#313 Evan Reiser: How Abnormal AI Protects Humans with Behavioral AI

0:00Let's have you introduce yourself. Tell us how you got to abnormal AI and what abnormal does and then we'll talk about the current threat landscape. Sounds great. Well, first of all, Craig, thank you so much for having me on the show. My name is Evan Reiser. I'm the founder and CEO of Abnormal Security, which is now called Abnormal AI. And my background is not in cybersecurity. I spent the grand majority of my career getting people to click on ads. So if you imagine all those annoying, you know, the companies that maybe, you know, suck up some of your data and follow you on the internet with ads, showing you things you looked at on different websites.

0:33That's almost certainly my fault at some level. So I apologize. And now trying to fight back a little bit by, you know, stopping some crime and putting bad guys in jail. So I'm a bit of a cybersecurity outsider, but I've been an AI insider for a long time. As you probably know, behavioral eye targeting, it's all about looking at large sets of data, understanding, creating behavioral models of people, trying to understand what they do and they're exposed to different types of stimulus. So that's kind of how I got into cybersecurity just through, honestly, I love the technology. I've worked in AI for a long time.

1:01I saw the technology obviously getting way better even seven or eight years ago. It's changed a lot in the last 20 years. In the last two months, it's changed. So the idea for Abnormal was actually not really an email security company. That's what we're most known for today. It was really to create a behavioral security platform that could basically understand humans' identities better than humans and then predict their behavior. And the idea was if we could do that, we could probably help businesses and stop some crime. We started with email because phishing is the number one cause of breaches.

1:30Business email compromise, the number one cause of financial loss in the enterprise. And a lot of the conventional solutions are overly reliant on threat intelligence, different signatures, patterns, and heuristics. And what we see, especially in the age of AI, and due to very sophisticated criminals, there's new attacks every day. They're personalized, they're targeted, there's a lot of social engineering. Every attack's unique, you can't use some of the conventional techniques. So rather than kind of studying known bad, we focus on known good. We create behavioral models of every person enterprise by integrating directly into their IT systems.

2:01And then when a new email comes in or other activities happen for other products, we analyze it to say, is it risky and is it normal? And does it look weird? Kind of same way a human would look at that if you had 100 hours instead of 100 milliseconds. So we do that. And it turns out that technique is very effective for stopping different attacks. So we've been doing that for about seven years. We have a bunch of other products as well. The company is over a thousand people today. And today we protect about 25 % of the Fortune 500. So excited to go stop our crime in the future that's a quick yeah that's uh that's really interesting one thing about about phishing and i keep i'm always amazed when there is a major breach or attack and it turns out that it all started with a phishing email because you would think yeah at this point people would know not to click on things particularly on their work email but But the attacks are getting increasingly sophisticated.

3:02So the attackers are also using behavioral psychology, I would guess, to craft their attacks. How do you see the... First of all, isn't phishing kind of archaic at this point? And aren't there new attack vectors that are being developed? or is it still, does it still boil down to getting the person to click on a link? Yeah, I mean, some stuff is old, right? The days of like the Nigerian print scam, we don't see that. Which interestingly enough, my wife's Japanese American, her mother has passed away, but she was a Japanese citizen. And my wife just got a letter from, you know, it's the Nigerian print scam, but it's in Japanese from someone in Japan.

3:56I was like, wow, they're reviving that? I think. Yeah. But anyway, go ahead. So, I mean, we see some of those, right? But they're certainly less sophisticated, right? That's not probably the best way to go hack the bank. Same thing. We don't see typos in emails, right? There's no bad grammar. Every criminal obviously can use ChatGPT to send out perfect messages that are super personalized, right? They know all about the person that read every webpage on the internet about you. They know your coworkers. They know the business process. So things have changed a lot. What hasn't changed is still about kind of tricking humans, right?

4:28And that's going to be a consistent thing. And this, I'm going to make up a totally fictitious example here. But Craig, imagine you and I went to go create a cyber gang. We kind of gave up on our day jobs. We said, we're going to go hack the biggest banks in the world, right? How would we do that? Well, we're not going to break through their firewall. We're not going to hack their satellites. We're probably going to send a bunch of like, you know, text messages and LinkedIn messages and emails to all their employees, right? And a big bank might have 10 ,000 employees. Someone is going to do something or say something they shouldn't be, right?

