Intro/Outro 0:01

This is AI Native Cybersecurity. Field notes from security teams defending against and operating like AI adversaries. Because the best defense is a good offense.

Shahar Peled 0:16

Hi everyone, my name is Shahar Peled, co-founder and CEO of Terra Security, and your host of the AI Native Cybersecurity Podcast. Welcome to the first episode. This is where we'll host some of the most innovative cybersecurity leaders with AI. Stay tuned for incredible thought leadership about how the best innovative cybersecurity leaders innovate with AI. And for the first episode, I have the incredible honor of hosting CJ Moses, the CISO of Amazon. Looking forward. Welcome everyone. CJ Moses. Thank you for joining me today on the AI Native Cybersecurity Podcast. It's a true pleasure to have you. I can't wait for this episode. Thank you for being here. Exactly. I'm excited for that. Before we jump right in, I guess most listeners know you already, but if you can tell us a few words about yourself, that'd be great.

CJ Moses 1:09

So my title is uh uh CISO of Amazon, primarily of our integrated security. And what that actually means is that uh uh we look at security a little bit differently, and that is that we don't look at security just being cyber or personnel or human or you know those types of different intelligences. We pull it all together into one, and that's been the uh the job of uh you know this is over the overarching Amazon, the 148 plus business units that we that are aligned, um, is basically being responsible for the security integration of all of that information to make sure that we keep bad things from happening to good people.

Shahar Peled 1:48

That's incredible. And you know, we met, I think almost two years ago as as Terra won the AWS CrowdStrike um partnership or accelerator in partnership with Nvidia for the most innovative um cybersecurity AI startup in the world in 2025. And ever since, you know, watching from the side on your journey and the amount of responsibilities you have, it's just it's been immense. And uh having you is a pleasure. So let's uh jump right in. We have so many things to cover today, and I'm sure the audience would love to hear your perspective on how the world is changing. Everyone knows it changed already, but it just keeps happening on a daily basis. Um, so what's the biggest change you have seen in cybersecurity since Gen AI was, you know, went mainstream? The pace, the new risks, the new discovered vulnerabilities, what's the one thing you can put your finger on?

CJ Moses 2:40

Yeah, so obviously the number of vulnerabilities could be one of those things for sure, but I think um the speed which you need to respond to them. If you think of the whole uh paradigm of you know 30, 60, 90-day patching windows, um, that seems so old, but it really was only like six months ago that's the mindset, and maybe for many it may still might still be that way. But the reality is is the time of exposure between when a vulnerability is you know known or you know of it to the time it is patched or otherwise mitigated. Um, you know, if it was 30, 60, or 90 days right now, with the threat actors being able to use uh the new technology mythos and other types of things to be able to not only find but also then turn around and do something about it, um, is something along the lines of, you know, um seconds. Um we actually have seen in some of our uh uh some of our telemetry, we have this big uh honeypot infrastructure we call mad pot. And in one case, we actually, you know, saw not only the intrusion happen rather quickly, but the entire takeover lateral movement across everything that those systems that they had initially intruded upon and all the access that it had to everything else. Um it's complete beginning to end, everything that they could do. It was uh 12 minutes and 42 seconds. So you're going from nothing with a something with a vulnerability, but no, you know, no access to um, you know, hacking the system, you know, doing all of the reconnaissance work on the system to figure out what accesses are there, then acting on each and every one of those things across the totality of not only the system, but the totality of the access that it has, um, and then continuing to spider web that out because it's not just one system, it's then when you get into the next and the others. Yeah. And in the past, this was something, this would be a campaign by uh, you know, an actor, uh, uh an adversary. Weeks, sometimes months. Weeks, months would go on forever. This was 12 minutes to, you know, just shy of 13 minutes. And uh those types of things are really kind of the big change. I think um specific to Amazon, one of the things that we were just kind of in my you know intro there, I was kind of hitting on. Probably the biggest thing that we've seen to change on top of working to mitigate those threats is integrating the security across the totality of Amazon. Um, having learned lessons from uh the 9-11 commission that you had intelligence that could have stopped a bad thing from happening, you know, 9-11 attacks, um, those types of things, we do not want to be Amazon and have pieces of information across the company that could stop bad things from happening to good people that we didn't put the pieces together in order to do. So we have now integrated our capabilities to have everything from our cyber environments, so our normal threat intelligence space, all the way through to workplace incident management investigations to all of the HR actions and stuff like that. Um, not only are we witting, in many cases, I own the responsibility to uh to conduct and to follow through, and therefore it gives me the ability to change things. So it's not just like I'm aware of them or my my part of the uh company is not aware of them, we're aware and can react immediately without asking because that's my responsibility to take care of it. So it's a little bit of a different uh mindset than what you've seen elsewhere.

