Welcome back to the Fraudology podcast.
I am so glad you're here and I'm so glad that Assaf
Kipnis is joining me again on Fraudology.
I had thought that it had only been like maybe a
year, year and a half since Assaf was with us.
It was back in 2023 when I looked up the episode.
Uh it was after he had left his job in Facebook and he talked
a lot about how bad actors and state actors were using social
media platforms for influencing political regimes and their
political influence using social media platforms.
Which was really interesting. It was a two-part episode.
It was really interesting.
I believe it was episodes 195 and 196 from back in May of 2023
if you want to go back into the archives and look for them.
We're now on like episode 430 something.
So um so it was a while ago but it was really fascinating.
Uh and I have kept tabs on Assaf's career
since and am really interested.
He's doing a lot in the world of AI um using AI for
fraud investigations as well as other things.
And he has a lot of thoughts on how companies should be using
AI and what they need to think about when using AI.
And that is why I wanted to bring him on the podcast today.
So Assaf, welcome back to Fraudology.
Thanks so much. Thanks for having me. It's been a long time.
And uh you know so I was just going to have you
start maybe introducing your background.
I know I asked you this question the last time you
were here but it was three and a half years ago.
Um how did you get started in fraud?
What is your, where have you where have you worked and what
have you done at a high level um that brings you to now?
So I've been in the space for about 13 years now.
I started out in security engineering um and then early in
my career got to LinkedIn where I was there for about three
years working on trust and safety investigations.
And when the team was very very small.
We're working on every type of actor we're working on scraping on
uh child safety we were working on financially motivated actors.
And then I moved uh in 2018 I moved to
Facebook where I spent In five years.
I was uh an investigator in the e-crime team.
Then I managed the e-crime team.
Um working on specifically on financially motivated actors, ad fraud,
uh a political influence for profit and a lot of scams.
Uh after that I was at Google for about a year and a
half working on uh scams, how Google products can identify
scams at scale across all of their products.
Uh then I went to ElevenLabs which is an AI startup.
I was there for a short amount of time
running their uh investigations team.
Uh and then over the last year I got much deeper into
consulting and found out that I really liked it.
Uh I started my own consultancy uh ASK consulting.
Where I do both investigations and AI builds and advise companies
on how to merge AI into their processes, how to better do
investigations, how to superpower their uh analysts uh with AI.
Um I will say working at ElevenLabs was a very big eye
opener for me because they're an AI first company.
So, I've I've dove deep into uh building with
AI and that's 90% of what I do now.
You've you've gotten to be on the front, you know, when you've
been on the front lines, you've gotten to do some really
interesting things and see the scams that are new, right?
Well, there really are no new scams,
but there's new usages of scams.
There's new ways that scams are being, you know, um, used, right?
And there's new technology to superpower
those scams and industrialize them.
But you've been really on the front
lines of seeing the newest thing.
Because you've worked for some of the
biggest companies in Silicon Valley.
Um, you've had to be, you know, on your toes, so to speak, to
find new ways to identify them as fast as they're creating them.
Which makes yeah makes you a perfect candidate for a consultant.
Because you've you've seen the hardest stuff and
the worst stuff and you've tackled it right.
So to now be able to help other companies, especially you
know small to mid companies, some large companies.
Um, you know, be able to identify problems
Um, I think that puts you in a really good position.
And one of the biggest things I learned
from these larger companies is scale.
Is what works at scale, what doesn't work at scale.
Beyond being um these companies are always the first
guinea pig for any threat actor to try things.
So yes, you get to know the new things.
Um but you also learn you get to see what breaks at scale
or what works at scale or if you do one thing how does it
the cascade um a year later into something that breaks. So
That is a huge thing about working with these companies.
That's a really good point.
Yeah, you get to see what solutions work and
which solutions break, you know, at scale.
That makes perfect sense.
Yeah, it's it it's been very interesting.
And a lot of times you after you've been in these places, you get
to be the you get to be kind of the person who says, "Ah, yeah.
" You can do that, but go see what it will affect later. Just
It's going to blow up somewhere else that
you're not going to be touching. And we all we all been there.
Because fraudsters and scammers, they're
constantly looking for the gap, right?
They're constantly looking for the vulnerability.
And so if what you build at scale has a gap or
a vulnerability, you know, hey, they're going
to break this once they figure it out.
They're going to find that gap.
So, why don't we fix that gap now?
And that's not always what, you know, what they want to hear,
what the what the big companies want to hear, right?
They want a lot of times they just want to fix the problem right
now, not fix the problem that's going to happen in six months.
It's I want to fix the problem that's right
in front of me and I can measure today.
I think measure is the big one.
That's a challenge for people in fraud.
Because you know sure we can we can stop the challenge
today but we want to stop the challenge tomorrow
and the next day too at the same time.
And leadership often especially I would imagine at some
of the bigger companies, it's all about standing out.
And so they want to be able to measure it so they can
put it on their resume so they can get a promotion.
It it it's it's somewhat of a perverse incentive
that some people don't like to hear, but
Yeah. No, I can see that.
