Stopping Hiring and Recruiting Fraud

Transcript

There may be transcription errors: we apologize for those in advance.

Jason Zoltak is the CEO and co-founder of Tofu, a company using AI and machine learning to tackle fraud and security problems in recruiting and hiring.

Rob: Hey, thanks for joining us. Welcome to what I call Won't Fix.

If any of you have worked in software, you may be familiar with "Won't Fix" as a bug status. It basically says: there's a problem here, but we've decided it's too messy, too complicated, too difficult to fix, so we're just going to leave it as is.

I like that as a framing because it summarizes a lot of the problems I've worked on in my career around trust, security, and safety. Sometimes there's a real complication — it's difficult, there are trade-offs. But other times, the "it's complicated" excuse just isn't a good reason not to work on something or fix it.

So that's the framing for the conversations I want to have here. I want to talk mainly with practitioners, but also with other people involved in these issues: executives, victims of scams and trust problems, journalists who've covered these things in a meaningful way. And by the way, there's a lot of non-meaningful coverage of these things, but there are some folks who have been able to get very deep on these problems.

That's what we're going to be talking about. So stay tuned. It's going to be interesting, hopefully educational in some cases, and I'm sure there'll be a lot of AI WTAF. We're really at a point where some of these AI tools, processes, and the ability to automate things like never before are going to lead to some pretty crazy problems. But there are also things that are just new versions of old things that have been happening on the internet for years and years. I'm hoping we can cover a lot of that here.

Today I have a really interesting conversation with Jason Zoltak. Jason is the CEO and co-founder of Tofu, which uses AI and machine learning to tackle fraud and security problems in recruiting and hiring. I think you'll find the conversation very interesting, and I'd love your feedback — this is obviously the first episode of this podcast, so please send it my way. I'm guessing you probably know me if you're watching this first episode, but if you don't, I'm also looking forward to getting to know the kinds of things you'd like to hear in the future.

Enjoy. Thanks.

Rob: We met recently and I thought the problem you're working on is really interesting. You're a startup founder working on this, you have investors and everything else. Maybe just tell me a little bit about what you're working on, and then how you got into it — how the company has evolved over the last couple of years as you've been attacking this space.

Jason: I worked for a company called Shiftsmart before Tofu, which is a labor marketplace for blue-collar jobs. Circle K goes to Shiftsmart and says, "Hey, we need 10,000 gas station attendants," and Shiftsmart places them.

I had this idea for Tofu at the time, which was basically bringing on dormant supply of talent, but for the white-collar space, whereas we were doing it for blue collar. We started working on that, and I left Shiftsmart to go all in on it.

We ended up pivoting into AI for recruiting, just because the marketplace — good idea, tough to scale. So that was end of 2023, early '24; call it January '24. We had this AI resume screening product, and when we gave it to customers, the few we gave it to came back and said, "Hey, this is great, but you're screening people named Rob Leathern and we're getting on the phone with them and they don't sound like Rob. They sound very different from what you might think Rob Leathern would sound like."

We didn't really understand what they were talking about. So we asked customers to send us video recordings of their interviews and screenshots. My co-founder and I even turned the caller ID off on our phones and started calling different resumes to catch people in lies. You'd call someone whose resume said Rob, and you'd say, "Hey, is this Matthew?" And they'd say, "Yes, it is."

That uncovered this whole thing of — okay, clearly there's some identity misrepresentation going on. And then doing more and more research, you start learning about North Korean IT workers and laptop farms, and how since 2020, when everyone went remote because of COVID, a whole new type of fraud exploded, with AI making it super easy to fake credentials and all of that.

We put out a very small feature for it, and the buyer response was such a stark contrast to AI resume screening. For the people who knew it was a problem, it was a hair-on-fire problem. They said, "I want it now." It was only a few customers we were talking to, but the difference in how they reacted gave my co-founder and me the confidence to say: every single company is building something in the AI recruiting space, and candidate fraud detection feels like a zag when everyone is zigging.

So we decided to go down that path and bet on it becoming a much more pervasive issue than it was at the time. It took a year for people to care. Then when CNBC and the Wall Street Journal and the larger media outlets started reporting on it, that's when things really started to take off. And yeah, here we are.

