AI has made deepfakes cheap and convincing enough to fool an entire video call—voices, faces, and all. In this episode, CTO Adam Blue talks with Jeff Scott, managing director of Fraud Intelligence at Q2, about why verification itself has become the attack surface, how behavioral and journey-based monitoring can catch what voice and video can't, and how AI-driven detection shared across financial institutions is starting to close the gap between attackers and defenders.
Account Takeover Is a Journey Problem: Here's How to Solve It
Stopping Account Takeover: The Power of Modern Fraud Intelligence
"Flash Boys: A Wall Street Revolt" by Michael Lewis
Thanks to AI and its ability to simulate voices, construct video, and construct very, very reasonable copies of popular websites and communications in the form of emails and text, fraud is getting more and more challenging to detect and to prevent. Join Jeff Scott and I as we talk about, in the fraud landscape, how we take AI and make it a defensive tool to help combat fraud and what the evolution of the fraud landscape is going to look like over the next few years.
Adam Blue
Hey everyone. Welcome to Cut to Context today. I'm joined today by Jeff Scott, general manager of Fraud Intelligence here at Q2. And we are going to talk about what is happening in the world of fraud and AI.
It's been interesting. Some of the things that we really counted on to distinguish human identity have become fully commoditized and reasonably easily reproducible. Voice, active video, even long form kind of integrations.
I watched a really cringe worthy video of Elon Musk that I assume was AI generated dancing around like he was at the beginning of Austin Powers. And I was so fundamentally and deeply embarrassed for him when I watched it that I could only imagine that someone used AI to generate it and shame him into, I don't know, doing something different.
But it really draws into stark contrast, what does it mean to authenticate an end user? What does it mean to secure digital banking? And so Jeff is joining us today because he leads the team of people that is at the very front bleeding edge of how we might use AI as defense and not just offense around things like dispute tracking, account takeover protection, and just general payments fraud. So welcome on today, Jeff. Thanks for joining.
Jeff Scott
Yeah, Adam, thanks for having me.
Adam Blue
So given that this notion, this thing we all talked about for a few years, the indistinguishable simulacrum of a human interaction, the deep face, has now become available for a fist full of tokens, what does that mean in terms of the way the threat has evolved in the fraud space?
Jeff Scott
Yeah, really I was with a customer recently who was talking about this story, you might have heard it, early 2004, where you have a finance employee at a big engineering firm, gets an email about a confidential transaction. He's kind of suspicious, he's got good instincts, he does the right thing, he asks for a video call. CFO gets on, colleagues are all there, faces, voices, you name it, wires $25 million. Every person on that call except for him was a deepfake.
Adam Blue
Wow.
Jeff Scott
And so the channel was the fake. And I think that's really the shift maybe that we're living through is verification itself has become the attack surface, if you will. And so the point that you just made, it's like what's changed isn't that deepfakes exist, it's that they've collapsed to basically be zero dollars to implement.
Adam Blue
Yeah, I think that's an interesting way to think about it. The other thing I wonder is in deepfake an interaction at some level, but in terms of how people talk and how they interact and that kind of thing, I think that's more difficult I think to overcome. And it occurs to me even in deepfake video, the longer the video is, the more likely there is to be a behavioral tell that the person is not actually the person they intend to be. Now, if you don't know who that other person is to begin with, you have no ground truth over what their behavior is, I think that's pretty tough, right? But I'm pretty confident that if Jack McBee, for instance, decided to deepfake me on a video call with you, four or five minutes in, you would probably feel pretty weird, right?
Jeff Scott
Yeah.
Adam Blue
There's a certain level of social nicety where you might not say anything because you don't want to seem rude or whatever. Power distance has a big impact here. This is why people pretend to be the CEO and go to somebody in accounts payable, right?
Jeff Scott
Right.
Adam Blue
Because they're less likely to question. But it's interesting, I think you can maybe deepfake what I look like and what I sound like, but still my behavior, the way I think, the way I talk, the way I assemble sentences, my rambling, my unnecessary metaphor, those things kind of belong to me, at least for now. And so where on the frontier does behavior end up being the way we dig in do you think maybe and try and get underneath some of this social engineering, deepfake AI driven technology?