Um, so how does AI fit into your work that
you're doing as a consultant now?
So I'm just building things constantly.
I think usually I have six terminals open building all kinds
of stuff together cuz that's how my brain thinks.
I got to wait on one. So let's start five more. Um,
I think 80% of us have ADHD in fraud.
And I'd be willing to stand by that percentage.
And, um, so one of the things that I enjoy the
most that I built is an investigative system.
So when I left the big companies, it was I it was
on my mind when I was at the big companies.
I said, "How would I ever do an investigation again?
" I I sit here sitting on all the data ever.
And if I don't have this data again, how do I find anything
like how do I do open source investigations?
And open source investigations for people who
are listening who do them, they are lengthy.
They take a really long time. You got to document everything.
So I started building myself a tool of
okay, how do I do an investigation?
It started with stuff that I did in the past about how do I find from
an ad what are the other ads that are running the same fraud?
Who are these actors? Where are they living?
How do I find uh affiliate marketing? So you can do that.
It just takes a really long time of opening a lot of websites.
So uh and finding patterns.
So I started building a system that I provide a seed and then
it extrapolates everything and continuously comes back to me
and says okay I found this which direction are we going?
And or I say okay we found this is interesting, but if I
was in the weeds I would cluster here and cluster there
and then find the similarities between the clusters.
So AI, the its best uh feature that it it finds commonalities.
Just help me find patterns.
So um I've been using AI for a lot of automation of processes
that are tedious and take up a long time.
Uh and then into finding patterns in um in abusive behavior or
patterns in data, that will give me um what I probably would have missed.
Or I remember looking at graphs that were 5,000 nodes.
I can't look at a graph that's I can't make sense person.
But AI can make sense of the of the graph, give me something.
And then I can I can evaluate the assumption that AI makes.
Or I can say no throw it back try again.
Or I can say oh okay this actually makes sense with what I found.
And I built a bunch of other tools that kind of make my life
easier, but the investigative tooling has been a big one.
And the other one is for my own work and my own consultancy
I struggle, And I'm sure a lot of people are
you like I will struggle managing my work.
Uh I remember asking people how do you record everything
that you're doing and do you use a spreadsheet?
So it took a year, but I built a system for myself that essentially
I just get up in the morning and say what do I have to do today.
And, or what did that person say to me.
And that that, I go to that client.
I say what did that person say to me yesterday, and
what was our conversation about that we recorded,
and what is the actual work I need to do here.
So that has changed the way I work because I don't have to well I
don't take notes anymore which is really nice, but I don't have
to go have the mental load of everything spinning in my head.
And then that's extremely hard when you're working on multiple
projects, which was always really hard for me.
Uh, but now I can work on six projects at the same time.
And I'll just stop this one.
I'll say, Okay, where did I stop on this one?
What did these people say?
What were the emails that came in since
the time I last spoke to them?
And what changed and how did they uh like what I wrote?
And then continue from there.
So those are the main things uh the investigation stuff and
the knowledge base uh has been extremely helpful for me.
And and clients that I've been working with.
So, talking about doing investigations and all of that, what are
into scams or ad fraud or whatever you're working on, what are
some of the trends that you're seeing on the bad actor side?
Uh, I mean, I just did a presentation at MFA that I then
redid for the podcast two weeks ago or 3 weeks ago now.
That where we talked about, or I talked about, how
bad actors are using AI to industrialize fraud.
And they're they're scaling. They're scaling at at a rapid rate.
Um, a lot of, you know, every measurement that we have
of specific types of payment fraud, especially.
Because payment fraud is easier to measure
than some of the other fraud.
Um because of chargebacks, but every measurement that
we have, you know, around like card testing, well,
that's up over a 100% in the last few months.
And ATOs is up over 78% in the last like quarter,
because it's much easier easier to do with AI.
Um, are there other things that you're seeing from the
bad actor side that that are changing because they
now have access to these superhuman computers?
Or these supercomputers really.
I think the industrialization has been going on since 2019 2020.
The fraud as a service and all of that.
And if you asked me a year ago, I would say eh,
I don't think they're using AI that much. Um,
Yeah, even six months ago, I probably wouldn't have said that.
And now I'm it's AI is this is the interesting part.
I think AI is used for scale versus uh changing anything.
So you have you have the ability.
So the same way that I have the ability to now create tools
that I want to play around with, anybody else can do that.
So you can just build thing in a in a in a much faster way.
And you can innovate on the speed versus innovate on the tactic.
So in a in places where you don't have the capacity
or the capability to get to a very large amount
of people, then you've got to be real crafty.
Um because there's only x amount of stuff that I can do.
But now I can have AI write the code for me to do stuff a lot faster.
And that also superpowered the the fraud economy.
Because now everybody can make tools.
So now you can sell the tools because in the end
not everybody is going to be working with AI.
Uh I feel that uh on the defender, it's really it's it's a really
weird dichotomy that on the defender side you you hear people
saying oh thread actors are so good at AI and they're amazing
at it and they're doing all these crazy hacky things.
And you ask so how is AI on the defender side? I said backer side.
How do we Oh, it's awful and we can't use it and we hate it.