Shahar Peled 6:01

Yeah, I always say that cybersecurity, every defensive control we have, and most of them, their goal is to either block, stop, or slow down an attacker. I feel like this last part doesn't make sense anymore. So now we either block them or it's too late. And so if I if I have to to minimize your your your answer to one thing is is the time. Like we have to be faster. I at least haven't seen these mega zero days that we've never seen before. I see better pace and velocity and volume.

CJ Moses 6:33

Yeah, so time it's velocity, um, and then if you also take it down, it's um the speed by which the actors can act um and do is do a lot more than they were able to do before in that period of time. So yeah.

Shahar Peled 6:49

Okay. And you talked about Amazon. Amazon runs one of the most robust cybersecurity programs in the world, protecting billions of users and transactions every day. Is there one specific thing that not something that surprised you, but something that actually was a major shift in the threat landscape that you could put your finger on? And maybe if you can share that, how you dealt with it?

CJ Moses 7:11

Sure. Um, so I think uh I think one of the things that uh probably surprised, and that's the thing is that uh not surprised very easily these days because there's so much that we've seen over the years. But um, you know, with the recent uh DPRK um, you know, uh types of things, we actually saw them scaling their uh insider threat operations using AI, which that's uh normally wouldn't think of the DPRK as being frontline AI innovators or users, um, but they're actually using AI to generate personas, backstopping entire identities and uh using that you know kind of uh resources um to backstop their intelligence operations um in using AI to let them do it at industrial scale. And to some extent that's problematic for companies like ours because a lot of what they're trying to do is create all of that stuff in order to get jobs at places like Amazon. And we've actively um stopped at least 2,900 uh DPRK actors from you know in the hiring process coming to Amazon. So being able to keep those people out of your company not only uh obviously you know uh keeps them from being able to be further threats into your space, um, but um also from the standpoint of a lot of what the DPRK is doing is actually using the funds that they get from being employed, and I air quote that because obviously they're doing it uh falsely across many, many different places, um but uh which are going back to fund uh you know the DPRK's you know wartime mission, if you will. That's not something that uh we want to be funding ourselves. Um so you know, looking at this, we've you know seen a lot of this, you know, as far back as April uh 2024, and they're playing a long game. They're they've been trying to do a lot of the stuff. And uh now with AI actually allowing them the backstop, it's just a completely different world that you live in. Yeah, um deep fakes, uh deep fakes and all these types of things. It makes it so much harder for us to have any type of remote workers, you know. Uh the return to office types of discussions and stuff like that, you know.

Shahar Peled 9:27

The return to office also has something to do with that.

CJ Moses 9:30

Well, that's I mean, it's not the main reason because uh quite honestly, I think the main reason is culture is to maintain the culture and have people work together because quite honestly, people do work better together when you're face to face. That's why we're talking face-to-face today. We could have done this virtual, it wouldn't have been the same. Um, same thing applies with the work that we do at Amazon, but at the same time, from a security perspective, being the security guy, I look at it from the standpoint of it's a lot easier to know that you're showing up for work that it's actually you than you know a remote worker that you really don't know, don't see, may see video of, and again, could be completely deepfake the whole time. Um, or you maybe, since you didn't know the person to begin with, could have somebody different interview and then somebody else show up on the you know on the virtual scene.

Shahar Peled 10:13

One of our biggest customers was on the news for uh accidentally hiring a North Korean engineer. Um and that was not pretty. They changed a lot of things in how they hire since then.

CJ Moses 10:23

Yeah, and that's I mean, you just heard all the things we saw that they were doing and just using AI to do that. So obviously it's it's not as easy to uh to not do so, or not to not hire someone like that as it would sound, or you would think.