Jeff Scott
Yeah, I mean, think about it's like the deepfake's the payload and social engineering is the delivery system. When you were just saying that about the nuances to you, there was a case that we had where somebody was faked because the accent of the individual was so close and so compelling that they were duped into sending the urgent payment or the transaction, as it were. And so it was everything around the voice that was right and made it contextually on point. And so that one detail was the thing that pushed it over the line, if you will. And so I think that when you combine those two things together in terms of the payload and the social engineering system, it becomes difficult.
And that's where it's like you've got to be able to monitor the behavior of the login or the behavior of the individual. We always joke about the analogy of when you get into your house, if somebody freely gives up their passcode to the keypad to the front door, the fraudster still inside the house is going to act very differently than the human was going to in terms of what they're used to doing. But you would have to be monitoring that person and that pathway inside the house on a regular basis to understand that Adam generally doesn't throw his keys down like that. And so if you're just monitoring at the transaction level or you're monitoring for one specific type of anomalous behavior or event, any one thing by itself isn't going to tell you much. It's like you've got to be able to understand the nuances of the behavior and the context for those specific journeys, if you will.
Adam Blue
Yeah, that's a powerful word in today's world, context. We talk about it constantly in all aspects of AI. And an account takeover doesn't start when somebody puts in the $25 million wire. It's not how account takeover works, right? That is the end of what I believe in the business people call the kill chain, which here's a sub-question while we're on this other question. Why do we talk about fraud like we're in a terrible Michael Bay extraction shooter movie?
Jeff Scott
I don't know. People like fear.
Adam Blue
I guess. Kill shot, kill chain, attack surface. I mean-
Jeff Scott
Right. Right. I tried to make it nice by saying journey, but-
Adam Blue
OK, journey. All right, let's use journey. So the attacker's on their journey, and there's just so much information that gets accumulated over the course of the beginning of the attempted crime and then the actual commission of the crime in terms of exfiltrating the funds. And so what are we leaving behind in that? What's there that the industry is not digging into? Because I’ve got to believe for all the deepfake stuff and the voice simulations and video simulations, the vast majority of funds lost are plain old, I got the password, I compromised your email, I went through an MFA loop, I did an account takeover, I moved the money out.
I mean, not every loss in the fraud space is some kind of Mission Impossible heist. They're the interesting ones to talk about, but I’ve got to believe 95, 99% of the money lost is still basic account takeover, and it feels like there's a lot that can be done there with this journey concept. So maybe expand on that a little bit.
Jeff Scott
Yeah, I think you're right. I think that the way that we build AI detectors now is around those behavioral pathways and journeys. And in the early innings of this, we're able to be highly precise on what a typical account takeover pathway looks like in terms of alerting it. And I think where this will head is as those pathways are taken and we can learn from every single event and get smarter, I think it will be harder for fraudsters to come up with different and new pathways and new ways to do account takeover. I mean, we do see it evolve.
I mean, the other analogy, I mean just to stay with the house alarm that we've sort of used is it's like community watch. It's like a burglar is going to case a house for a while where a specific thing that they're doing, like going up to the front door and seeing if anybody's home, in and of itself isn't going to trigger an alert. But over time, a smart monitoring system might say, "Hey, that person's definitely casing that house and we can monitor that behavior." Now everybody in the neighborhood has the information about how these attempted break-ins are happening and we can get smarter about that for the whole community. And I think that the place that we miss in the industry is that we don't have the ability for all of those neighborhood watch systems to talk to each other.
And so we're basically saying go put a fire marshal in every single living room instead of having a smoke detector alarm system that somebody could be monitoring at a central location. And so we've got to be able to connect the dots between systems so that they can all talk to each other and we can understand new and different pathways for different social engineering scams really quickly and act on them.
Adam Blue
Yeah. So there's a couple ideas that are intersecting here in a way that's pretty interesting. So you've got this network effect of sharing the data across people. So that's powerful and that's different. And then you've also got this notion of I think velocity really, right?