Attackers are people like you and they might hate it, too.
I think the the the bigger difference here is that
they the attacker has a lot more freedom to kind
of play around and try, uh and innovate.
While defenders have to continue doing their
job while they're trying to innovate.
But I haven't seen something that's like
or I'm seeing chat bots running a mock.
I I just uh I think people using chatbots a little
bit more for fraudulent transactions.
But then also the defender side that doesn't really know
how to digest chat bots yet, or virtual assistants.
But it, I am seeing the attacks just being more.
Before if they were they were, let's say in order to
get a fake account you would have to reach out to a a
fake account farm in Bangladesh and they would create
accounts for you and they would sell it to you.
You can now do it yourself.
You can go buy the tool or make the
tool and kind of make it yourself. And play around with it.
So I think it's just um the velocity of the attacks became faster.
Because it's easier and it uh it kind of promotes this community.
That you can just sell new tools, that before had to be expert.
That you had to have an expertise in it to make it.
And oftentimes there were multiple people
involved in one attack, right?
Because you needed multiple people's expertise
to complete the cycle of one attack. I just hit my microphone.
Uh the the thing is the expertise still sits there.
You still need someone who knows what they're doing.
Yes. But not maybe not as many, right?
You don't need like one expert in this and one expert
in that and one expert in that as much.
The thing is you can proliferate it.
You can build a tool that has the understanding.
You have it has the knowledge and I'll just use it.
Um and that's, this is where I see also a gap in defending.
Is the the before we always had every attack
always came from some sort of playbook.
Always. And nobody got that then.
But I don't know if anybody gets it now, but now
the playbooks are are a lot more robust.
They're a lot more um available.
Because you can replicate you take something that someone made and
replicate it for yourself and just move on or sell it yourself.
Yeah, those are all good points and
things that I'm seeing as well.
And then, you know, you started a little bit to talk about
the trends and themes that defenders are seeing.
You know, um, on the good side of fraud.
And I'm seeing several, seeing and hearing about,
you know, several different themes about AI.
That, um, I mentioned to you a little bit before
we recorded, but I'll mention a few now.
You know there were companies six, seven, eight months
ago that were giving directives to every department.
Saying okay by this date you have to be using AI for up to
25% of your or for 25% of your job for your entire team.
Or you were tasked with reducing human hours.
Um how whatever that metric was they
were like doing these guidelines.
And then you know the problem was is that nobody
knew what to use AI for at first, right? It took a little while.
Uh and then now there are some companies that are using
it for things like, you know, chargeback responses.
Which are still pretty manual.
But that's another, uh they're not able to get the data from
one place to the next to autopop populate everything yet.
Um and so there's still some manual work in that.
But um they're also using it for you know
uh data extraction and running reports.
Or um you know different different use cases.
And um what I've noticed happening is the hut.
Now that these companies that had this very aggressive directive
are realizing, oh we told them to use it and use it as
much as possible to reduce, you know, the amount of
human, you know, hours that we needed to pay out.
But now we have to pay the tokens.
Uh so now they're saying, "Wait, wait, wait.
Now we need you to pull it back.
" And I think that's that's one thing I'm seeing.
Um do you have thoughts on that one?
Oh yes. Um so I want to start from where we started.
And I think the problem is a mis is a is a fundamental
misunderstanding from leadership.
And sometimes it's also fraud leadership or abuse leadership.
That's abuse fighting is blocking. And I will die on this hill.
The thing so I've been thinking about this and I think things are
just like as we're talking kind of falling together into place for me.
One of the things that I really hate
about abuse fighting is blocking.
Uh blocking anything, blocking content, disabling
accounts, uh uh blocking websites. I I find it useless.
Uh and perform performative.
So the thing is if you're as a and I'm I have a
feeling I'm going to make some people unhappy.
But if if you're as as a team uh a profession and not necessarily
a profession but if if your job as a team if how you
see your job as a team is blocking things, then your
leadership thinks your job is blocking things.
And blocking things is easy. It's straightforward. It's menial.
And when when you talk to leadership when you talk to leadership
and you'll say what can you do with AI the first
answer is always how can I remove menial work.
And abuse fighting is blocking I'll just reduce that.
And then the the it hits that that's not what that is.
Blocking is a tool that you can use when you when you're fighting
abuse it's not a very effective tool but it's tool you can measure best.
So, to a degree, it's it's it's self-inflicted.
We've spent so much time measuring how many accounts
we've taken down and how much content we remove,
And then to an outside reviewer, it seems menial.
You're just blocking stuff. Cool.
Or I've heard that you're the blocking team, which makes me ill.
And then sure, I'll bring AI and it will block stuff.
But then you start working at it and that's not the job.
It's part of the job. But um blocking needs finding. How do you find?
Yes. How do you identify?
Blocking needs understanding of patterns.
And you say yes AI can find patterns.
AI doesn't do doesn't do anything by itself.
If you don't know to tell it to look for the patterns.
I I for example I can give my investigative
tool to my daughter who's 13.
She'll find some stuff but it won't be very effective.
She'll find stuff you can block.
