Ep. 349 AI Is Changing the Speed of Federal Cyber Attacks Today

Speaker 1 0:00
Hey, John Gilroy here. Today, we’re going to discuss insider risk on steroids. Hit the music, Manny. Welcome to the Federal Tech Podcast. My name is John Gilroy, and I will be your moderator today, face to face. Adam Myers, head of counter adversary operations at CrowdStrike. Adam, how are you?

Speaker 2 0:22
I’m great. How are you?

Speaker 1 0:23
Listen to the background here. We are recording this from Billington Cybersecurity in lovely downtown Washington D.C. and Adam just got out of a group presentation. It was fascinating, wasn’t it?

Speaker 2 0:35
Oh yeah, it was a good time talking about living off the land techniques. Yeah,

Speaker 1 0:38
listen to this title: “Hunting for living off the land attacks for the federal government. You can just gone for hours about this, couldn’t you? Oh,

Speaker 2 0:46
absolutely. So many

Speaker 1 0:47
things. Well, I want to bring up some salient points that you made that I wrote down that I was really fascinated by. You talked about breakout time. Yep. And I I wrote that down and I said, Nah, not believing this. But the breakout time is just incredible once they get in, isn’t it?

Speaker 2 1:02
It’s crazy how it’s accelerating too. You know, we started tracking this years ago, and it was hours. And you know, in 2024, the average breakout time-how long it takes for a threat actor to go from initial access to moving laterally into the environment-was 48 minutes. 2025, it dropped to 29 minutes, and the fastest breakout time that we saw was 27 seconds.

Speaker 1 1:24
Seconds, wow. Okay, I’m going to ask a softball question. Do you think that’ll be impacted by AI?

Speaker 2 1:30
Well, it already has been, and I think we we just announced last week at our Falcon event in Las Vegas that we could see this getting to zero, and we’re we’re actually looking at new metrics to replace breakout time because the speed, which is you know gotten faster through the years with humans driving it, now partnering with AI or or even letting AI kind of take the wheel, it’s gotten even faster. And so this is getting to to ridiculous amount of speed. And one of the things that we saw that I talked about was that we we noted one of the threat actors we track, which is Vault Panda, has brought AI into as like a copilot for their intrusions, and we can watch it learn. We could see it tried to do a web request, and that web request didn’t get anywhere. So then it tried to do a post, and then they realized that it it needed a cookie. So it went through each step, but it went through the steps in in you know seconds, and that’s the thing that really I think represents the speed of AI is just how quickly these things can learn and adapt to this environment.

Speaker 1 2:30
Yeah, it’s amazing. Many brilliant people on your panel. There’s a guy named Taylor Hanson there, and I think someone from the audience or the questions. Well, we have telemetry inside federal systems, and we can kind of figure out what’s going on, can’t we? And and he said, you know, there is a wealth of telemetry, but that doesn’t mean you can do anything with it, huh? Yeah, I

Speaker 2 2:49
think it was the CISA guy was saying it’s like you know he he was equating it to a needle in a haystack. Yeah. And what I told him was I think it’s more like a needle in a needle stack because when you’re looking at that telemetry, you’re looking at that data, it all looks like it could be interesting. So figuring out what is that particular needle that you’re interested in that moment in time is really you know the the sweet spot that you need to be in.

Speaker 1 3:11
I think you’ve been around long enough to know 10, 15 years ago I started talking about big data. I didn’t say good data, effective data. It’s almost like telemetry. Yeah, we got a lot of telemetry now. Now what? I mean,

Speaker 2 3:23
data lakes and data oceans and data moats and yes.

Speaker 1 3:26
Now let me bring you up to speed here. The federal government knows there’s insider threats, and they have procedures and processes for controlling that. Now there’s gaps in there, so so we know there’s gaps. So so how can they address some of these gaps, these challenges?

