Sponsored: Attackers need to be right more than once

0:03James Wilson: Hey everyone, I'm James Wilson from Risky Business, and welcome to this sponsored interview with Dropzone AI's founder and CEO, Edward Wu. You know, there's an old saying in cybersecurity, attackers only need to be right once while defenders need to be right every time. But that idea misses something important about modern intrusions. As Ed explains in this interview, getting initial access is no longer the same as getting the full prize. Attackers now have to map an environment, move laterally, escalate privileges, establish persistence, and find a path to the systems that actually matter to them.

0:39James Wilson: But if all of these additional steps emit a lot of signal for defenders, why is dwell time actually increasing? Why is time to containment increasing? And how do we prepare ourselves for the onslaught of fully autonomous AI-powered attacks? I'll drop you into the conversation here where Ed explains that in order to be able to answer these questions, you first need to understand that a modern cyber attack is a bit like the sort of heist you'd see in an action movie. Enjoy. Like every time I think about this, I'm envisioning, you know, one of those Mission Impossible movies where getting to the building is like the initial 10 % of the excitement.

1:18and then moving laterally through the layers of physical defenses and getting to the ground rule is actually the vast majority of the challenge. And I do think nowadays, most organizations have firewalls, they have different systems. Obviously, there is still a very long way to go for us to get to true micro-segmentation, least privilege. But folks are doing enough stuff where getting RCE on a publicly facing server is no longer sufficient. So typically what we see is that hackers have to first perform reconnaissance. Some of that is reconnaissance on the network layer, IPs and ports. Some of them are reconnaissance on the identity layer.

2:12this is where you see tools like Bloodhound or Sharphound being used. Some of them is reconnaissance. If it's cloud, reconnaissance on the cloud workload layer. And after performing the initial reconnaissance, this is where attackers have to start their lateral movement. And obviously there are different ways and techniques to pivot across different assets within the organization. But with modern enterprise complexity also brings in the challenge of how do you plot a path to your ultimate quantum drill. Very few initial breach get domain admin access or you can say EC2 or AWS root access. So this is where attackers had to take steps to slowly escalate their privilege while moving laterally to get close and eventually get to the crown jewels.

3:15And all of those typically involve multiple steps.

3:19James Wilson: Yeah, so if I follow the Mission Impossible example here, which I think is a good way to think of it, right? Gone are the days where you breached that first door to the building and then, aha, the prize is right there, right? Now, physical security was well ahead of us here. The crown jewels are now deeply buried in a safe that is under multiple layers of concrete through many, many access doors. Those access doors have got man traps in them, security cameras, biometric readers. The same things happened in application security, network security. There's just layer upon layer upon layer. And no single layer defeats the attacker necessarily.

3:52James Wilson: Sometimes they slow them down. But to your point, the attacker's now got to do that initial breach. They've got to do that credential with those credentials. They've got to do the discovery, the lateral movement, establish persistence so they can get in and out and around all these things. But every one of those, just like in a physical intrusion, would be generating some degree of signal, right? Like, again, if we follow the physical parallel here, the glance of someone on a security camera, the quick beep of a panel that says someone's moving through here. surely the same sort of stuff is happening for each of these lateral moves these reconnaissance these establishing establishing the persistence but if that's true why is it the dwell time is in fact going up and detect and contain is going up you know we've we're reversing the five years of improvement here in the industry why why is this in my mind it's two factors one is ultimately I feel like we still don't have enough detections.

4:53There are a lot of degrees of freedoms attackers could take. And one challenging cyber is there are not that many, as many natural choke points compared to physical security. We have things of outdoor, the secure elevators that require retinal scans. these are all physical choke points that make it very easy to deploy detection. But in cyber, most systems are networked, connected, right? And this is where there is not a lot of physical and significant choke points. What that means is attacker can approach the crown jewel through a number of different ways. And in order to detect those, we really have to deploy a lot more detections we see in Mission Impossibles.

5:50So that's kind of one. The second thing is even when we have sufficient detections, frankly, most security teams do not have the sufficient analytical capacity, you know, eyeballs on monitor to really pay attention to every single detection or tripwire that are being tripped. and part of that has to do with the challenge of detection. You might notice in these movies, most of the time, Tom Cruise is trying to break into those environments in the middle of the night, so it's very easy to identify anomalies, because no tripwires should ever be touched at night, when the system is in a lockdown mode.