Which is great cuz I built her the thing.
But she won't be able to find like what's actually going on.
How do we actually go to the to the root of this and
finding and finding and making it difficult to act?
Uh and that's what happens.
You have either abuse teams that are now less people
who are doing the same blocking job they were
doing before just people are overworked.
Or teams that have completely been gutted and all
the things that people know how to do is gone.
Because the the the leadership is like oh you just block stuff.
AI block stuff. Great. Find the stuff, block the stuff.
Do uh image image searches, which is also useless.
Or uh word matching and block it or make do a classifier.
And to me, that's that's the real crime here.
Like you're losing all the actual knowledge here.
And you're not going to get it back.
I get what you're saying.
That it is easy to slip into the trap of
thinking that our job is blocking.
And sometimes I slip into that trap too.
I think our job is identifying for sure. And finding.
Ideally think that it should be stopping
the root cause of the issue, right?
So finding how they're getting in finding the vulnerability
fixing that vulnerability and changing changing that.
And unfortunately a lot of times that takes
resource extra resources than you have.
A lot of times online companies aren't looking at those signals.
They aren't or they they aren't ingesting those
signals the signals that they need to identify
the bad actors earlier in the stream.
And that's also something I talked a lot about in my presentation
was how fraud is no longer happening just at the
point of transaction or the point of login.
It's happening throughout the process.
And if you're able to identify it from the moment they step
on your platform, or your website, all the way through, and
you're able to ingest all of those signals all the way through,
you can have a better picture of what they're doing.
And how you can stop them.
Um, but not every company is there yet.
Not every company has those signals or
has the ability to look upstream.
And so that does turn them more in the default is blocking.
Um the default is well we'll just keep playing whack-a-mole
and stopping them where we can find them.
And that might mean stopping them at account creation.
And then we measure how many accounts we you know we closed.
Um and it does come down to measuring.
Because you know it's people's jobs right?
They want to be able to tell their leadership they did something.
And if you know they still can if they're
do if they're finding the root cause.
And changing the you know what they can ingest
and how they can identify them earlier.
They can say hey we've had you know x% less account
login, or you know account creation fraud,
or account takeover fraud, or whatever.
But it's just a different way of looking at it.
Is is that basically what you're saying, as far as
what should be done instead of the blocking?
No, what you're saying is is kind of how I'm thinking about it.
Sometimes you have to block stuff. I get that.
Um that that also erodes of the on the threat that erodes the threat
actor's abilities and makes it annoying to be on your platform.
And I think that's the main goal.
Your main goal is to make it annoying to
be on their platform and expensive.
And and when I have when I see companies saying, "Well, we don't
need a fraud team anymore because we're just going to put AI.
" I think it does two things.
One, it just does not fight fraud at all.
It's just it's like I'm not really sure how you're
bringing in something that's doesn't understand your
company, doesn't understand your customer.
And even if you teach you the processes,
first of all, who's doing that.
You're still going to struggle with having it go on its own.
Uh and and and do things and make decisions that are are the correct
decisions for your use case or correct decisions for your company.
On the other hand of that, it also causes
people not to want to use AI at all. I've spoken to people.
I said, "I'm building investigative tools with AI.
" And the first answer was, "Well, it's not going to take my job.
So, the I can't blame those people because that's what you hear.
It's We just want to reduce headcount.
And I still hear people uh in startups for AI startups and companies
that say we're going to use this to reduce headcount.
Because that's the easiest way to get
into a CFO's heart saying I'm here. I'm going to reduce headcount.
The way I see it and at least the way
I use it for me, it's a superpower.
I can do things that I couldn't do before.
And yes, I can do four, five, six investigations at a time.
Are you telling me that you only have four
to five to six investigations at a time?
As as as a company that fights fraud? No. You have six analysts.
If each of them can do two to six investigations
at a time, they'll never run out of work.
Because, and and I think that comes with another perception that we
talked about earlier, that some people who are less who are some
in leadership that are less understanding of what fraud means.
It's like oh we took down this cluster of accounts.
So it's done, right? It's never going to come back. We won.
And if that's how you think about it, then
we don't really need a police force.
Because we took down the guy standing on the
street or girl standing on the street.
Selling drugs. We're good. We won, right?
We don't have a drug problem anymore.
We took down the drug dealer. Yeah. That's a good point.
The the perception is is is wrong because you're you're just looking
at it as and I still see that and it still boggles my mind.
People still think it's that one to five dudes sitting somewhere.
Running scams like that hasn't been the
case since the last decade at least.
No, we're talking about scams compounds, which actually
my last guest on the podcast escaped a scam compound
in Laos, which was very interesting.
Um and crazy. Um is a crazy story.
But um you know, we've got, you know, criminal organizations.
We've got state actors that are funding their com their country's
military abilities based on, you know, transaction fraud and
and other types of fraud and banking and and other places.
I mean, we it's industrialized.
It's no longer a few guys sitting in a basement at
their mom's house like you know people picture.
The perception for some reason is still that.
And oh I've I heard people oh I blocked this person's um device.