Rob: Was there a key moment for you? For me — I've worked in this area before, but the thing that made me want to work on AI for scams was seeing a video of a woman feeding money into a Bitcoin ATM. I just thought, this should not happen in our society, someone feeding their life savings into an ATM for scammers. Was there a moment like that for you, or was it more a gradual realization?

Jason: We stumbled onto the fraud opportunity, but the catalyst was recognizing that I just find AI a much more interesting problem. I was captivated by the whole essence of it.

On the internet it's always been a little interesting to figure out whether the person you're talking to is who they say they are, right? Chatbots, MSN — since the beginning of internet chat and discussion, do you really know who's on the other side of the keyboard? But the concept of identity, of misrepresenting who you are, with artificial intelligence making it so much easier — that felt like a much more interesting problem, especially since these models are getting insanely good month over month. I think it's going to become really difficult to discern the difference.

So it was that, plus the fact that when we were talking to customers we noticed how differently they reacted to that problem versus the one we were working on. From a business point of view it just felt right, and like that was the right direction to go.

Rob: It's interesting, because there's always been a bit of an incentive for people to embellish their resume or their LinkedIn profile. So you're in a space where a hiring manager is already a little bit circumspect, and now you're going even deeper. Do you find you're also running into the lighter, less serious kind of fraud — embellishment, shall we say — where they're not faking where they're coming from entirely?

Jason: Yeah. There's such a spectrum of what people define as fraud, depending on who you talk to. A recruiter might characterize someone using AI to help them get through an interview as fraud, or as cheating — or maybe not even cheating in this day and age, because companies ask their employees to use AI. So, can you use AI in an interview?

We often get asked, "Hey, can you tell us if the candidate is inflating their experience?" That's hard to do unless you have the job description from when they got hired, or you actually talk to someone who worked with that person. We never stake the claim that we can do that. We're focused on identity theft and identity misrepresentation rather than skill embellishment.

But depending on who you ask, people will characterize fraud differently, and we often have to bring folks back to: that's not really fraud — this is fraud, or at least this is how we characterize it.

Rob: That makes a lot of sense. About 18 months ago, some friends of mine from Meta and Google — one of them was running a startup engineering team — sent a message to our group chat saying, "Hey folks, just be careful, we're starting to see resumes in the pile that appear to be North Korean." That was one of the first times I thought, wow, this seems like an actual real issue. Can you talk a little bit about the North Korean problem?

Jason: It's funny — well, it's not funny, I guess. Every single company I've spoken to is affected by it. Literally every single one. If you're hiring for technical remote jobs, there's a 99% chance a North Korean has applied to your company. Whether or not they've gotten far is a whole other story, but there's an email, a resume, or a phone number in there that links back to the DPRK.

Certain companies get hit harder depending on the type of information they have, and whether that lets someone get paid more — because that job in that industry pays more — or gives access to more sensitive customer information, where getting in and holding the company ransom is a more lucrative payout.

A year, year and a half ago we were seeing a lot of fake LinkedIn profiles and fabricated identities. Now I'm seeing tons of identity theft. There's a really interesting article — I think it was on Trend Micro, I'll send you the link — about how LinkedIn is a massive hotbed for stealing data and cybercrime. So one thing we see now is LinkedIn account takeovers, whether through a password leak or something similar.

But also, a lot of North Koreans will assume the identities of people whose LinkedIn profiles don't have a profile picture. If I go to your LinkedIn profile, Rob, and you don't have a photo, I don't know what you look like, and I can't describe you unless I go deep diving on the web. So they'll apply using those identities and create fake emails and fake phone numbers, and when they get on the phone the recruiter has to assume that's what the person looks like, rather than being able to corroborate it against the LinkedIn profile. I see a lot of that.

I'm also seeing a lot of deepfakes, more and more, and they're getting really good. Depending on how deep you want me to go we can get super in the weeds, but that's the shift I'm noticing. And there are a lot of different entry points these people are calling in from or operating from — Nigeria, Islamabad, Russia, China, places in South America.

Rob: Where's the blame? If you had to say what's most broken that leads to this situation — is it that hiring companies are trying to move too quickly, or trying to hire remotely when they should be optimizing for something different? What's your assessment of the root causes of why this is suddenly a thing? You mentioned COVID was a watershed, where now people hire remotely and have the infrastructure to support it — mailing laptops to people wasn't as much of a thing five or six years ago.