Shahar Peled 10:38

Yeah. That's that's super interesting as a use case. Moving from Amazon to Amazon Web Services. Yeah, yeah. AWS. Um AWS invented the short responsibility model, which I'm a big fan of. Um we have something pretty similar with our customers about guardrails for AI agents. Yeah, um, it held remarkably well. Has anything changed about the shared responsibility model now, people working with AI agents, the workloads, the differences of how you actually secure the new threat landscape? Or is it the same?

CJ Moses 11:08

So um this is where I get to go on a little bit of a side story because the shared responsibility model or the shared uh security responsibility model um actually is something that um I had responsibility back in um 2008. One of the things that Andy asked me to do, uh, that's Andy Jassy, um, he'd gotten a questionnaire in from a customer, and it was a lot of security-related questions. And the Jesse is the CEO of Amazon for those who do it. CEO of Amazon, who's who was at the time wasn't even the CEO of AWS because there wasn't a CEO, he was the leader of AWS, so it was the senior vice president of AWS, you know, it's a long time ago. Um, and he actually uh asked me to literally answer this questionnaire of security questions. And it was easy, you know, pretty easy stuff, safe harbor, which is probably not even a term most people even understand anymore, but these types of things. Um and it took me a long time to get all the answers put together, and we created out of that, rather than just answering the questions, created a security white paper. And the first version of it came out. Um, and the thing that we got questions from customers all over the place, anytime that the doc was read was okay, but what are you responsible for and what are we responsible for? Where is the the you know, where is the separation? What do I what do I need to do versus because the uh you know in the cloud there was always the magic of the cloud, put it in the cloud and safe and secure. And you know, most people understood that there was some responsibility they had. So I took that and figured out that we needed to have a paragraph or two in the security white paper that was um a you know shared security responsibility model and kind of explain that and then go on and have a matrix that was a separate doc. And it took me a couple months to actually get it staffed through, which isn't normal at Amazon. Things don't get staffed, it's kind of like you've reviewed, but it was reviewed so many times because there was people weren't comfortable, especially on the sales side of the house, saying that the customers had to do anything. And I was like, no, they need to know. And we worked through it. And so the first version of that took six months to get that new version of security white paper out. It included this the security uh, you know, responsibility or the shared responsibility security model in there. Um and it was written generically enough with the intent to be able to adjust because we have so many services in AWS. At the time, I think there was a handful. Like, you know, when I started in late 2007, there was four, then there was five soon thereafter, and there might have been 10 by the time the first doc came out. There are a couple of S. So we made it, you know, we made it past EC2 and S3 and a few other things. Um, you know, but uh the model then was so intentional, understanding that we need to grow, that today, even with AI, um, you know, that model still uh holds because it's really the matrix of what is responsible, who's responsible for what uh within the AI you know environment is kind of the add-on of so the infrastructure, if you're running um in AI in bedrock, of course the data center infrastructure, the security of that's my responsibility, not yours. And you can work your way up the stack to you get to a point where where you have access, you have responsibility. That's always been my tagline for whenever anybody asks, if you have access to something. So you if you do online banking and the bank does all the security over the entirety of your your account and all of that stuff online, and you take your user user ID and password and put it on a sticky note, stick it on a computer, and someone takes it, uses it. Whose responsibility was it to secure that user ID and password?

Shahar Peled 14:49

Yep.

CJ Moses 14:50

So it can do all that stuff. So in in any environment, there is this model, and it plays, same thing applies in the AI space. So I say it still carries, um, and the one thing is is where's the boundary? And that changes depending on what level of what services, if you're using, you know, in our case, if you're using uh bedrock, um, you get more things that we're responsible for because we're putting the some of the guardrails in. Um, but then it there is that still handoff point. So I I still think it holds true, and I'm I don't know that I'm proud. I'm probably just as much uh bewildered that it actually is still holding. Um that something we wrote that long, you know, that long ago and debated a good bit, um, but it's still been the answer to many times the questions that customers have is you know, there's early days we started creating a matrix, a literal matrix of what it was, and as the services grew, we were like, each service can can call out that that boundary.

Shahar Peled 15:48

I had no idea. You're literally one of the founding fathers of the shared responsibility models.

CJ Moses 15:53

So at least as uh viewed by Amazon, because if you think about it, the this shared responsibility model exists in many things, but from the standpoint of from a security perspective, how we implement it. Yeah, yeah. That's so I I didn't know that. It really, I would say the founders are our customers that were asking the question.