He'll be back with a proxy device. Yeah,
And that's another thing that AI AI made things even cheaper.
They were cheap before. They're cheaper now.
A lot cheaper. And that allows them to scale.
Yeah. The, and the thing is it allows the threat actors to scale.
Right? You said that. What are companies doing?
Replacing instead of scaling.
So scale your people. Make them do the job faster, better.
Uh make them not do the stuff that they're, for example,
the thing I hate doing most is writing the report.
I would spend days writing the report.
I now spend 10 seconds writing. I did the investigation.
And now I hate that part. Write the report for me.
I'll read it to make sure that it's what I
said, which my process does that anyway. And off we go.
Um, it releases me as an investigator from a lot
of the things that take me a really long time.
And allow me, not just by being able to automate things, it
allows me to just not do things that I don't want to do.
That would take me so long because they just take me a long time.
Yeah. Um, you reminded me a minute ago of this analogy.
That I feel like everyone in fraud has heard at one point now.
Because I used to present about it at like so many conferences.
Um but it was the fact that so many people who fight
fraud, or their leaders, uh you know their leaders
think of fraud as fighting a dragon. Like it's one single attack.
And you use all the same tools that you used
for the last dragon that you fought.
And it's pretty predictable how you're going to
slay that dragon, because you did it before.
And you, once you slay that dragon, you can go
back to your village and do another job.
But what it really is is it's like fighting zombies.
That continually are, you know, reiterating
And you kind of have to just grab whatever tool you can.
Or you have to, you know, or maybe you take a minute to be
strategic and say, Hey, we need to have, you know, specific
tools to find these zombies sooner or whatever it is.
But you never have time to go back to
your village and be celebrated. You never have time to.
Because they're always gonna come back and they're usually
going to come back bigger and stronger and faster.
And they've adapted to the defenses that you put in place.
So, you're never going to be done.
Yes. And that's the crux of being a defender.
I remember being in the military border defense.
If you are if your country is has a hostile
border, you never leave the border.
That's a very good point.
You're always there. And you have the much much harder job.
Because you have to figure out is that a bad guy testing
me or is this just a dude walking around?
As uh as a defender, you have patterns.
As an attacker because you have to you have
to systematize what you're doing. It has to you have to do it.
And that creates patterns.
And the attacker just needs to sit there and
view your patterns and figure it out.
It's kind of it's like it's like escaping.
I'll make it less dramatic like this. Maybe like escaping a prison.
You sit and the light goes like this and like this.
Okay, I can time it. I think that speaks to people more.
And the prison guard always has the hardest job.
They have to look everywhere.
Uh and the the person escaping just needs to find oh, I found that.
Um, but think about it at massive scale.
That you have everybody trying to figuring you to figure you out.
But something you said really uh got me is the slaying the dragons or
or getting the zombies and then every time and this is how I work.
Every time I fight something or I find something
Or I fail, I say, "Ooh, new tool. Toolbox. "
Now, next time I see something that looks
like that, I'll say, "Ah, right.
" So, the thing is what I'm saying right
now works for me as a one person. It doesn't work at scale. Right. It
You sit at a company and the way I've that happened to me in the
past is I would say, Ooh, this username is really annoying.
And then from the other side of the room someone will say, that guy?
I worked on that 5 years ago. Let me tell you.
And and that's the thing that you're losing if you are
allowing people to leave or having people leave.
And the funny part and, or replacing them with AI, and
the funny part that's the best thing about AI.
You can create a living knowledge base, that someone fights
something, they figure out a tactic, you have it in the knowledge base.
And then when someone else like sees the exact same thing,
the knowledgeist will say, this person found that.
This is exactly what they did.
Now we can learn from that.
The challenge with that that, and this is a challenge that a
fraud leader for a large company told me, was that they're
having a really hard time with their team members, their
employees being willing to share their knowledge with AI.
And I don't blame those people.
Because they're worried, and maybe rightfully so. I don't know.
But they're worried that if they give up that domain expertise,
those years of everything that they know, right, and how well
they know the systems and how they talk to each other and and
how to fix the problem and oh yeah, I remember this guy in that
department can fix that for me or you know, just all the things
they know, because they've worked there for so long.
There's they're not going to be able to transfer all the
things, but they're worried that they're going to transfer
enough, that the company is going to look at it and say,
"Okay, we've now basically downloaded your brain.
And we don't need you anymore, because
it's now in the knowledge base.
" And that's a struggle that employers are having.
And I and then a struggle that employees are having.
Like I don't want to tell my boss no, but at the
same time I don't want to dump my brain into this
thing that's going to remember it forever.
And I'm not going to be as useful anymore.
So I want to challenge that.
And that this is not a new thing.
How many, you've worked at many places.
How many co-workers did you have that were that coworker
that was did not want to share with anybody.
Because they were they were like I am the only one
who knows this thing and I'm not going to tell
A few of those. Yeah. Mhm.
That is a cultural problem.
If you if you convince your employees that they're only
as good as what knowledge they have now, and if they don't
have it anymore, or they're not special anymore, you'll
get rid of them, and is that's how they'll think.