Jason: Really good question. I think about this a lot.

Because of how interconnected and also disconnected these systems are, going platform to platform makes the association of information harder. As an example: you'll speak to a recruiter and they'll say, "Well, I looked at the LinkedIn profile and it was verified." Okay, but when a candidate applies from LinkedIn, you then hop off into an ATS. You've moved from one system to another and you lose that connection. It doesn't really mean anything anymore that a LinkedIn profile is verified, because once you move into the ATS you lose that interconnectivity. I know LinkedIn works on this problem, but the lack of ubiquity or interconnectedness between these platforms makes it harder to stop once you move from one thing to another.

And obviously there's no good internet passport or digital ID that someone can use — which may or may not be something we're trying to do. So first and foremost, the problem is: what is the source of truth for identity information? A lot of companies have tried to solve this — "apply to companies through this one platform" and so on — but those platforms have such low utility other than finding a job that they don't stick.

The second issue is that this is a security problem where the recruiter is the first line of defense. Security teams — CISOs, directors of security — have been aware of this for a little while, but I think they initially put talent acquisition teams on the front lines to solve it. Then you get into the budget issue: TA teams may not have that budget, so it becomes a decision about whether to spend on it or not. And you have TA teams that think they can do it without something scalable, so they become this house of cards that lets these people through.

Then there's confirmation bias as you move a candidate from one stage of the funnel to the next. By the time they get to final-round interviews, you're more and more bought in on the candidate. You don't think it could happen to you. And then the person gets let in.

So I feel security is being brought in more, or at least taking ownership of this problem now. I'm starting to talk to more CISOs. A year ago it was definitely being pushed onto TA. When we first started reaching out to companies they told us to kick rocks, and honestly the majority of them came back later to say, "Hey, it's actually a problem for us."

Now it's being brought to light and everyone's in a reactive position. But between popping off different platforms, the budget struggle, and being forced to meet your hiring goals while also adding this level of security with finite budget — those are competing priorities. I think this should probably start moving toward the security budget.

Rob: That makes a lot of sense. You talked about identity, which is a huge can of worms. What do you think of the efforts by folks like LinkedIn to get people to verify? You mentioned that compromised accounts make verified accounts more valuable if they can be taken over. LinkedIn's working with Clear and others — what do you think about that whole area?

Jason: Unless it's real — unless LinkedIn forced everyone to validate their profile or they can't use the platform — I honestly wonder why they don't do that. I think you have to. There's a give and get, right? There has to be something driving the behavior.

I don't think I'm verified on LinkedIn. I've never felt a reason to do it. And I think that's possibly a microcosm of the whole challenge with getting people to verify themselves: why should I care to do it? It's not regulated. If there's nothing forcing you to do it — especially with finding a job, which hopefully isn't a month-over-month thing, it happens in a very short spurt and then you don't do it again for years — because it's so infrequent, it's possibly a harder behavior to enforce.

Rob: One thing I've noticed, having started companies and worked at big companies like Google and Meta, is that some folks who've been at these big companies a long time have a LinkedIn profile with almost nothing on it. Just "Google, 12 years," and you have no idea what they worked on. How do you think about people with really thin online footprints? Does that create issues in itself?

Jason: In terms of identifying them?

Rob: Yeah — feeling like, is this actually real? This could be the best engineer who ever worked at Google and they wrote almost nothing on their profile. We sometimes overemphasize that information, even knowing some people just never had an incentive to put it there.

Jason: I guess it depends on the use case. If you're looking for a job, you're really silly not to make it look like you're legit. Some people don't think about it, though — a few really great engineers I've worked with didn't have LinkedIn profiles at the time, or they took them down.

I think it behooves every single person applying for a job now to display some information that gives the person on the other side confidence that they're a real person. The whole candidate fraud thing has become really insidious-feeling — you question everything, you second-guess everything, especially as a recruiter, maybe if you don't have a security background. So if you're looking for a job, beef up your information. Make it seem like you are who you say you are.