Jeff Scott
Yeah, totally.
Adam Blue
Which is it's terrible when anyone gets compromised and they lose funds. And then if they're consumer under Reg E, the bank's probably going to give them their money back, bank or credit union. If it's commercial, a lot of times that works out the way it works out. But if you're the financial institution, what you're really doing is you're operating on the curve of how quickly when a new vulnerability, exploit, attack vector, issue, compromise, password breach, how quickly can I react to that? Because the first couple days you're going to take some losses and that's terrible.
But if you can put countermeasures in place in those first few days, within a week, within two weeks, that's radically different than taking a month, two months, three months to respond in terms of the total loss set.
And so AI has made it very, very inexpensive to generate really, really good phishing sites, really great phishing emails that don't have the telltale grammar and usage and awkwardness issues. You can create high fidelity voice, face, all those kinds of things. And so how does AI on the defensive side, how is that asymmetry problem getting resolved as we bring that technology to bear?
Jeff Scott
Yeah, I mean, I can't deny that there's asymmetry in the fact that they have to be right once and we've got to be right all the time. But I think it can work the other way as well. And so as we learn about an attack or a different pathway that's new and we can build an AI detector around it, we can put that detector into production super quickly for a financial institution. And then all the financial institutions that are on that model should get the benefit of knowing that this specific pathway that was detected in North Carolina can now be used for a financial institution in Texas. And I think that network effect, as you sort of explained it, is what's new and novel and different. And so AI gives us the ability to watch all the houses all the time and all at once.
And if things get through, we learn from it very quickly and we can adapt very quickly, which I think has to be the name of the game. There's no more waiting for us to go code a model that can take a new pathway that has merged and put it into production. We've got to be able to do it in days or hours in order to keep the asymmetry in check. And I think it actually works against the fraudster because now if you think about what they're trying to do, it's like an ROI on a fraud attempt just got a lot more expensive because they've got to find all these different ways in order to hack the system because once we see it, we're able to basically lock it down across the whole network.
Adam Blue
Yeah. Yeah. So it's clear then that you don't have to stop every fraud. In fact, the level of friction, the level of investment there, that might not even be practical, right? There is some reasonable level of loss where the marginal cost and the marginal benefit equilibrate out. If we can use technology to reduce the amount of time it takes to react to new things that happen and if we can use the network effect of a lot of people participating in a broad fraud intelligence fabric framework, that you could make really pretty substantial inroads.
So maybe talk a little bit about what do you think at equilibrium as these technologies get deployed, what is a reasonable expectation for the fraction of fraud maybe that can be stopped for an FI? Where would they think about good being? What does good look like for them as they measure? Because I think this notion of, well, it's zero, that feels unrealistic.
Jeff Scott
Yeah, I agree. I mean, I think there's an ROI calculation that is quantitative, but then there's also a piece that's qualitative in terms of the customer experience that they're trying to deliver for the institution. And that's where we really think this is headed in the long run is if you're able to monitor with this level of sophistication and you're able to restrict bad guys and girls based on those risk scores, we also think that you should be able to enhance permissions for your really great customers and that should help from a customer experience standpoint.
And so I think you're doing this whole calculation of what's the friction level that we're willing to apply to get our fraud losses below some X number, and then you're weighing in the factor of how much does this total system cost us. And I think from our perspective, the financial institution that's around $5 billion, if they have a $5 million fraud loss surface, it's like if we can get that into a manageable range, call it 7,500 grand and you're paying some reasonable price for that protection, you could probably get to this place where you're like per login or per customer that we serve, what's the cost to protect and what's the cost per deposit? And you can get that into some sort of ROI that makes sense.
Adam Blue
Yeah. You said something interesting there that's a little paradoxical, but I think it's true, which is if you can figure out how to turn your approach to managing fraud into a control surface where you twiddle the knobs and dials, then you can make choices about how much risk you can bear for customers maybe that actually make it possible for customers to do more than they could do previously. You can actually improve the experience for them. You can grant exceptions when exceptions are necessary. I think that notion is really interesting.