So they thought that way before and they'll think that way now.
But now there's a new monster, a new a new
thing that will steal your knowledge. So that's the first one.
So I don't think that's a new issue.
That's that's true. Yeah.
They wouldn't want to give it to other employees.
They wanted to be the one that knew the thing, right?
I'm the one that knows that geography or I'm the
one that knows that system or I'm the one that
knows, you know, ATOs or whatever it is.
That fraud attack, you know,
But that also downplays your own expertise.
That you you're downplaying your you're saying the AI brain
is better than my brain and I can't compete, which yeah, it's
what it's what uh frontier companies want you to think.
That AI knows everything about everything
all the time, which is just un untrue.
Companies should, companies hired you for your judgment.
Not your, sometimes for your knowledge but mostly for your judgment.
And your judgment is going to continue with you.
Your judgment is going to be the part that
tells you how to go find new attacks.
Because on the other side there that the the attackers
are not saying let's just AI do everything and
we're just going to sit here and make money.
They're not able to do that. They need to continuously iterate.
So you need brain on this side.
It's like the silliest thing is let's fight AI with AI.
Which makes no sense. Because what does that even mean?
But let's fight innovation with innovation
and innovation comes from people.
The thing I've noticed about so it might change in the future
but me my working with AI is I noticed one thing.
People think of AI the same way of AI thinking, as the
same way that they think of a person thinking.
So if I ran out this room and the door was closed
and I smashed into the door, I would fall.
Trust me, I will open the door every single time. I will watch.
AI will smash into that door every time. Or 70% of the time.
Because maybe the knowledge transferred, maybe
it didn't transfer, maybe it transferred
Maybe it didn't understand the context.
Or it didn't you know it didn't understand the context
or the you know contextual like exactly how you
know how things go together with each other
And you can teach you can you can explain
context in a way that AI can consume it.
But it's not the same way you perceive context.
Context for AI is just flat data.
For you, you start saying oh I did this, and then
I did this, and then I fell on my face.
I'm not going to do that and that. Yeah.
And and that happens to, when you would use AI, you're like, Hold on.
We just did this yesterday and it blew up
in my face and it's like I I don't know. That's what I know. I It's
It's surprise. It's funny.
On one hand, we get surprised when AI does things
that you're like, why did you do that? That makes no sense.
And then we give it this like it can do anything.
Well, and you know the other thing is that I've learned
is that you know AI never says I don't know.
Or and it never says I'm not sure, right? It
It will just give you an answer.
I mean uh Frank McKenna and I were playing around
with Google AI like a few uh months ago. I don't know.
Time is a construct I feel like but.
It it was yeah probably a couple months ago.
And uh we were just saying like, what is pizza fraud?
Or what is like, we were just making up titles of fraud tactics.
And then it would make up a totally weird, almost nonsensical
definition, of this type of fraud attack that we made up.
Um you know, so it's not going to say that's not a thing.
Or I don't know, or I'm not sure that that's a thing.
There's, so the way I think about that is again
the way people are told how to use AI.
You're told to open the chatbot and ask questions
So the, if you think about it, the way Google does it.
Google doesn't just give, usually, doesn't just
give you answers give you here links.
And maybe the links will will derive something that it will tell you.
Um when you ask a chat a question about pizza fraud,
it only has the context that that that it probably,
the probability is available to it right now,
And it's let me build this together. I never do that.
I never ask my Claude tell me about that. I say go research that.
And research me, find articles specific to this field by experts.
Go find all that stuff. And to me it will say I found nothing.
I don't see anything of what you're talking about.
And and that's where, and I think it's just a matter
of having an understanding of how to use AI.
And it's just in the business space and specifically in fraud and
safety, It's not the same way that you would use it as a consumer.
Well, consumers probably use less, but it needs to
be a lot more thoughtful when you're using it.
Mhm. That that does make sense.
And yeah, we weren't we weren't giving it any context to be fair.
But um that's a good The other concern I have as far
as asking Claude or any of the other tools to research
fraud topics, is that most of the fraud knowledge
within our industry lives within our brains.
That's why conferences are still so important, right?
Like that you can talk to each other one on one.
Because a lot of us are, you know,
a little cynical and skeptical.
And we're not going to write it in an email and say, "Oh yeah,
that's what happened to me when I worked at Facebook.
Let me tell you all about it.
" All all that we did, and all that we saw.
Like no, you're going to say, you know, "Hey, let's
hop on a phone call," or something like that.
Or like let's get together.
Because you don't, you know it just, or maybe you aren't
even talking about a specific company, right?
It just, we don't feel comfortable doing that.
So, a lot of what's out there in the in the OSINT world,
as far as fraud, you know, trends and fraud tactics
and things like that, are written by vendors.
And, you know, a lot of them are blog articles that might
might turn into a sales pitch at the end, right?
Like, here's the problem and the only
thing that can solve it is my tool.
Um, and so I I worry about that too, is like if if we're
relying on open information about fraud, like if I
say, "Hey, research everything you know about account
takeover", you know, account takeover fraud.
It's mostly going to find not expert, you
know, research or anything like that.