I even see people putting their actual email addresses in their LinkedIn profiles. You'll see people write in their bios, "I'm not applying for a job right now" — I've seen that countless times, because people impersonate others. But again, there's this lack of incentive. People are working at their jobs, they clock off, and they're not really thinking about someone pretending to be them. And for the people who don't want to be found, that's deliberate.

Rob: Totally. What happens when a candidate — this must happen, it happens with every trust problem — gets flagged as an issue, as a bot or as fraud? Does that happen often, and how do you deal with it? False positives seem to happen in every kind of trust and safety problem.

Jason: Yeah, totally, that can happen often. Genuine people can unintentionally show signs of suspicion without knowing it.

For a very weak or light candidate fraud detection product that isn't looking at multiple things — if you index off email age, then people who apply to jobs with an email they only use for job applications will get flagged. Those emails won't show up on the internet. They won't be caught in the crosshairs of a data breach or something that might indicate someone is legitimate.

It's probably hard to educate the entire job-applying population that email age is a sign of authenticity, because you obviously want the person to be sharing that information. We have things in our product that get around a possible lack of digital footprint across email and phone. But again: if you're applying to a job — especially if you're an engineer and you know about this problem — why would you want to make it easy for someone not to look at you?

Rob: Google just announced they're letting you select another username for your Gmail account while still retaining access to the old one. I'm curious whether that'll mess with some of this, because all of a sudden you have a never-before-seen Gmail address on an old account. I wonder if that's going to throw some of these algorithms off.

Jason: That is interesting. I feel like a good OSINT tool will have to be able to connect those two — going through password reset flows or whatnot to connect a username to a previous email, then link that back to get a breach count, first-seen date, things like that. But yeah, that's interesting.

Rob: When it comes to interviews, I've talked to a bunch of people who say we should be doing all our interviews in person — the person could be using ChatGPT to answer questions, it might not even be the same person. Is that something you're seeing, where you have a different person interviewing as an expert, helping people? What's the range of things you see in the interview process specifically?

Jason: Maybe this is a good time to announce it: we're coming out with deepfake and proxy detection. It's actually already being used by customers, and it solves for two things.

One is deepfakes — making sure people aren't using AI to change what they look and sound like. The other is what we call proxy detection; you could call it swapping. Basically, continuous monitoring of identity across the interview process, so you don't have a stand-in taking the technical round for someone who did the initial phone screen. That happens all the time.

The more enterprise you go, where you have larger orgs and less likelihood that one interviewer talks to another, the more that's going to happen. If you and I both worked at the same company and interviewed someone, it's less likely we'd find time to chat and say, "Hey Rob, what did Mark look and sound like on that call?" Those conversations don't come up. They're also a little PC and maybe taboo to have in the workplace. That works to the benefit of someone trying to get through.

My personal opinion is that six to nine months from now, the reaction to deepfakes is going to explode in the same way it has for fake applicants, because these things are getting so good.

A few weeks ago I had to reach out to someone because we saw a deepfake that was so good I don't think 98% of people would have been able to tell the difference. The guy actually ended up knowing he was being deepfaked — I think a few other people had reached out to him too.

What happened was, at a customer of ours, a new hiring manager joined and wasn't aware Tofu was a thing at the company. So they were just working inside the ATS and moved this person forward. The candidate was doing a technical screen and sharing their screen, and they hit F3, so all the windows expanded — on the Mac. You could see AnyDesk, DeepLive, an interview response assistant. There was Google Maps open on San Francisco with the person's address, so they could answer questions about where they lived. You had all this person's information. It was wild.

What was even crazier is that in the video this guy is getting up, standing up — you can see his whole torso and every single frame stays intact. There's none of the waving-your-hand-over-your-face stuff that a lot of people use as a tactic to identify a deepfake. This was so good.

So I think that's going to be the next holy-crap moment: it's going to get really good, and then teams will be scrambling for a solution.

Rob: That's really scary, and unfortunately not unexpected, because these things have been getting so good. Did you see the — I think it's a Chinese company — app where you can fully change what you look like? I think it's aimed at social media influencers. It's scary good. I'm sure that's going to make its way here really soon.

In terms of the incentives for the people doing this: I assume it's not just "I get hired for $80,000 a year and I pocket $80,000." There are a bunch of other ways they're trying to extract value from getting inside these companies. Can you talk about what that looks like — what's the benefit to a North Korean, or someone else, in getting one of these jobs?