Shahar Peled 16:11

That's a good politically correct answer.

CJ Moses 16:13

No, it is, it's but it is literally, it wasn't something I dreamt dreamt of. It was literally they asked questions to the point where I had to remember or I had to come up with a way to be able to answer it, and the answer was this model. It still helps today. And so I'm glad that it does.

Shahar Peled 16:28

We have something pretty similar. We don't call it share responsibility anymore, maybe we should because we didn't trademark it or anything like that. We put it in the hands of our customers and some of them who wanted what we call TerraTorch, which is basically an AI hacker in your palm, right? You tell it what to do, you give it the scope. You have to be responsible for the actions you give it for the guards are removed. So something pretty similar. Um let's move on to the next one. I think it's very known that within Amazon and AWS, you're big users of AI, whether it's within engineering, security, and and I guess beyond that, sales, marketing, everyone. Efficiency is important. As more and more code and tasks are being done and generated by AI, the new attack surface, the way we see it, is is really intent, not just code.

Speaker 17:11

Yeah.

Shahar Peled 17:11

Does this change anything in your point of view and perspective of how we need to secure the modern threat landscape?

CJ Moses 17:18

Yeah, for sure. I think you know, AI generated code um still has vulnerabilities just the same way human code does. Um as the models were trained by human code. So you're going to also be trained on the mistakes, you know, and you would think that AI would learn from that. And in some cases it does. Um the training data that's used, the AI will recreate that which it's um you know been trained on. Um, and I think that the way we deal with that is we run faster CI CD pipelines. Um, and because actually slowing them down in order to try to mitigate some of the stuff is actually the antithesis of what you want to do. Yeah. Um, because it'll make you more vulnerable for longer. So in this case, we'll actually speed speed up or run those um faster and in that process, we'll actually do as we're doing um code development, especially patching, um, we will actually iteratively be scanning the code in its in its iteration. So as it's being created, it's being scanned with the same types of things that we would scan the code for after the fact, but we're doing it in line as much as we can. It's part of CI C D's part of the CI CD pipeline. And in doing so, by the time you get to the end, you're pretty close to where you're you're feeling good about it. And then that's where at the end, when you have the encode set, um, once you have that final scan done, you can then um do do the types of testing, you know, pen testing or otherwise, that you actually want to do on the end code. But do this all just like we we talk about, the uh the adversaries are actually able to do things so much more more quickly. We're doing the same thing, um, embedded into the environment so that we can actually iterate uh and keep the speed. Because if you slow down that process, and that's where a lot of people are like, hey, we need to slow down because things have changed so much. The reality is if you slow down, you're actually playing into the hands of the thing we talked about in the beginning. You you know, the time to deployment needs to or the you know, the convergence needs to happen where um you were, you know, I I three years ago when I was doing the AWS, uh it was the AWS CISO, um, one of my main keynote items was actually the mean time to defense. Because it was, you know, you always had the mean time you you found a vulnerability, and then it was always, you know, um, you know, just that you found it. How long did it take to find a vulnerability? Uh it doesn't matter. What's the time between the vulnerability and getting it mitigated really mattered? Really kind of plays the same today. And in this case, the CI CD model that we're using with a enabling AI to enable us has allowed us to continue, um that path.

Shahar Peled 20:09

And you actually probably have data to prove it. Do you see any do you see a decrease in the amount of critical vulnerabilities in production that you have to remediate if you do CI CD well, but then also pen test at the end, as you as you mentioned?

CJ Moses 20:23

So yes and no. Uh, I think what we're seeing is so by way of by way of numbers, I think we are seeing right now um more because there's more new software. So I don't know that it's apples to apples. And maybe the ratio. The ratio, yeah, the ratio likely, and that's I don't know that we've done the math on the scale of things. But in the last few months, with the number of even just basic CVEs that we're patching, that's part of that iterative process, um, you know, we're covering the coverage and the amount of things that we are uh patching for, um, you know, using you know AI assisted is many, many fold more than we've ever had in the past. Um and knocking on all kinds of wood, including my head, um, it's it's been rather successful. Um and that's the thing is you want to keep the attack surface down. Yeah. Um, and as much as we can, we're doing that. We're also informed by the Mad Pot um you know, environment where we can see what uh adversaries are going after. So that helps us to prioritize what we're going to patch, how quickly in the environment as well.