It's going to find blog articles from 17 different
fraud companies that were hoping that you
know someone would click on it for SEO.
Like you know, so that I think
Our industry is different than other industries,
like marketing for example, where you know there's
tons of open-source articles and stuff.
Because the other reason we're not going to say it
publicly, is because we don't want the bad guys
to have any advantage that they might have.
Like we talk about fraud all the time on fraudology.
Yes, this is going to turn into open source information.
But you and I are not talking about specifics.
We're not talking about specifically how we identify
this type of fraud or what's being used. So, it's okay, right?
So, but it's not going to be used for, you know, you're not
going to be able to be a company who fires their fraud manager,
and then just ask Claude what what they would do.
And there are companies that are considering
And said, "We'll just ask Claude when we have a fraud question.
" And I'm like, no, no, no. You're going to get
Bogus blog articles that don't have any real answers.
You're also going to get um decisions,
or directions, that you cannot test.
An example is, so the whole issue that I have with working with
Claude or with AI is that it's not hallucinations hallucinations
it's decisions that it made with whatever context that it had.
And then it just made that decision and went that direction.
And part of my work, and the main thing
of my work, is saying no no no. Why did you make that decision?
Why'd you go there? We're going that way.
Because, and it will confidently make a decision, and
it's making a decision on the context that it has.
And maybe I didn't give it the all of the context, and I,
does it doesn't have the context that I have in my head.
It will say, "Oh, I finished the investigation.
I found this and this and this and this.
" And then I sit there and I say, "Okay, what about the
connection between here and here and here and here?
" And of course, it will say that that is a great idea.
But the way I think about it is it would not make the same
determinations that I make at least 50% of the time.
I, but the the thing is is that you can't trust it to be consistent.
Uh and people really like to harp on on um hallucinations.
Hallucinations again are decisions that went awry.
So I want you to consistently make decisions like I would, right?
Because it doesn't have all your context and also another
person's not going to make another the same decision
like you because you think differently.
Um, and it's not going to ask itself
questions and say, "Oh, am I wrong? "
Right? That's a good point.
So my question that I had written out was how can you trust
that the output of working with AI is accurate?
So if you're only looking at the output from AI, you can't.
You can't trust it to be accurate.
What you can do is bound it with code.
It's saying if this happens, um so first of all I'll go back.
It's it's nondeterministic.
So if you have something that you want to know
if the answer is yes or no, don't ask AI. There's code for that.
Code does that. Scripting does that really well. And you know what?
AI will write the code for you.
Um but it's a lot of times it's also identifying
where are the limits for AI.
So an example would be we are working a lot with spreadsheets, right?
It's the space works a lot with spreadsheets.
If you put a spreadsheets in, spread a lot of spreadsheets with a lot
of information into AI, it's very possible that it will make some
decisions there that you can't see that will corrupt your data.
So don't have the AI read the spreadsheet.
Have code parse it into documentation that then is easier
for the AI to parse and then make decisions.
Um, and within the, generally within the the Claude codes
harness, I have a lot of things that if Claude tries to
do that, uh, the the code says no, you can't do that.
I'm not even call my harness doesn't even call the the LLM at all.
Or it says you can only do this.
Or if the AI tries to do something, it blocks it. It says nope.
If it tries to delete something, it just blocks it.
Cuz you cannot do this action.
So I think this is why I say I think about it like a thread actor.
Because I it needs to have parameters of what it can it
can't do and they need to consistently be enforced
And not enforced with skills or you telling it not to do it.
People, oh but I made a never hallucinate or always do that.
Yeah, I learned pretty quickly that doesn't work.
Like you, it will make a a determination to, oh, I needed to
do this faster, so I decided to not read your whole skill.
Okay, that's where code comes in.
Okay, if I really need you to not to do something specific,
you can write code to say just do that.
And this is where it gets more a lot more
complex than talking to a chatbot. But this is where chat bots fail.
You cannot have run a business with a chatbot.
Huh, that's a very good quote.
I think that should be like your professional coat.
You you can quote you cannot, you know, maybe it's on your
tombstone or it's, you know, your your whatever your quote,
but you cannot run a business with a chatbot.
I think that's very good.
Oh, you know, one other thing I wanted to just talk about really
fast was, um hopefully we can talk about this fast.
So, that was in the AI boot camp that
we had for Merchant Fraud Alliance.
That was something that came up quite a bit.
Was okay, now that we're using AI and everybody's
using it in a little bit of a different way, and
you know different flavor and and all of that.
Like you know, it's now what do we do for AI governance using AI?
Um and what advice would you have for them?
I think it's the same advice.
Is have guard rails and guidelines that are enforced
in code about what you can and you can't do.
Um, and I think this is what happens, right?
You bring in AI, you tell people use all
your credit, they build a lot of stuff.
And this is an a software engineer's worst nightmare, right?
Because nobody looked at the code, whether it's good or bad.
Nobody looked at the code.
Nobody saw how this engages with production.
Even with policy, Uh how does it how does it
when it meets real life, how does it work?
You have to have the same thing that you have for people.