4:54Someone will maybe think we're someone else and maybe reset a password or send us an invoice. Can we pretend to be their vendor? Right? I mean, that kind of human judgment is still like the weakest link in most organizations. You know, if you're a criminal, like why would you try to go up against the big bank? They spent a billion dollars in cybersecurity. They got every piece of tool, technology, every FBI threat to investigate. It's super hard. But if you email the poor support help desk person and say hey i forgot my password or you email their accounts you know accounts payable and say hey like i'm the ceo's blah blah blah and if you don't do this like we're going to delay our service and the flight won't take off and god knows what will happen to you like you know people are you know it's like 99 people are like good citizens are trying to be helpful and so i think you know criminals are really trying to kind of exploit the good intentions of humans they're trying to use our own you know people in kind of it systems against us that's very hard to defend against because unlike your you know network you can't shut off the fire you know shut off the ports and block access right you need to keep emails open for visits phone calls open you need to tell your staff that they should like you know try to listen to customers try to help them out and help out the co-workers and you know attackers are using kind of our good intentions against us yeah have you seen presumably you see these attacks have you seen them evolve and when you talk about an attacker knowing everything about you if it's corporation with you know 20 000 employees or something they're not doing that is it do they have a comparable ai system that's putting together all the various data points about each individual and crafting you know personalized attacks for each of the 20 000 or however many employees are the answer is kind of like yes and no right let me i'll show some context right just to add, you know, appropriate and hopefully not too much or too little credibility to my upcoming answer here.

6:43So I've been doing this for, I don't know, eight years or so. You know, we protect thousands of companies. We process more than a billion emails per day. A hundred thousand attacks get blocked every day, maybe a million. But I would say at least every day we find one novel attack that's never been seen before in the history of civilization. So, and I talk, you know, I personally, I talk to at least one CISO a day. I probably met half the global 2000 CISOs at some point in the last couple of years. So like while I'm not a cybersecurity expert in some ways, I am kind of like Wikipedia for like email scams and fraud and social engineering.

7:14I've heard about everything. I could someday I'll write a book, a very long blog post about the top thousand attacks of the century or something. So the answer is like, yes and no, right? No in the sense of like, you don't need hyper sophisticated AI to get people to like do some they shouldn't be doing, right? Sometimes it takes a basic email, one line, right? You go register fake Gmail accounts. You create a fake LinkedIn profile. You send one little message. And it's like not that hard. You don't have to be that sophisticated. You don't have to have a PhD in computer science or machine learning information retrieval or like high dimensional math to be effective.

7:43So I think from a volume perspective, you don't see that many high risk sophisticated attacks, which is not necessary. However, they do exist, especially when you look at kind of organized criminals and nation state attacks. They are using kind of all the technologies you can imagine, right? Again, the same things that you and I would use if we were the bad guys. But I think that the asterisk there is like, they don't need to go spend a billion dollar training custom ai bottles right you can use off you know off the shelf open source kind of large language models any of the foundation models that are available for free on the internet and you can just ask it hey i'm a small vendor i have some financial difficulties one of my customers the big bank you know you know my bank account like there's some issue and i had to open up a new one can you help me write an email to explain why it's really urgent that they need to change my routing number because if they don't do this right you know i'm not going to make payroll to my three employees and like, you know, that's what I rely on.