6:39But, you know, imagine trying to secure something in a busy mall where the tripwires, the infrared sensors, you know, the vibration sensors are continuously getting triggered by all the foot traffic and different activities or legitimate people getting lost and wander their way, you know, to the vault. So I think that's kind of where the lack of true monitoring and analytical capacity is another reason why even in cases where there were sufficient detections, defenders were not able to stop the attackers promptly.

7:19James Wilson: Right. And so that's the, I guess, deficiency on the defender side. Let's flip it around and talk about the attacker side, right? Because it's not all sunshine and rainbows for them either. We hear a lot about this, you know, attackers can move at machine speed. It seems like whenever there is one of these AI-assisted hacking campaigns, we always sit back and go, wow, this is what it looks like at machine speed. But you provided me some interesting stats here, which is that I think this was data from Sophos that found that one verified case of an attacker AI use across 661 investigations, time to exploit was flat, right?

7:51James Wilson: So where's this machine speed? What do you make of this? most attackers still need time to really understand the environment. And yes, you can use AI to help you with that, but that actually still takes time. And when it comes to AI, most of the time, AI is drastically accelerating on the initial intrusion because now when you have a CVE, You can have AI draft up working exploits very quickly and really take advantage of this window of opportunity where defenders haven't patched. I think the impact of AI for initial breach is much higher than the impact of AI at this moment for post breach or post intrusion steps.

8:45It still takes AI time to run M-Apps and to reason about the environment. One big difference between the Mission Impossible movie versus cyber attacks is 99 % of the time for movies like Mission Impossible, they already have the building schematics.

9:04James Wilson: Yeah, right. But I would say most cyber attackers have no idea what this environment is like. One could even argue most defenders don't have these schematics of their cyber, you know, cyber property. So that's where like post intrusion still takes time. And also what we have seen is, yes, there are attackers like using AI to assist with post intrusion steps. but most of them are not yet at this moment letting AI completely run loose and just take it all the way from initial intrusion to the crown jewels. Why is that though? Like what's the limiting factor? Like how long until they are letting them loose and what would need to be true for that to be just the new emerging tradecraft?

9:56Yeah, I think part of this is model capability. Like right now, if you use existing open source models and put them on a particular task, you might notice they will spend a lot of time, you know, hammering the same set of things over and over again. And it's very wasteful and very inefficient. However, when you look at, you don't see experimental models that Z Frontier Labs have been playing with. It's pretty clear that probably within the next one to three quarters, a lot of these newer experimental capabilities or more capable models will be democratized to the open weight world. And that's where higher model intelligence brings exponentially higher cost efficiencies.

10:55using large language models to have it run mmap over and over again without really realizing what it needs to do next is very expensive. But as the models get smarter, they wouldn't waste as much token as the average cost of running end-to-end attacks or running these models in a lot of people call the goal mode. Here's a remote access. Here's my goal. go use Linux shell to make this happen.

11:26James Wilson: Right, and don't stop until it's done. Keep going. Iterate, iterate, iterate, yeah. Yeah, right now, I would say the cost of iteration is still not very economical for most attackers because the open-weight models are not as capable. But I will not be surprised within the next one to three quarters, this changes materially. And with better models, attackers will be able to afford run attack agents in the goal mode and and really um you know get an additional factor of leverage and amplification yeah get that full extended automation and and uh uh you know almost like a hands-off sort of thing of like i found the target we've got the initial access here our agent you do the rest that seems to be around the corner which is a good point to look into okay so if it's coming around the corner in one to two quarters time what what do we do in those one to two quarters yeah in my mind it's a combination of two things one is better basics yes we talk about this asymmetry you know attackers only have to be right once while defenders have to be right every single time but keep in mind historically attackers were also resource constraint.

12:41This is why, you know, there's one school of thought that says, historically, most organizations are secure because of herd protection, right? By hiding in the middle of the herd, you know, attackers don't have all the time in the world to go after you. So this is where I think historically, we are able to get away with a lot of that because attackers were resource constrained as well. But now with AI agents, you don't get to hide within in the middle of the herd anymore, right? Everybody is 360 degree exposed to the adversaries. And this is where I do think when you have a thousand teenagers checking every crevice in your house, you better be sure you do a good job, not leaving any small gaps or holes.