Shahar Peled 21:33

Yeah, PPT, dark web, threat intel. Yeah. As a daily user of Amazon, I hope you keep being successful in that. Um you know, for decades, security has been mostly defensive and reactive. More and more security controls to slow down or stop an attacker. Stacking tools, you know, encompassing controls, it worked mostly. I personally, and you said pen testing, I believe offensive security or the umbrella of pen testing. Red team and basically proving the exploit, mimicking the adversaries is where the world is about to go. But obviously, you know, that's from my perspective. In today's world, how do you see the ratio between defensive and offensive going forward?

CJ Moses 22:12

So I think the defensive math of stacking controls on top of each other has been overcome because part of the idea there in that defensive math was that an attacker would only have enough time to go through parts of that and they would get befuddled by one part of the defense in depth that you're you're deploying and then move on to somebody else or never be able not be able to do it. That's where AI is chaining things together and able to go through that. So that math doesn't work anymore. I'm not saying defense in depth, there isn't still some benefit to it, but in the case of chain daisy chaining from an AI perspective, chaining vulnerabilities together or being able to go through the different defense in depth things. So we have flipped, you know, we're still doing some of that stuff, but for the most part, uh testing your own environment, uh, very much akin to what you know what Terra provides, is the path that we're on. We're testing, you know, doing the code remediation as part of the CI CD pipeline, but then also on the backside, continually testing the environment and then acting on uh what you find before someone else has the chance to do so is the model. Uh right now, and uh obviously things can change very quickly in the space, but that's that's the way that we're basically using AI to prevent AI from you know uh harming our our environment or our you know our people and data.

Shahar Peled 23:40

I think it's it's crucial because eventually if you keep playing the numbers game with the amount of vulnerabilities, you can never win. Eventually you have to find this very small number of vulnerabilities that are exploitable with business impact, yeah, and that's what adversaries can actually achieve today, because there's just too many vulnerabilities out there, most of them theoretical.

CJ Moses 24:01

Yeah. Well, that's the the theoretical one. That's the that's part of it, is that uh understanding and prioritizing what is so not only what is real, um that's the thing is though, we've been able to, you know, be able to use AI to be able to um identify vulnerabilities, but then once things are identified, also have it create a proof of concept uh code right there before they even tell you about the vulnerabilities system so fast. It's here's a proof of concept that it that we that it's already tested. Well, wait a second, how'd you test it?

Shahar Peled 24:32

No, no, no. And everyone is so excited about autonomous systems now, right? Yeah, and and as they should be. But I think there is some some lost premise right now in autonomous in cybersecurity or offensive security, whatever angle you want to take it to, just like uh what self-driving cars. Yeah, how autonomous is it? What happens when, God forbid, there's an accident or it's been stopped by a police officer? Is it really autonomous? Heart surgery and other things that we're not ready to have autonomous. What's your take on autonomous systems in offensive security or cybersecurity?

CJ Moses 25:02

So, yeah, as a whole, you know, uh Amazon has this model that's worked really good for us uh by way of making decisions. And uh decisions that are two-way doors, you can make much more quickly. What do I mean by that? Anything that you can back out of rather easily or roll back quickly without any impact, why not move forward? So in the CI CD pipeline, if the blast radius of the change A can be rolled back and doesn't impact a large number, doesn't take down AWS, doesn't take down the internet, doesn't um, and that's where the judgment comes in as to what do you allow, in this case, AI to actually autonomously essentially make the decision that you can move forward because you have a means by which to back out of that decision in a in a quick order. Um, if you are making a one-way door decision, something that has a huge blast radius could really impact a lot of things, um, or there is no backing off. Once you've done it, you're you're you've done it. Um it's there's a very different thought process that goes into that. And I think that human-assisted um in many cases and a lot of times in a lot of pen testing environments, um, at least having you know human eyes on some of these decisions. And when AI is able to daisy chain or to chain um, you know, attack vectors together, um, that's where you start to get into, okay, it's found one. What is the blast radius from now finding another one? And wait a second, now this is a root uh level access. Maybe it should be asking, should I actually act on it and what is the impact? And it should be able to provide an assessment for a human real quickly, to you know, if we were to do this, this is what what I have access or what we have access to then. And um, if at any time that were to say, oh, if we do this, it potentially could get to customer data or to anything, time out.