8:32Hey, here's the person. Here's every email that we've had in the past. Here's kind of like some information. Here's their Twitter page or LinkedIn page, right? Like, can you help me imagine like, what do you think the process is at a big bank, right? To change that, right? What are the compliance, you know, hurdles? There may be ways to get around it. Come up with 10 ideas and draft me a hundred different emails, right? That you might think might be effective to get that person, right? And then you can go send those, right? And so within five minutes, you're sending a hyper sophisticated, targeted, personalized phishing email that probably like five years ago like that you could only deliver that right if you were like a nation state actor and now a petty criminal right even with no computers you know technology knowledge can be kind of working at levels of efficacy so i think that's the challenge last comment i'll stop talking about like that's the challenge i think with the these attacks in the age of ai is that ai allows non-criminals to be criminals because i'm not sorry non-cyber criminals to be criminals right it allows criminals to kind of increase the scale of these attacks where they can automate it but it's through fancy python scripts or ai and the third piece which is most disturbing is that ai allows for a level of sophistication these attacks that are beyond kind of human imagination right we literally see i personally see maybe no maybe not every day but once a week there's some new attack where you look at it like i don't think that's an attack and i would say probably yeah almost every day a customer complained to us say hey you blocked an email that was a safe email and we go investigate we say no actually this is sophisticated let me explain what they're doing to kind of trick you so like we're now at the stage where these attacks are kind of surpassing the human's ability to perceive them wow yeah that's frightening yeah there seems there seems to be an uptick in i mean it seems like everybody knows you know three or four people that have been scammed out of a bunch of money the but also it's like it's not like it's kind of like not the most interesting or sexy thing to talk about.

10:26I think if you look at the FBI has this interesting report where they publish for people when you lose money, you don't you know, you don't call the New York Times always. You don't go on Twitter X. You do call the FBI. Say, let me get you. How do we get my money back? So the FBI provides this really kind of, you know, I would say like a most a more objective view on kind of the where what cyber crimes are causing what type of damage. So I have this report which describes I think they call it the IC3 report for the Internet Crime Complaint Center. And it basically kind of talks about when people call us and they say there's money lost.

10:53We ask them, how did you lose it and they kind of track the money lost by cyber crime and so the number one cyber crime i think outside of like i think ftx kind of messed this up so maybe minus ftx and investment fraud but the last seven years the number one has been social engineering through email there's actually more money lost for that for i think then credit card theft identity fraud and like ransomware combines right yet if you look at like the visibility of some of the you know social engineering via email right in in media and even cyber crime kind of news it's very underrepresented right so it's kind of interesting how it's kind of like boring right one said if a team clicked on a link where it's kind of not super fun to talk about yet that is still you know the number one cause of breaches right so it's the difference between like what's actually happening versus you know where kind of our attention as an industry and society is yeah and when you say that you see attacks that you've never seen before there or this example of an email that a customer complains has been blocked and then you show them can you give us an example that is sort of out of the ordinary that demonstrates that sophistication i'll give one that i think is extremely ordinary but my guess is for most listeners they've not considered this that i've heard about it right so one of the most it's not the most common attack by like you know the number of these we see is pretty small but if you look at like if you look at like okay if money was lost how did it get lost it shows up very highly there and that is thing called well i don't know what it's called we call it a vendor account compromise and so imagine you know you're you're a big bank super hard to hack in the bank right so if i'm a hacker i don't want to you know it's like it's silly it's a waste of time i'm not going to hack your firewall or break through cryptography what i can do is i can find out who your maybe water supplier is new york city right and i'll go hack them right because they probably don't have you know 2fa and like you know any cybersecurity training.

12:40So I'll break into their account, right? Maybe I'll go and LinkedIn, I'll search for the accounts receivable person. The person's kind of demanding invoices. I'll break into their accounts. And once I'm in their account, I have access to all their old emails, right? And so what I can do then is I can take every email I've ever sent, which might be, I don't know, 10 ,000 emails, and I can load it into, you know, a large language model, right? With a chat GPT or, you know, an open source one like Lama, I can do it on my laptop. I can say, hey, go through every email I've ever sent and find every one we've ever worked with, everyone we send an invoice to, find out their account number.

13:09And then I want you to send an email to everyone that we've ever worked with in the past and explain why they need to send us some payments, right? Use their real account ID, use their real name, use their email address, and then look through our past conversations and make sure you include in that message, right? Any, some sort of indicator that I know them personally. And so we'll see these kind of personalized messages. And it's really hard on the victim side. You get an email from somebody worked with for 10 years it's the real domain right it's it's like the real email address it's replying to an old email right that where you talk about something and in that email references i don't know hey evan it was great seeing you you know i hope you're recovering from your sickness from like the holidays or something i don't know right yeah and it looks your personal and there's some like reasonable pretext or excuse hey we're at you know sorry repeat this all is there but like you know hey we were at svb and our bank account got shut down and you know blah blah blah.