13:35And this is where I do think the cost of kind of operational, you can say hygiene configuration mistakes is much higher now because previously you might be able to get away with some publicly facing unsecure stuff because nobody had the time to really pay attention to everything on the internet. But now with AI force multiplication, a small mistake can immediately lead to a much bigger breach because now attackers have the capacity to literally check every single thing, like check every single port, check every public facing IP, check every single subdomain they could find. So I think this is where security hygiene, whether it's defensing dApps, whether it's segmentation, least privilege, having good visibility continue to be very important.

14:35Right.

14:36James Wilson: The thing that I was hoping you and I could sort of debate in the open here is what's the role of AI, right? I think we both agree that you've got to get back to the basics, the allow listing, the access controls denied by default. So let's take that as established. But my question back to you, Ed, is what's the role of AI versus AI here? And I come from this problem from the perspective of observations that we know that guardrails to instill safety in models don't work particularly well. even input classifiers putting a large or a small language model in front of the big large language model it doesn't work it can be worked around you know we've seen some really novel prompt injection attacks that that's just a problem that's never going away so with those observations in hand when someone says to me oh the you know the the solution here is just have ai fight the ai right i always go i i just don't think that's going to work because you know if we can't solve prompt injection if we can't solve classifiers if we can't solve guardrails because of this indeterminate nature of an ai model isn't just pitting ai against ai in this attacker and defender sort of equation here just a like it just boils down to a zero-sum game eventually or is is there i guess convince me there is a lasting role here for ai on the defender side even in the case of pitting your defense AI against your attacker's AI?

16:00I think of it less about using AI to fight AI. I think of it more as software-based staff augmentation and how do we enable security teams to operate as if they have 10 times more headcount and 10 times more budget. I think nobody can disagree that the overall sophistication capability and risk reduction increases as the security budget and headcount increases. In my mind, AI is like the best worker bee we cyber defenders can find on the planet. And in order to prepare ourselves for the onslaught of more attacks, probably more intrusions, as well as let's also not forget that the cyber, you can say, property square footage or the complexity of modern enterprises continue to increase drastically.

17:04I think if you look at a typical organization, whether it's the number of IPs they have in their internal network or looking at the numbers of identities they have in their active directory or the number of even SaaS apps that they're using, that has been drastically increasing over the last decade. And now with AI agents, like we as defenders have a whole lot more we need to defend. And I think what most people could agree, you know, the overall complexity of the digital space or the digital footprint of most organizations is growing much, much faster than the overall budget headcount of the security team.

17:50James Wilson: We've sort of strayed a little bit and sort of veered into talking about things like budgets and headcount and multiplying the effectiveness of the staff you've already got, which is starting to talk to the things that would be on the mind for a CISO. So why don't we just sort of come up for air now as we wrap this up and say, if this is resonating with someone that's listening and they're thinking, you know, this is the kind of defenses that I need to be thinking about how we invest more in, the first thing they've got to work out is how would you even measure, you know, success of kitting out an analyst with more AI tools, whether it's from Dropzone or from anyone else?

18:27James Wilson: Let's just start from the basics of if you were thinking about augmenting your defender and your security posture with AI, how do you even begin to frame this as what would we measure, what does success look like, and how do we even frame a business case around this? When we talk to prospects, CISOs, security teams, one question I always ask them personally is, what would you do if you get 10, 20, 30 additional security engineers? What will you have them do? Tell me about what that looks like, and then let's work backwards. like what would it take for us to get you to that um you know without 10xing or doubling your budget or headcount and a lot of that a lot of times that boils down to two things like efficiency and effectiveness right ultimately more capacity means that you are able to work on more projects there might be security practices projects whether it's you know increased hunting or kind of operationalizing threat intelligence more or actually spending more time looking at security alerts the low and informational alerts that historically were not paid attention to at all like these are all net new tasks and projects that are that could be done by the AI agents And then the second vector is effectiveness.

20:01And a lot of that comes down to a decreasing time or decreasing latency, because every minute matters when it comes to whether it's operationalizing threat intelligence, whether it comes to completing threatens, drafting detections in response to a latest attacker technique shift, as well as obviously looking at security alerts. So reducing the latency of these activities from hours or days down to minutes is another big effectiveness gain we see in real world agent-exercive deployments.

20:44James Wilson: yeah that makes a lot of sense ed well let's let's wrap it up there it's been great to walk through i guess dispelling the myths around the attacker versus defender asymmetry and of course the increasing role that ai will play on both sides it was great to meet you thanks so much for dropping by for this chat thank you for having me Look out. Look out.