Shahar Peled 26:59

That's where the human is definitely gonna be like, well, wait a second, we're and these are pretty much guard reels saying here's what you can do autonomously, and here is where you need to have human in the loop or human-governed autonomy or whatever.

CJ Moses 27:10

And that's I think a lot of it is is setting the system up ahead of time such that you know that's where it knows it can't go any further, it has to ask for a human. Yeah, and a lot of that is through, you know, trial and error trip testing over time, figuring things out. And we're still very early in those days, as as you know, to understand where's what it's best. So the humans are more involved now. The gotcha with having the humans involved is that initial thing we talked about, the speed, the time.

Shahar Peled 27:37

Yep.

CJ Moses 27:38

You add a human in and automatically you slow things down by a good bit. Uh, you know, even if the human's sitting there waiting, it's gone from seconds to minutes at a minimum. And very rarely do you have humans around the clock sitting there waiting for AI to deliver.

Shahar Peled 27:50

This is where the ICD pipeline is a great way to inject things. And yeah, I mean, the recipe is how can we exponentially increase velocity, speed, and accuracy, but not risk safety or flat or uh stability. And that's that's that's crucial. And I think this recipe is going to change over time as the technology gets better and we trust it more.

CJ Moses 28:10

And the trust, that's the experience. It comes with experience, just like anything else. Once you have the experience, you get more comfortable. We we've done this with just normal non-AI types of automations. Um, I actually I'll give you a physical example from the old days again. When we first started doing, uh most people won't even know what this is because it doesn't exist anymore, I don't believe, a SAS 70 Type 2 um accreditation or uh, you know, we do a document to basically get audited against it. When we first started doing it back in 2008, 2009, for a company to see our results, um, used to hand carry it out to their office, let them read it, and I'd have to take it back.

Shahar Peled 28:53

It feels very, very old.

CJ Moses 28:55

Doesn't it? And it is, because I'm very old apparently. No. Um but um, and then the walls got pushed out, and they realized, wait a second, there's not this isn't like damning information or anything like that. So we created a reading room online, and then it's kind of gotten to the point where reports like that these days are business norm for us to share, and there's not it's not a secret. Yeah, so we have a means by which to share lots of accreditations, lots of reports, lots of compliance-related regulatory stuff in a much more easy easy fashion. But that's because we had experience with it and got used to and comfortable with um it's not going to harm us um by putting it out there.

Shahar Peled 29:34

Same thing applies here, and um two things, yeah, two things need to mature: the technology and the market with it. Um you talked about speed. I think you know speed better than most diesels around the world for sure, as you race competitively. Um, I think some some analogy can happen, you know, between racing and an autonomous system in cybersecurity. You have your team around you, you have the car, um, you drive 150, I don't know, 200 miles per hour. Um do you see any any correlation between your role as a security leader in this fast-paced Amazon environment to how you would manage a race?

CJ Moses 30:11

Absolutely. So um on a track, um, you're actually building trust in the system, aka the car in this case, um incrementally. You don't just go to the brake zone and just say, okay, I'm gonna try braking here. You actually work up to a point then you realize, oh wait, I blew past the braking zone, and I guess I need to back it up just a little bit. So with that experience of understanding where um where you can operate, um uh, you know, the other thing is that you know the data from the crew is accurate, you know, what they've done with the car, you've got trust built in the car, um, the telemetry that's being provided, um, the way that ultimately the car responds to what I do is important. And therefore, that's kind of the analogy. You know, you start to earn um trust and therefore speed uh in the race car. And the same thing applies. We were just kind of talking about this. Um, you know, if you're not used to the car and things like that, even when I get into a car for the first time, even if it's my car that I've driven a bunch of times, I'll get in the car the first or second, first and second lap are okay, does everything feel right? Does it do the same thing? So after the first lap, the second lap, I'm gonna push the braking zone to make sure that stuff all works. And on the third lap, we're gone. We're we're qualifying, man. We're going you know, full out. Um, and that's because it's a build, built trust. Same thing applies in this space. Um, you know, and I think that uh it kind of it helps me in that way to you know understand it in the environment.