14:01Right. So like, you know, you just wouldn't expect to get a an attack from someone you've worked with for a long time with a real email account referencing things. And so like when you imagine from the attacker side, it's not too incredible. But, you know, if it's Tuesday morning 8 a.m., you get that email, it seems super reasonable. You wouldn't even think that like you should think about that attack. So that's an example, right, where people are obviously using, you know, generative AI technologies, but they're really kind of relying on just getting good intentions and trust and digital identities that we're so kind of accustomed to in 2025.

14:31Yeah. And now that we have sort of agentic automation percolating through all these systems, presumably, you know, the criminal could set up one of these systems and it sends emails daily, you know. Sure. Yeah. Yeah. Non-stop, right? I mean, it's just seems like it would be an overwhelming amount of attacks coming now that AI is hooked up to it. Yeah. And like, I think the only reason it's not overwhelming is because you don't need these old specific techniques to like make a lot of money as a scammer or a criminal. And it's because like the basic stuff still works, unfortunately. And there's a new attack every day.

15:14You know, I think we will see more, right? I mean, one, I don't know our official data point is here, but like I'll give like the plus or minus 25 % version. When we deploy, we are kind of the last line of defense. So we sit behind every other line of defense. So there's things that process that may filter emails. We are kind of the last kind of check before it gets in front of a human. We so that means that if things would get blocked higher up in the stack, we can't see it. We can only see stuff that like would normally get sent to the user. So in the last like last 12 to 18 months, we've seen about a 2x increase in the number of attacks that we see, which means that with abnormal removed, it's right there'll be two x more getting in front of kind of our customers employees mailboxes so and i think that all the systems around around us right they're also getting better but i think the criminals are kind of outpacing like good guys and so just these attacks are very effective and you know i don't know if it's because of ai or other reasons but we are seeing just you know higher volume of attacks and for the attacks that are for the top one percent top one percentile most sophisticated those are now past the point where humans can recognize them we have yeah like I said, we get complaints about people saying, hey, you missed this attack.

16:16And sometimes we do, right? But like a lot of times it's because like it didn't miss it, just the AI picked up on something that, you know, the human analyst didn't understand or doesn't even really kind of believe even when it's explained, you know, by the AI. Yeah. And so you're using, you're training a model for your system, which presumably is a bunch of models to understand the norm in an organization and flag things that deviate from that norm. Is that right? That's right. And like, it's, you know, in some ways, like it's not that smart, obviously, like, I think it's pretty cool technology, but you know, most of cybersecurity historically has been focused on just detecting kind of known bad stuff.

16:58And so, and that's not how things work outside of cybersecurity, right? I think like it's intuitive to most people how credit card fraud gets stopped, right? And so let me use that as kind of the analogy here. So that's probably more relatable. So imagine like if every time you use your credit card, the bank said, okay, is the vendor that Craig just sent money to, is that a known scammer? If so, block it. If not, then it's probably okay. That's how most of cybersecurity works. Now, as you know, that would not be sufficient because there's new scammers targeting new people with new kinds of trends to steal their credit card, steal money all the time.

17:28And so obviously, if you get a credit card charge on your card and it's from a coffee shop in Brazil and the charges for$10 ,000, your bank doesn't need to believe that's a scammer, right? They just have to say, Craig, I don't think Craig's in Brazil. I don't think coffee shops historically haven't spent $10 ,000 for like a latte. And so we're going to block this one to play it safe. And maybe we'll just kind of send, you know, Craig a note. If it's on the fence, we'll just ask him. We're going to default to no. So that's kind of more of a behavioral-based approach. So that's kind of the technique that we do, right?

17:57So when we see a new email come in or maybe a new login, right, to kind of Microsoft or Google, what we're going to do is we'll go, what do we know about this person. We have a pre-computed behavioral model. Well, here's who they work with. Here's the time of day they're active. Here's their tone and their personality, their style, and the projects they're working on. And we say, okay, what do we know about the sender, the receiver, the topics, right? Like we were talking about some business reports. If I send you an email saying, hey, Craig, thanks so much for that analysis you sent me, right?