Speaker 2 3:41
Well, I think that it’s about visibility into the right sets of data, and you know, to your point about there being so many, so many, so much data or the needle stacks that are out there, it’s really important to understand that there’s different domains that security teams need to be responsible for and take ownership of. Everybody kind of thinks about endpoints, right? The desktops and the servers is kind of the main thing, but in recent years, identity has been one of the critical gaps that organizations have. They need to have identity threat detection and response. They need to be able to hunt across that identity data. Did a new user show up, or an existing user show up with a new multi-factor authentication device. Did they get a new phone, or did a threat actor get access to that account and provision their own MFA device? We’re seeing this in the cloud as well. Cloud control plane, a huge area of opportunity for threat actors. And if you’re a modern threat actor, you’re not hacking in; you’re logging in. So you’re going to log in with an identity. You’re going to move into the cloud control plane. You’re going to move into the SaaS environment. You’re going to go to all these places, and guess what? You never touch an endpoint. So having the ability to cross-domain hunt is absolutely critical, and I think that’s one of the big gaps that we have, not just in the federal government, but across the entire. Cybersecurity defense space is that organizations they don’t have the right visibility. They’re hunters. They’re they’re threat analysts. They don’t necessarily have expertise in cloud or identity, and so they kind of focus on the area that they’re comfortable. But by the time it hits an endpoint, there’s already been an identity compromise. It’s already moved through the cloud, and so you need to be able to have hunting across all of that to be effective.

Speaker 1 5:24
During your presentation, again, I’ve never heard this before. This is going to shock many of our listeners. I mean, set this up. So I taught at Georgetown for many years, and we have an annual hackathon. Oh yeah, hackathon! It’s big fun. Kids come in and have all kinds of fun. Now you describe what’s going on in China, where they have hackathons, it’s not that benign, and they offer cash prizes. So tell our listeners about this. I mean, this is a great story.

Speaker 2 5:47
Well, in China, they’ve been working on something called civil-military fusion for a while, and this is bringing together academia, private sector companies, and the military and intelligence agencies across China. One of their kind of premier, what we call pone to own. Think of it as you know racing for pink slips or something like that. Is that what it is? A pink slip for the car? I don’t know. I’m not a car guy. But effectively, what they’re doing is they’re saying, okay, VMware, or you know, pick a product and and whatever that product is, identity or firewall, whatever it might be, they offer a cash prize, and they say if you can find a vulnerability that you can exploit in this product, you’re going to win a cash prize. But those vulnerabilities, they’re not being sent to the vendor for going through the proper, responsible disclosure process. It’s going up to the CNITSEC and ultimately places like the Ministry of State Security, who are now going to use it as part of their offensive program. All of this has been backstopped with national security law. So China said, and I think it was 2018 or so, that if you’re doing security research, you don’t give it to the target entity to fix. You have to give it to the government, and if you don’t, there’s going to be consequences, and they have social credit and corporate social credit, so they can actually make your life very difficult if you don’t follow that process.

Speaker 1 7:06
Now at Georgetown, every class, even the math classes, have to have an ethical component, and we just talked about some ethics here. So in America, I would imagine if you find a vulnerability, there’s a process to let companies know. But over there, it’s exact opposite. It’s just it’s it’s brewing all kinds of problems down the road.

Speaker 2 7:24
Well, there is a process. It’s just not aligned with the public interest, right? So

Speaker 1 7:30
there’s a process. It’s a bad process, right?

Speaker 2 7:32
Here we do have a process, and it’s called responsible disclosure. So if I find a vulnerability, I contact the vendor. I tell them what I found. I tell them how to reproduce it. They go investigate it. They figure out if they need to make a patch. They file what’s known as a CVE, a common vulnerability exposure, and we set that all up so that everybody is safer. With China, they’ve kind of subverted that process, and rather than going through responsible disclosure in that sense, responsible disclosure there means that you give it to a government agency, who then decides what to do with it, and they may tell you to send it to the vendor, or they may say thank you, we’ll get back to you, and then they never get back to you. So it’s it’s all kind of it’s almost like crowdsourcing cyber weapons for their offensive programs.

