recovery are the key right. And in most of the cases when you say resilience, we have systems of systems interconnected. And every now and then we see that we are not planning well. When part of the system goes down, the whole system goes down

and people trying to figure out what happens. Plus it's very hard to recover. And if you go back to the NIST framework, there is a lot of work around govern, identify, detect, yada yada… there is less work around recover. Right. And I think now in the era of very advanced AI, we need to a very fast paced resilience and recover because we will see more of those issues, not less. It's unfortunate, but that's the reality that we need to deal with. All right. My theme ten.

A few key takeaways and trying to connect back to my 1886 book: friend and/or foe. So AI will accelerate attackers. I showed you examples of the mapping and from the reconnaissance all the way to exfiltration, but we are hoping that AI will also strengthen the defender side through now detection, patching and all of that, hopefully all the way to the “secure by design”. I did mention earlier that detection alone is not enough. We need to have solutions for fast paced patching while the system is operating.

Resilience becomes the deciding factor. And ending with the future is proactive, secure by design. Back to Doctor Jekyll and Mr Hyde. In my view – I'm commenting on a book that's 120 years, 150 years, whatever the math – I think what he missed is basically he thought the good side of him will be the primary and the Mr. Hyde, the foe always secondary. And they got just a flick, a switch to switch that off when they want. I think that's the wrong way to go about it.

I think what we should do (probably, in my view) is: it's both there and we should assume they are independent entities, that they are moving together, moving forward. So whatever we are doing on the defender side, on the attacker side, we need to make sure that's “uncontrollable” means they will have the capabilities, they will attack those systems. So what we need to do with that, exists there. Finally, that's my last slide. We wrote a paper with our colleagues from the UK Alan Turing Institute and others.

If you are interested to read more about the topic, that's the paper. That's my talk, thank you.