18:23Ari, I'd be like, that's not a thing we've ever heard about. So you basically kind of look at all these entities, right? The identities and the people kind of all together. And then we're using it as context to make a decision. And the decision we're making is really kind of across two dimensions. One is, is it risky? If I got an email from someone I've never seen before and just says, hey man, great seeing you last night, that's not risky. So we're going to kind of ignore it. But if it is risky, right, saying, hey, can you pay me soon? And it seems abnormal, then we're going to block it. So we look at the kind of utility, the multiplication of these two things, right?

18:54Is it risky and is abnormal? And if so, we don't know that it's bad. We just know it's not good and we block it. And And so that's kind of the difference. Most cybersecurity, like we said, focuses on kind of known bad stuff. We basically kind of flip it where we say we're only going to let known, you know, kind of known good or known normal stuff in. And the reason that's like really important in this age of AI is that if every attack is unique, you don't have like training data for your machine learning models to help you stop the next thing, right? If I knew every attack in the world and I built a model to say, you know, is it an attack?

19:23I can't predict the next one because it's never been seen. It's outside the trained data set. But we have a lot of data on what is normal behavior, right? But every day, all of us are kind of leaving behind this giant exhaust stream of like every click and every keystroke, everyone we talk to, all things we do for work. That's a lot of training data for known good stuff. So we use all that stuff to kind of train our machine learning and AI models to figure out what's known good. And then if it doesn't seem to be good, we block it. And that turns out to be extremely effective. In some cases, it's like two to 10x more effective than the conventional threat intelligence-based approach.

19:54Yeah. And you were saying this is the last line of defense. so it doesn't replace other filters or other, you know, alerts that a system may already be using or a customer may already be using. Is that right? Kind of like yes or no. You know, almost every business in the world has some form of what's called a security gateway. And that is kind of like the conventional email security technology. If you use Microsoft or Google as a business for your email, you have that built in, right, to provide native security. made, you know, sometimes bigger enterprise will buy third party secure gateways. We're kind of agnostic.

20:31We say you got to use some secure gateway because there's some unique technology there for stopping spam and stuff that you're going to need. And you can pick a great one that's probably a third party or you pick the free one that comes with, you know, Microsoft and Google, which is our recommendation. But whatever one you pick, we're going to make sure that anything gets through those things. We're going to stop. It isn't good for any users. So, you know, there is kind of need of security in Microsoft and Google, right? That is the process of these messages. And, you know, it does stop a lot, you know, a lot of more of that kind of spam, maybe the more commodity type attacks.

20:58But we do replace kind of third party solutions, right? That's kind of like more conventional third party email security solutions. And that is one of the reasons why people buy us, right? They want to stop more attacks, but also like not pay for the old kind of legacy thing. Yeah. This, so you're coming from the world of ad techs and you, what did you study in school? I studied stuff in a school I've never used professionally. So I went to school for electrical engineering, which really is, sorry, computer engineering, which really electrical engineering for the purpose of building, you know, computer stuff, right?

21:28So I took classes on CPU architecture design and networking and things like that. I then decided that I couldn't stand the slow iteration rate of like building CPUs and circuits, right? I couldn't wait six months for it to be fabricated. So I really, I got into kind of really web-based software development and, you know, just fell in love with, you know, programming and software engineering. I didn't really kind of study that in college. I never programmed before. I went to college, took a couple of computer science classes. I just loved it. I wrote, built my own video games. It was like so fun.

21:55And then after school, I lasted about two years in the industry and then kind of have been doing startups since, but just really love kind of building stuff for people. And working with kind of machine learning has been just really fun. You can do kind of these magical things and like the industry's been changing so fast for the last 20 years, which is very intellectually stimulating. So yeah, my background is in mostly software engineering. Then I kind of got dragged into kind of product management because I love building stuff with customers. And then you've spent a long time in ads and now get to build cool stuff with AI for our enterprise customers in cybersecurity.