Speaker 1 8:18
So you use the phrase cross domain hunting, and I’m sitting there thinking about that. And last time I was in San Francisco, saw that Salesforce. I go, wait a minute, that’s you know I have a system here, and I tap into Salesforce, and then maybe we have some in Amazon and some over here maybe in Google. Wait a minute, these are the cross domains. Just because John Gilroy has an expertise in AWS, what the heck that got to with Copilot? I mean, so cross-domain hunting. There’s almost like okay, put on your Philly hat. Now put on your Yankee hat. Now put on your Indian. I mean, it’s just that skill set is almost impossible to find.

Speaker 2 8:49
Yeah, I talk about that actually all the time. In fact, you know, I kind of found there’s eight different hats that your average threat analyst needs to swap out, and it’s not even domain specific. It’s things like, you know, one one minute you’re threat hunting, and then you find something. Now you have to go into incident response mode. So you have to switch your hat. You have to switch your tool. And each time that you’re switching that hat, it’s not just going to mess up your hair. There’s cognitive load that goes along with that, and you’re going to contribute to the burnout of that analyst. And if they’re switching constantly between these eight different jobs, I think about it as you know, what’s one of the most stressful jobs you can think of, right? What’s got the highest burnout rate?

Speaker 1 9:26
Oh, I’m guessing EMT.

Speaker 2 9:28
Okay, I would have said air traffic controller. Oh yeah. And so, air traffic controllers-they kind of have it easy, right? They only have to look at one screen and keep a bunch of things from bumping into each other. These analysts have eight different screens that they’re looking at, and each time they look at a different screen, all the all the all the things move right. So they’re they’re constantly going to be on the edge of burnout, and you know you get through a day and you’ve barely gotten the first thing that you set out to accomplish done. Now here’s the good news, is that AI can actually help those analysts. Can remove that cognitive load. We can allow them to switch those tasks more effectively by bringing AI as part of the solution.

Speaker 1 10:07
One of the distinguished members on your panel was Robert Carros from CISA, and he said, “What’s going on now is well, people see a problem, so they bring in a tool, then another tool, then another tool, and you get your old tool bag going. But the hunters can identify the tools now, and maybe modify those. So, it’s it’s it’s like whack-a-mole here, huh? Even the tools are being compromised. Oh

Speaker 2 10:27
yeah, it’s it’s asymmetric warfare for for the defenders. But again, you know, I think for the first time we’re at a point where the defenders now have tools that allow them to move with the speed and the scale that they need to keep pace with these adversaries, and that’s through bringing agents in. So rather than me analyzing a piece of malware, for example, and spending eight hours reverse engineering it with, you know, a disassembler and a decompiler and trying to figure out how this whole thing works, I can bring a malware analysis agent into the into the equation, and I can analyze that malware, find all of its related brothers and sisters and cousins, and I can take immediate action in seconds, not days or hours.

Speaker 1 11:09
We’re in Washington D.C. I’m sure if you get an Uber and drive around town, you’ll see all kinds of pillars. And for some reason, the federal government likes to talk about six pillars, three pillars, and five. And when they well, it’s on all their

Speaker 2 11:20
buildings. Yeah, well,

Speaker 1 11:21
that’s why. And so when I talk about zero trust, guess what I talk about? Talk about pillars. And the first pillar is identity. Right. Yeah, we know that. And so during the presentation this morning, Taylor Hanson said, “You know, identity is complex and poorly understood. I thought that was a guy who’s really been experienced and seen this. Huh? It is not that easy. Show me your driver’s license. Oh, good luck with that, huh?