Shahar Peled 31:45

So I'm gonna get to see a race very soon, so I'm excited about that. One question before we we go to the closing. Um a few months ago, something changed in cybersecurity and AI. And again, even before that, you were one of the early leaders that saw the effect of AI on cybersecurity. Um, yes, there's cyber for AI, there's AI for cyber. Let's talk AI for cyber for a second. A few months ago, uh, Mythos was launched.

Speaker 32:09

Yep.

Shahar Peled 32:09

And ever since we've seen other mythos-like models from other providers and new improvements. Every security team, every board is bringing them up. They asked, are we protected against a mythos ready world or against a adversary is using mythos-like models, et cetera, et cetera. Yeah. In your read, how much of it is is actual exponential changed in the technology, and how much of it is marketing um or people just being secured of the amount of hype around it?

CJ Moses 32:39

Um I think I'd say that the impact is real. Um and it's not going away. Um, and what I mean by that is that um if you look at the amount of vulnerabilities that Mythos uh is you know finding um so quickly comparative to what we had pre pre-Mythos. So kind of that's that's kind of a change, a watershed moment, if you will. And we've gone from you know this number of CVEs to 10x that, you know, like overnight almost, it seemed like, and in that in that quarter after release, it's been been crazy. Um and uh it really has changed things, all of the things we could just kind of talked about to some extent, the speed, all of these different things. I think that there were there is a uh uh there was and potentially might still even be, a bit of a misunderstanding, somewhat in the security space, um, definitely in the boardroom, that this is a one-time uh change, that we've hit, you know, we've we've gone to this, we'll get through all these vulnerabilities, and once those are gone, then we can go back to the way things were. The reality is that that isn't the case. We have now hit the next plateau of this is, you know, the latest AI is the best and greatest that you'll ever have, but it's also the worst because the next one's gonna be better, and there's gonna be a new step change as we kind of go forward. So this is the uh new way of doing business. And those that try to respond to this moment in time by surging people or doing more of what they used to do, trying to patch things in the same manner, um, are always going to be behind. And in some months' time, they'll realize that if you take this moment in time and you change the way you respond by using AI to defend against the AI, use AI for the patching, the quick CI CD models, things of that nature, um, and embrace that that this is the new way of business using uh that technology as well as pen testing uh continuously to make sure that the things that you've done are actually enough using the intelligence that you garner to be able to understand what are the things I need to really get patched now and put the time into it. Um, you know, using the one-way and two-way door decisions to get as much of the one-way or the two-way door stuff done. You can get the two-way stuff done by AI, that means that you may have humans that can actually pay attention to the one-way stuff to make sure that the human is in the loop. So kind of going through all of that, there was a lot of hype, and I think there still is you know hype around it. But there's actually real things that need to be done. There's real impact that needs to be uh looked at mitigated.

Shahar Peled 35:23

That's a really good answer. But how do you explain that with all this hype, with all this technology um improvement we see? We haven't yet seen a crazy new zero-day AI adversary. Yeah, we heard about the Hagen Face OpenAI incident that happened. I don't know how much business impact was there. Do you think it's coming and we'll see immense business impact, breaches like never before? Or because currently it doesn't have a lot of effect in reality on the organizations getting breached and impacted.

CJ Moses 35:51

That way, that's the thing, that we know is always the thing that worries me when we get into this space. Because the threat actors that you need to be, you know, if you're a big company like ours, um, you the threat actors that you need to be worried about are aren't the ones that come storming in using the most the most recent zero day or even something that needs to have been patched years ago. Um, it's the ones that come in quietly that you don't know are there. And um, you know, some of that you know could be happening. I'm not saying it is, but my history and experience tells me it is, um, and that there are things that are occurring that we need to uh um be mindful if if when we're pen testing, we're getting in.

Shahar Peled 36:38

Adversaries can do the same thing.

CJ Moses 36:40

Adversaries can do the same thing. So that's why we're pen testing and obviously mitigating those things in our environment.

Shahar Peled 36:45

It's also why it's important to pen test on production or something very similar, because otherwise you're not mimicing adversaries.