22:24Yeah, I was asking because I thought maybe you came from a behavioral psychology background or something. So I don't, but actually my dad is a psychologist. And because of that, I've got really, I've always been very interested in behavioral psychology. I've, I don't know, I've read, I don't know, dozens, if not a hundred books. And I've always just, you know, my specific area of interest that was actually even pre, you know, predates my cyber-dressing interest is I love reading about cognitive biases, right? Understanding like the evolutionary basis for why your brain works a certain way and like why it's like really set up to work in like a world that is not the modern world and just kind of understanding why we make some decisions versus others.

23:04That's more, I would say that's more of a hobby interest, right? I have like zero academic training and probably don't know half the right kind of like words for things. Yeah. And, but do you think as, I mean, LLMs, for example, as they become more nuanced in understanding human psychology, which they clearly are. Do you think that, I mean, everyone talks about it's an arms race, you know, the attackers get a little ahead and then the defenders catch up and then the defenders get a little have back and forth. Do you think that this will plateau, that there is only so much social engineering that you can do and that as new defense systems like yours develop, that the attackers will run out of new ideas?

24:01It's an interesting question. I don't think so because the AI is going to generate ways of attacking us in ways that we can't imagine. I'll give you one example. The other comment I'd say is it's a bit of an asymmetric war. And it is true, there is kind of like a cat and mouse game. They try this, we block that. They try this, we block the next thing. And it is true that there is a new attack every day. And that requires companies like us and all of our peers to have new defenses every day. And that is the bare minimum, just like tread water. You can actually still be losing. You can be getting better every day and still underpacing the attacker.

24:34And so it's a very interesting industry. And frankly, it's very intellectually stimulating to work in an industry where it's actually kind of like challenging and fun to like know that if you don't get better every single day, you lose. The reason it's kind of asymmetric is because attackers, they don't have to, they have a very high tolerance for being wrong about stuff. They can try a thousand attacks, you know, they only need one to work and they don't have to worry about privacy or security or compliance or like obeying the laws. They can try crazy stuff. They usually don't have to deal with like supporting a million customers, 100 ,000, you know, coworkers.

25:04So they have to like only get it right at once. The good guys on the other side, they have to be right every single time, right? The one time they mess up, and it's not just a breach. It could be like they roll out the AI model too fast. They don't have the right compliance for the new privacy law. They can only get, you know, they have to be right every single time. So it's a little bit unfair, right? Even just the adoption of AI, right? Criminals are adopting this technology instantly. You know, ChatGP came out three years ago. I'm sure there's some phisher using it the next day. And there's some enterprises today They're still trying to figure out, hey, how to roll down a safe compliance way.

25:33And so it's a bit of an asymmetric war. And so it's kind of unfair. Right. And I think that's the, you know, I do think there are some advantages that the defenders have. But I think to your question, like, I think it's going to be, you know, the we're not going to get to it. I think there's going to be always new types of attacks that are going to escalate. I do think the defenders will have the data advantage. Right. We have, you know, we'll have more data about inside of like how the business works and how people behave and the bad guys. Right. There's more data inside the enterprise and outside.

25:58And so I think that's why defenders will win long term. And just the final thing I'll say is like the future of cyber attacks, even cyber warfare is going to be very strange. The same way if you explained your cybersecurity, someone from a thousand years ago, they wouldn't know what you're talking about. I think it's the same thing. Like 10 years from now, there's going to be cyber attacks that we can't imagine. I'll give kind of one example. I talked to one of my, maybe a friend is too strong of a statement, but like there's a customer I'm close with, who's a CCO of a big bank that we've all heard about.

26:25And I said, Hey, what is your biggest worry about kind of AI? What's the biggest AI attack you're worried about, you think no one is talking about? And what he said is that, he's like, look, there's a lot of stuff I'm worried about that we're trying to actively face, but there's new class of attacks I think we can't even imagine today. What if like one of these models had a small cultural bias, right? A malicious cultural bias that was not observable, you know, on a day-to-day basis. But every day, right, it was giving answers that would make all of our employees maybe 0.00001 % a little bit more socialist.