Jason: I'm happy to answer, but I'm also curious how you think about it — you've been in this a really long time.

I think if you get ten jobs and you're hired for two months, that's a lot more than $80K a year. Multiply that by however many people are being forced to do this, whether under the North Korean regime or something else. Even if you're only employed four weeks, eight weeks, that's 40 grand in a month, possibly more. It adds up. These people don't care about getting fired. If you can scale yourself and shop the work out, maybe you get away with it for a while.

Then it's the exact reason companies have insider threat teams. Once you get inside a really large organization, there's IP theft risk. Depending on the company you infiltrate, there's a lot to be gained by ransoming that company for large sums of money. What are some of the top reasons — other than stealing secrets, corporate espionage — from your friends who work on insider threat and things like that, that these things happen?

Rob: Some companies are just very powerful vantage points for things happening in other places. If you're one of the big platforms, certain kinds of data can give people all kinds of insights. It could even let you insider trade on the stock market, depending on what flows they're accessing.

So I can see both angles. There's the direct monetary benefit. I was actually talking to someone about a different problem — scraping video data and reselling it as AI training data. Their point was, "We couldn't figure out the economics. Why would they be scraping us?" And I said, well, they're probably selling the same data to multiple different partners. When you can do something in parallel like that — which is the point you were making about holding multiple jobs at the same time — that's obviously a big benefit.

But I also think there are companies at a strategic crossroads where they have software on a lot of different machines and could potentially be a distribution point for malware. I worry that we're barely scratching the surface at understanding some of these things. And I only worry more when agents — pieces of software — are autonomously updating open source packages from multiple different entities. The surface of the threats we face seems to be multiplying exponentially.

Jason: What you just said makes me think of whatever movie has a villain who tries to get access to the sewage and water system and poison it, because that thing goes through the entire city, into everyone's homes and faucets. You're right — a company that's so interconnected and embedded in people's lives, if you can get access to that, you can do a lot of damage.

Rob: And what I'm also hearing from you is that there's an incentive-slash-effort-slash-expertise issue: the folks on the front line are incentivized to move people through the process. Hiring managers want to — unfortunately, we're at a stage where sometimes hiring managers want to close candidates before the headcount is taken away from them. So the people on the front lines of this problem aren't always equipped or incentivized to address it as quickly as an insider threat security team might be. That seems like it has to be the next way these things evolve internally at a lot of companies.

Jason: One of the things we bring up in conversations is: how many times does a recruiter need to send a hiring manager a fake candidate, where the hiring manager gets on a call with them, before that hiring manager fully loses trust in the recruiter and the recruiter's job is in jeopardy?

There are two opposing forces. The recruiter needs to find people and send candidates to the hiring manager, otherwise the hiring manager asks, "Where the heck are the people I'm supposed to be interviewing?" On the other hand, you want to avoid the conversation of, "Why are you sending me people who aren't who they say they are?" Now the recruiter is caught between a rock and a hard place.

Rob: I imagine recruiters and these systems are already deluged with candidates, and this is wasting a bunch of their time. Do you have a sense of how much time recruiters are spending dealing with fraudulent candidates?

Jason: Before I answer that, I want to acknowledge your earlier point: depending on where your jobs are posted and the business model of that platform, some are incentivized to just send the house — push all applications into your pipeline, whether or not they're fraudulent. So consider where you're posting your jobs. That's definitely a thing. If you're paying LinkedIn for a number of applications, that Easy Apply button is like opening the dam.

In terms of time saved, it depends how you think about it. You have AI resume screening tools with fraud detection embedded in them now. Time saved is this weird ROI thing — whenever we're positioning this, it's not what we lead with. But say it takes you 10 seconds to review a resume and you have a thousand resumes; divide by six, that's 166 minutes. If X percent are fake, maybe we're saving 30 or 40 minutes.

Where I think about it is opportunity cost: how many candidate slots are being taken up by fake people instead of genuine ones? How many fake people do you talk to in a week where you could have talked to a real person? What's a recruiter's time worth — say $120K a year, so $30 per half hour, $60 an hour that you're being paid to screen people? Multiply that out and it's in the thousands, or tens of thousands, depending on how many recruiters you have. That's more quantifiable than hiring manager time, and it's how we typically define it.