And so maybe talk through, as we have more tools available and we can increase velocity and we can leverage network effect and we can stack that value up, what does great look like at a $5 billion FI that they have people that are responsible for fraud, but maybe they don't have a 10-person enterprise fraud team, right? What kind of structure do you think maybe that team adopts and what does their day look and feel like? What are the opportunities for them to evolve into?
Jeff Scott
Yeah, well, I mean, and in my conversations with fraud operations teams or the digital channel team that's dealing with some portion of fraud operations, they can't throw more bodies at this. Clearly that's not going to scale. Back too, I think I said it earlier, you can't put a fire marshal in every living room, right? You figure out a smoke detector system that you can have blink and alert you at the most important times. And so I think historically the alerts that we've used in the digital channel just weren't sufficient for the level and the velocity of fraud that we're sort of seeing today. And so you've got to have an AI-driven tool that's watching real time all the time and is highly sophisticated such that the alerts that are triggered, they know are of the highest and most important value.
So imagine you're sitting in a radar room and if you've got hundreds of blinking lights. You don't know which one to go triage first. AI helps there because it can tell you this one's blinking and it's blinking for a very bad reason and this is the thing that you need to go focus on. So teams that have not as much resource, it just gives them a leg up because they've got that automation and they've got the tool working for them.
And I think the second piece to it is that we haven't really talked about is the blinking light is great, but you need to be able to integrate it in the platform in such a way that it can automatically do something and you have that sort of real time interdiction right there inside the workflow. And so that is also the other piece that I think is very different in terms of the interdiction levels that are now available, whether they're restricting a permission or just completely blocking an action, you can dial those in a pretty sophisticated manner.
So I think it takes some of the work out of somebody's got to go triage this case, they've got to understand what it is, they've got to go make a decision, they've got to go do something and block a transaction. It's like all that needs to happen automatically so that your limited resources in the back office are able to do the most valuable piece of the work, which is figuring out the most important cases that need some sort of intervention.
Adam Blue
Yeah. So maybe at the bank, the bank employee's job moves from being someone who's operating, right? I click here, I read this thing, I release a transaction, I stop a transaction. They stop operating so much and they start orchestrating. I look at the reports at the end of the day, I figure out this agent or this rule or this heuristic or this skill, whatever the manifestation is, I think it's effective, but I'd like to tune the edges, right? I'd like to do this.
Jeff Scott
Yeah, understood.
Adam Blue
So they kind of come up a level in the implementation. Yeah. And so the other thing that always strikes me is working in this space, it's interesting, and I've worked on and off in a lot of parts of the fraud and cybersecurity space in my career. And as you pointed out, having a platform that has the hooks in it to let you do what you need to do, dial up and dial down entitlements and permissions, perform real-time interdiction, that is non-trivial. But even if everybody had that, talk a little bit about the special sauce on the engineering side.
What does somebody who's great at building these kinds of tools look like and what kind of data do they get access to, do they have to have access to in order to approach a segment? Because I think it's really interesting when you look at the depth of skill and talent that's required to work in this space. It's even sometimes a little narrower than being just a great software engineer.
Jeff Scott
Yeah, well said. I think that the secret sauce, at least the really smart team that I have the privilege of working with, is that they really, really understand the pathways that fraud takes. And they've learned that, at least in our case, by sitting down with our fraud operations team that gets real confirmed cases of fraud. And they trained an agentic model on those confirmed cases of fraud and said, "Go figure out what's different about the very slight behavioral nuances of these pathways." And then they're able to tune the model to build detectors against those.
And so they have spent an inordinate amount of time understanding what I would call the policy build, if you will. So a lot of tools in the market that we see or come across on a regular basis, they would go to the financial institution and say, "What are the things that you want to happen and in which sequential order in order for us to create this excessive navigation alert on your behalf?" We did all that work for the financial institution via these engineers. And so they understand all the different various ways that account takeover can happen today, and they're constantly learning new ones based on studying fraud essentially, and taking any confirmed case that we get and making sure that we have a way to build a detector for that specific nuance.