Speaker 2 11:44
Well, it’s it’s the driver’s license is is the easy part of that equation, right? I think where we’re looking at identity, it’s human identities, it’s non-human identities like AI. It’s going to be you know temporary sessions that get established. It’s complex authentication flows like OAuth two, and because of all of the complexity with identity, people don’t you know they haven’t been trained in identity, right? People go to reverse engineering classes or they go to security incident response classes very frequently. You don’t find identity specific material or discipline or tradecraft. So a lot of folks are just kind of seeing that as you know an extra piece of the intrusion that they’re working. But really, it needs to start with identity. And one of the critical tools that I think a lot of organizations need but don’t have is what we call identity threat detection and response, which can identify things as simple as impossible travel. If I’m here in Washington D.C. and 10 minutes later I log in from a Wagadugu, then you know there’s a a clear issue that’s happened there, right? How did I get from here to there in an impossible period of time? And so maybe somebody compromised my account and logged in from there at the same time. So there’s all of these different things that identity threat detection response can help analysts start to identify that will be the first clue that there’s been an intrusion. Not working backwards and trying to go through the identity logs to figure out how they got in.

Speaker 1 13:15
Speaking of classes, we’re here at Billington. There’s 4,600 people here. There’s 250 speakers and 50 different, so all kinds of ways to learn things. These conferences, you you were at an event, and a lot of people in that room was jammed. They learned from you. Anthropic has all kinds of different things going on here. Earlier in the interview, you mentioned Falcon. Tell me more about that. Where does that fit in this whole solving the problem?

Speaker 2 13:35
Yeah, Falcon is the security operating system. Is kind of the way we like to think about it. It started off as EDR endpoint detection and response, and we added. You know, we we’ve kind of seen where the threats are going, and so we added to that vulnerability management. We added to that identity threat detection and response. And just last week at our Falcon event in Las Vegas, we announced Falcon Guardian, which is the evolution of AIDR, which is AI detection and response, because as we’re seeing more agents being used across all of the enterprises, we’re seeing more and more organizations implementing AI. Very basic questions like who’s using AI in my organization? What are they using to do this AI? What agents are they using? What platforms are they using, and where is the output of that going? And so these are the types of gaps that we’re starting to see organizations contending with. And you know, it’s it’s even you know unintentional these threats. You know, we will see AI are what we say are goal seeking, and we’ve seen incidents where a legitimate user was trying to build a tool to do something effective in their organization, and the AI said, “Oh, there’s a security tool or security policy in my way. I can’t accomplish my goal, so I’m going to work around that. And so, unintentionally, these AI are actually creating security incidents. In fact, I just announced this the other week, but for the. First time this this this year, we saw that the number of detections that are created by machines by AI versus humans is two and a half times. So two and a half times more detections created by AI than people.

Speaker 1 15:13
I wrote down the word guardrails, and so the unintended consequence is is we want to limit your car to going 45 miles an hour, but what if you have to speed up to pass the truck and you get hit by the truck? I mean, guardrail. I think this is the early stages of trying to figure out what the guardrails are, what they aren’t, and boy, so many different opinions on that, and it’s it’s so hard to figure it out.

Speaker 2 15:34
Well, great example of what you’re talking about there: the guardrail, you know, limiting the speed when you need to go faster. Hugging Face came out and said that when they were trying to investigate what happened with OpenAI, they were trying to use a frontier model, and they hit the guardrails, and they weren’t able to do it. So they had to pull one of these Chinese open weight models. I think they used GLM. They had to use that to help analyze what had happened in their environment. So those guardrails do sometimes act as the the third rail as a you know if you’re familiar with how trains work right so you don’t want to touch that one so it could be a problem

Speaker 1 16:08
I remember years ago renting a U-Haul truck and you couldn’t go more than 55 miles an hour causing accidents you know whether or not it’s good I’m writing notes here with ink probably shouldn’t because so many things are changing. I should probably use pencil or erasable ink or something. So when you look at first half of 2026, what do you think’s changed the most just in the last six months? A snapshot here.