And you have to decisions can only be
applied within specific parameters.
And that ends up being again uh the guardrails have to be in code.
You say to the to the to the LLM, you
cannot do that or you must do this.
Um, a lot of times, like I said, people try to do
that with skills, try to do that with pros. And that doesn't work.
And that's a general misunderstanding of how LLM's work.
Because they're just going to try to get it's just I'm trying
to get from point A to point B the fastest I can.
Uh but if I say I build a wall, and you you you I just and
not allowing you to pass this wall, go the other way.
Um or, and you and you, the thing is governance takes a really long
time because you have to figure out, because this is not a person.
This is some this is an adversary that's
going to try to get around you. So you have to say I built the wall.
Oh crap. It went under the wall. All right.
So now I got to go deeper.
Or now do I have to go 10 steps back and
figure out like what happened here?
Um so starting AI governance after having like full infra
that's AI dependent is going to be very very hard.
It's possible it's just a lot of work.
I was wondering as they were sharing that frustration
and that concern of like, uh now we have to figure
out the rules, around the guidelines, right, around
what we're doing and how we're using it.
I was thinking, is that backwards?
But we do everything backwards. We do security and fraud backwards.
Everything is like, because you get the shinies,
you see something cool, you build, I'm guilty.
And then you're like crap, I have three AI agents
that are writing to the same documents.
Or I have three analysts using it in
three completely different ways.
And we need to have some guidelines around like how, you
know the merchant that had used AI the most within this
boot camp was sharing like they've had to have a lot
of conversations around how do you define this.
How do you identify this.
Because she realized oh I have I don't
know 27 agents, or whatever it is.
And and they all find account takeover differently, like with
different SQL queries, or with different, you know things.
How do we tell, you know, Claude how to identify ATO
when 27 people are finding it in 27 different ways?
And how do you explain it to legal?
The this one file the right ATO's and this is where someone
said to me, I told you earlier, someone said to me
Well it doesn't stand up in court.
But yeah, if I can find if I do discovery and I find that you
have 27 ways to find ATO, tell me which one is right?
Especially if it the biggest problem is if you treat, If
you allow AI to be a black box and in governance and
safety and security and fraud, you have a problem.
You can't have black boxes. You don't know what's going on.
Especially if you're going to legal and saying,
"I know that this is a bad actor. Sends the send the FBI there. "
You can do that, but only if you put the appropriate
guard rails and processes inside.
But I think this is a problem that everybody is going to have.
Most people are going to have.
Like it was really cool and then it got scaled and then
you're like, "Oh no, now we have to build governance. "
Yeah, that's true. Well, this has been a really fun conversation.
We're just at the top of the hour.
I had a feeling that we would have good banter on this topic.
Um, I've definitely learned a lot about, you know, how to think
about AI in a different way uh after talking to you today.
And I think a lot of other people have too.
I'll definitely include a link to your uh LinkedIn
as well as to your consulting page.
I think there's a few companies I thought of that I wrote
down as we were talking that I was thinking, you know,
it would be really good and helpful for them and I think
they'd be really open to having us off come into the
their company and just do like a workshop, right?
Like a thinking out loud workshop of like, hey, this is how we're
thinking about using AI or this is how we're using it now.
Tell us where, you know, what we're doing wrong or how
we need to think about it or how we can put some code
around it to make sure that it's, you know, going
to do the right thing and have the right output.
Um, I think that would be a really good, I think
a lot of companies would benefit from that.
Right now they're they're not really, you know, they don't really
have anyone to go to to be like, hey, am I doing this right?
Or am I, you know, what am I not thinking of?
But because you've had to do it already for, you
know, two very large companies, you've been
through those processes that they haven't.
So, I think that could be a really good service to offer.
Um, I'm just, you know, adding to your your consulting offerings,
but I think that that would be really beneficial. Um, yeah.
Is there anything else that you want to mention
before we uh call it an episode?
Um, always go look for the root cause of whatever failed.
Don't let don't say fix it and move on. Say why did this happen?
I think that is the core of how I work with AI.
I always say, "Oh, you failed here. I don't want to hear fix it.
Tell me why. Do a root cause analysis. Tell me why. How can we learn?
" And I think that's the way people
should look at at working with AI.
Anytime every time it fails at the opportunity to learn
That's true in life too, right?
Every time we fail, it's an opportunity to learn.
But that's good to transfer it to AI as well.
Uh because they're not, it's as you said
it's not necessarily learning, right?
Like it's not necessarily going to say, "Oh,
I walked into that door yesterday. I'm not going to do it again.
" It's going to follow the path it's
been given every single time.
Um and so if if we are learning from it and then
helping it not make that mistake tomorrow,
that's only going to make things better.
Yeah. It doesn't feel the pain that you feel is not.
And it's not attached to a paycheck or
all the things that a paycheck means. It doesn't have a family.
It doesn't have, you know, the meaning.
I mean, it it's getting paid by tokens
whether they're right or wrong. So, it doesn't really matter.
Well, thank you again so much for such a great conversation.
I um hopefully it won't be three years
again for the next time you come on.
All right. Thanks so much.