And we've got hundreds of difference of nuances there in terms of what account takeover or other social engineering scams look like. And the thing that they did that was completely different in my opinion from what we've seen in the industry is that they created a way for the model to be able to create those detectors once they came across something new. And that was really the secret sauce, I think, for what they were able to figure out.
Adam Blue
Yeah, that's really interesting. I saw a presentation from a company that's building some AI tools, not for music creation, but for grading music, right? And so you can upload a song that you've recorded and it'll give you a lot of feedback. And it's kind of interesting because I like music. I like to listen to music. I play a little music. And so a lot of times I'll hear a song and I'll be like, "I don't like the way that song sounds." And then I met some people who do that for a living and they're like, "You don't like the way that song sounds because the upper end of that electric guitar is in direct conflict with the lower end of the vocal register of the vocalist. And when the vocal hook and the guitar hook hit at the same time, it gets very muddy and they didn't pan it or put it in two different channels. They just left the mix flat."
And it's like, "Oh, well, now that the benefit of your 20 years of experience has explained that to me, yes, now I understand why I don't like the way that song sounds." It's always interesting to me to go find people who are some of the best at what they do and just ask them, "Why do I like this song and not like that song?" And what you find out is different people, they have radically different visions of what good is. And sometimes I think I work with some very strong, brilliant engineers on your team and lots of other teams, and some of the best of them, I think they just have a different view of what good is. They just have enough understanding of nuance and complexity that their version of good and my version of good, theirs is much more thoughtful and much less instinctual, much less concrete, and their bar is just a lot higher.
And so it's maybe a little off-topic from AI, but it's interesting to watch people that have that taste, that discretion, that understanding, that grace with something. And then you give them AI and it's like, well, that's both impressive and terrifying. And so that intersection is pretty powerful. I think it's great. I wish there were less people that used it for evil and lived I would assume overseas and stole money from people, but I guess it gives other people jobs. So it is kind of what it is. Awesome.
Jeff Scott
Yeah, that was a great analogy, I think, to it. And then the other thing I'd say that's different about the engineers that are working on this is they want to make it as simple as possible for the customer. And so many of our customers are like, "We don't want to have to figure out what the policies are, man. We want you guys to do that."
Adam Blue
Yeah, I can imagine. Yeah, because even that skill, thoughtfully constructing a policy, it's a little like linear algebra when you get those problems in high school and you're like, "Here's three equations that define this plane, but here's one point in the plane and two equations derive the third." And you're like, "Oh God, it's like a puzzle kind of thing."
Jeff Scott
Totally.
Adam Blue
But the people that can do it, they do it very, very well. There's a lot of art in policy construction and trade construction and data discretization I think that gets lost. Those are the pieces that when you combine them with AI end up being really, really interesting, really powerful. All right. Well, thanks very much for being on today, Jeff.
Jeff Scott
Yeah. Thank you.
Adam Blue
This was really fantastic. Appreciate your thoughts and insight on fraud. I think everybody will learn a lot from this. On our way out, I like to highlight a piece of art or culture or something interesting like that. Have you read the book “Flash Boys”? I bet you have.
Jeff Scott
No.
Adam Blue
You don't know that one?
Jeff Scott
It's on my list.
Adam Blue
It's a good one. It's about high-velocity trading on predominantly the New York Stock Exchanges, and it's about one set of traders and software engineers figuring out how to take advantage of the cracks. Nothing illegal that I know of, at least written down in the book. Unethical maybe is another matter, but it's fascinating to watch the give and take and the attack and defense between these firms that are trying to be the first to trade and their use of dark pool trading and fights over getting one meter less fiber between you and a switch so you're faster than your competitor.
And it's starting to feel to me like fraud and security are iterating in that direction. So if you want a little preview of what that probably looks like going forward, kind of that integration of nuance and technology and high skill and high stakes, read “Flash Boys”. It's a fantastic read by a great author, and I think it gives you a sense of where the fraud and security landscape is going.
All right. Well, thanks again, Jeff. I think this was fantastic. We'll close out today's Cut to Context there, and you can find us wherever your fine podcasts are published and sold. So thanks everybody.