Speaker 2 16:30
Well, look, I’d say just in the last couple of months, right? The mythos moment was in April, right? They announced that there’s a cyber capable model for the first time that was so dangerous that they not that they released it, but that they didn’t release it. Right, that’s pretty important. They they they said this is too dangerous to release. And then in May they brought together kind of a coalition of domain experts. They said we we’re not cybersecurity people, we’re AI people. Let’s bring a team together to kind of figure out what the capabilities are. And that was in May. And if you take a look at the vulnerabilities that were released in June as CVEs, 96% more vulnerabilities that were published in June of 2026 than June of 2025. That is a hockey stick of of new vulnerabilities, and that is you know the lowest you’re going to see it. In July, we saw triple-digit numbers. In August, we saw triple-digit numbers. So we’re eclipsing the number of vulnerabilities that were found, largely driven with the support of AI.

Speaker 1 17:31
Yeah. So Adam, if someone is in a U-Haul truck now and listening to this podcast, they may want to get more information. I’m going to direct them to this CrowdStrikes threat hunting report, and this is-I mean, this is really a place to sink your teeth into. Maybe get scared or run away screaming or something. But this is a great link for the listeners, isn’t

Speaker 2 17:51
it? Yeah, there’s great statistics in there, right? We we talk about a lot of the things, so that that report is generated from. We see 7 trillion events every single day, and those events are what feed that Overwatch hunting team, which is able to use that to go hunt across all of the telemetry that we bring in, and that drives new detection logic. It you know we notify customers based off of that, so it’s a pretty important statistic that feeds that report, and that’s where all of the insights from that report come from. So we use that to talk about what products are are problematic, what types of things you should be thinking about from a defense perspective to make your enterprise safer.

Speaker 1 18:31
I’d say a video I would love is that story you describe, where you’re watching the malicious actors step by step. If you can show a little YouTube video on that, you get a lot of views because people want to see because you’re like almost watching a thief trying to get in the house or something. I actually

Speaker 2 18:44
put it in a slide last week when I was at Calcon, and you can watch it build out and you can see like minute over minute as the AI learns what it’s doing wrong and what it needs to adjust. Wow, wow, wow!

Speaker 1 18:55
Well, unfortunately, Adam, we’re running out of time here at the Billington Cybersecurity Conference. This has been a fun interview. I’d like to thank my guest, Adam Myers, head of counter adversary operations at CrowdStrike. Thanks, Adam.

Speaker 2 19:06
Thank you so much. Great conversation.

Transcribed by https://otter.ai

Adam Meyers serves as CrowdStrike’s Senior Vice President of Counter Adversary Operations and is one of the company’s longest-tenured leaders, having joined in 2011 to build the intelligence framework that now underpins the Falcon platform. Mr. Meyers leads one of the most advanced intelligence organizations in the private sector – Counter Adversary Operations comprises more than 500 analysts, threat hunters, linguists, reverse engineers, AI researchers, and geopolitical experts operating across a global footprint. Drawing on prior experience at the U.S. State Department, Mr. Meyers built Counter Adversary Operations around a foundational conviction: organizations don’t have a malware problem, they have an adversary problem.

Counter Adversary Operations’s intelligence flows directly into the CrowdStrike Falcon platform. The team tracks criminal, state-sponsored, and nationalist cyber adversary groups across the globe, producing adversary insights and actionable intelligence to protect enterprises, major financial institutions, and critical government agencies worldwide. A world-renowned intelligence expert, Mr. Meyers frequently briefs senior U.S. government officials, the House Intelligence Committee, the Defense Intelligence Agency, the FBI, allied governments, and boards of the world’s most influential organizations. He is regularly interviewed on adversary threats across major broadcast, print, and online media, frequently speaks at leading industry events, and is the co-host of the Adversary Universe podcast, a biweekly show on the latest adversary trends and techniques. In 2025, Mr. Meyers was named to the CyberScoop 50, and SC Magazine recognized CrowdStrike with its Best Threat Intelligence Technology award.