But the average fraudulent pipeline has 56% fake applicants or more. I've even seen some at 80%.

Rob: That's pretty bad.

Jason: It's a lot of time. But to me that's the second-order consequence. The first is that if they get past that first call, the confirmation bias of moving them through puts them at higher risk of actually being hired.

We onboarded a fintech company that had a candidate in final-round interviews. The head of talent decided to run fraud detection on everyone past application review, and this person who was in final rounds, pending offer, lit up like a Christmas tree on our platform. The recruiter reached out and said, "Hey, I think you guys made a mistake. We're about to make this person an offer. We just spoke to their references. Everything seems good."

We said, "Well, you should reach out to the candidate on LinkedIn and ask if they ever applied." Lo and behold, they do — and the candidate says, "Yeah, I've never applied to your company. You're the fourth recruiter who's reached out to me. Is this the email they're using? This is my real email."

I think for that company a light bulb went off: oh crap, we actually have a problem. So that's what we're trying to help you avoid — but also the time savings, 100%.

Rob: My final question in this area: some of this is monetary, some is time-saving, but there are also legitimate national security issues and lots of other problems here. Do you think there's a role for legislation, or for government to step in and put things in place to protect U.S. companies? Or is that premature?

Jason: I don't know. I'd actually put it to you — I'm curious, what's something you never thought would have federal or state legislation pushed down, where now there's something in place for it?

Rob: I typically have the opposite concern. I've been hoping there'd be a federal privacy law for a really long time, and of course there isn't. I was just in DC, and folks were talking about various anti-scam things they want to put in place. There's also such a patchwork of state laws around deceptive practices. So I think it's really tricky.

A lot of the time we unfortunately pin our hopes on regulation because we think it'll magically solve things, when in fact private companies doing stuff can move so much more quickly. So I generally think yes, there is a role to be played — there's a role for lawsuits and other things too. But honestly, a lot of it can be companies coming together and figuring out things that are in their interests and can also protect people and move things forward. That's often a better way.

Jason: That makes a lot of sense. When you talk about that, I think about Uber being so quick to upend the taxi industry, and then everything having to be reactive from that point. You could probably have a few of the biggest companies band together to do something like this.

I think if legislation comes, it's probably going to come from something really bad happening. If a defense contractor were suddenly infiltrated by one of these people and it somehow became a national security issue — now it's about weapons. I'm just throwing out random thoughts. But I feel like you have to be pushed to that point for something to happen. Otherwise it's not really seen as a problem.

Rob: You mentioned the deepfake work you're launching. Are there other things we should be looking for from Tofu in the near future?

Jason: Yeah, we're putting out something tomorrow that I'm super excited about. We can do a whole recon of your ATS. We can drop into your applicant tracking system and, in the same way a pen test does a vulnerability assessment of your website or app, we can tell you immediately — within 20 minutes — where all the fraud exists in your pipeline.

So it's either really scary, or really reassuring that you have no problems. Or you find out you have someone in final-round interviews, or that you've already hired someone. I'm both excited and nervous to see what people's reactions will be, but we're trying to build the security layer for this problem in this industry.

Right now, the way most products are set up, you have to do a lot of button clicking. You have to create an agent to go scan resumes. It's based on API calls, so you can't figure out if you have fraud until you go run something. We want to offer immediate value right up front, the second you integrate your system. So any customer who signs on to Tofu, or anyone currently on Tofu, is going to have this live, breathing, continuous monitoring recon system that's just swatting people away as they come in.

Rob: That's awesome. I think that'll give people a lot of either peace of mind, or a rush to fix the problem because it seems worse than they thought.

Jason: And it'll poke holes at all our competitors too, which is always a good thing. Nothing better than saying, "Well, you think you have a good product? Why don't you hook this up? It costs absolutely nothing." And we'll see who walks.

Rob: Jason, really appreciate you taking the time to chat. Really excited to keep track of what you're going to be up to. This is a very important problem, so I'm glad you're working on it.

Jason: Appreciate you having me on, and honored to be a guest, and in the presence of other people who've been in this industry far longer than I have. So thank you.