CJ Moses 36:52

Absolutely. So that's the thing is just because the big thing hasn't happened yet, and I I would actually say that the um open AI hugging face thing is kind of one of those things that was an eye-opener awakening type of thing because we knew it was gonna happen, something like that, but it happened really quick. It didn't take long before, and um, you know, even some of the things that we heard um, you know, from George Kurtz and the Falcon uh keynote in regards to the way that things were daisy chained together and the agents were working together. Um they're being sneaky about it. And they were being sneaky about these things that all of a sudden you start starts to make you think that um technology is there, it can happen. I think to some extent it may be happening. And we we have seen some of the AI assisted things I talked about with the DPRK is doing. Um, we've seen other other instances um, you know, where you know that's one of the nice things about uh our uh our nice uh uh honeypot environment, uh Madpot, because we can actually see what adversaries are doing, and we've seen adversaries using AI because they're uh you can tell because it's 500 milliseconds between commands that are just like one thing's done, boom, 500 milliseconds later it's already responded and moved on to the next part of things, which it wouldn't have known a priori to actually put that command in. You can't even type that fast. Yep. So it's you know, on top of things. So, yes, it's real. Um, at the same time, um, I think that how you respond is what matters. Um, people that read you know that react to these types of things rather than respond, reacting is essentially hair on fire. Oh my god, we gotta do something, we gotta do things, is not the way you took it. You can paralyze the business. Yeah, no, actually. We need to be enablers. Yeah, you need to be enablers, you need to respond, and the response that's prudent is the one that's best for the business, and that is mitigating the threat while enabling the business to move forward. And that's a lot of kind of some of the things that we've done in the past, you know, with the CICD pipeline that is faster, enabled by AI, yeah, um, rather than slowing it down so that we can have more humans look at it.

Shahar Peled 39:03

With what you're saying, I I think it's coming. And that honestly is a bit alarming. Yeah, let's close things out. I I have my closing question. Um, let me challenge you. 12 months out. I'm not gonna say three years because no one knows what happens tomorrow, but 12 months out. What's one thing that you believe in AI um and cybersecurity that most of your peers would disagree with today?

CJ Moses 39:30

Well, we've talked a lot about uh finding vulnerabilities and kind of how easy it is now and how many there are. Um I don't know if people will agree with me or not, but um I think you know um the measure of a good security program um won't be how many vulnerabilities you found. Because I think for a long time it has been. It's how many CVEs can you put out into the terrible metric. And the metric is exactly wrong. Uh I think, you know, kind of the mean time to defense or whatever we want to call the exposure time um will be the metric that is used. It is the metric we are using. So there's a small group of us at Amazon that have decided this is what the future brings, and that is essentially decreasing the time between a vulnerability being found and it being patched or otherwise mitigated. Sometimes patching isn't the only way, you know, high patching, blocking, other things, isolation um can be there. Um you know, and I think that um that that window will be the thing that everyone, the metric that means the most to everyone.

Shahar Peled 40:39

Um I don't know if that's all that much of a hot take that it's it is pretty hot take because for decades people are just counting vulnerabilities. Yeah, how often did you fix them? What's the business impact you you created and saved for? So your hot take's biggest lane, less people are gonna care about how many vulnerabilities you patched.

CJ Moses 40:57

Yeah, but so many of them who cares.

Shahar Peled 40:59

Mean times the defense. I love that. Beautiful. That's that's super in super insightful. Um CJ Moses, teacher of Amazon. Thank you so much for joining me on the AI Native Cybersecurity Podcast. This was incredible. I really appreciate the time, and I'm sure our audience uh would love to ask questions, please send them my way. And I'll try to forward the CJ. I don't reply to you, but uh thank you so much for being here.

CJ Moses 41:27

Thank you for the invitation. It's been uh wonderful to do, and also look forward to uh to coming back again when when things things are inevitably going to change. We're gonna have more fun things to talk about.

Shahar Peled 41:37

I'll bring the quote from the hot take. Let's save it after it. There we go.

CJ Moses 41:41

I agree.

Shahar Peled 41:42

Would love to do that. Thank you so much, DJ.

CJ Moses 41:44

Thank you.

Intro/Outro 41:47

That's it for today's episode of AI Native Cybersecurity. Your attack service probably changed while you were listening. So take something from this conversation and go test it. We'll see you next time.