Vijay Balasubramaniyan has spent twenty years on one question, and generative AI made it everybody’s.
“because fundamentally, generative AI has brought into question what is human and what is machine.”
Vijay Balasubramaniyan, at 10:42
Vijay is the co-founder and chief executive of Pindrop, the Atlanta company whose software listens to phone calls for fraud and for voices that aren’t human. Before it he was an engineer on voice systems at Intel and Siemens, a researcher at IBM and a software engineer at Google, and his PhD at Georgia Tech was on telecommunications security. In 2010 he published the paper that showed a call carries traces of every network it crossed, and in 2011 he started Pindrop on it with Paul Judge and his advisor, Mustaque Ahamad. By the end of 2018 its customers included eight of the ten largest U.S. banks and five of the seven largest U.S. insurers (Pulse 2.0). We recorded in the weeks after Pindrop’s May 2024 report on fraud and deepfakes, and published the episode on October 30, 2024, with 46 words of show notes. This page is the rest of it: a security person who is also a voice person, reading an encrypted call, the paper that started the company, the years the phone call was supposedly dead, a brother in security, what Pindrop does better than anyone, “welcome back, John,” one deepfake a month and then one a day, Giraffe Man, the first case of election interference, why people can’t hear the fake, a warranty that works like a bug bounty, being human again, five seconds of a TikTok, the best day and the hardest one, the mental model he’d give his younger self, the product he still wants someone to build, and what Pindrop did in the two years after the episode came out. There are four things to try along the way: a quiz, a scorecard for whether a cloned voice would get through your phone lines, a calculator for what a trust problem costs your growth, and a timeline you can filter.
A security person and a voice person
I opened with the question I ask everyone, whether he thinks of himself as a security person. He does, with an addition.
“I do, but you know, the interesting thing is I think of myself as a security person and a voice person.”
Vijay Balasubramaniyan, at 0:48
The voice part came first. His LinkedIn has him as a software engineer at Intel from August 2002, a senior software engineer at Siemens from July 2003 to July 2005, a research engineer at IBM from May 2006 to August 2007, a software engineer at Google in the summer of 2008, and back at IBM for the summer of 2010, where the work was on Twitter accounts that spread dubious information while looking legitimate (LinkedIn). On the tape, Intel and Siemens meant voice “for old landline switches,” and Google meant voice over IP, “the more modern Google Chat, Google Hangouts.” The security part was Georgia Tech, where he finished a PhD in computer science in 2011 (Georgia Tech Computing on X) with a thesis on telecommunications security, and where he picked up patents in voice over IP security and scalability along the way (USENIX Security 2018 speaker page).
“And then when I did my PhD at Georgia Tech, it was marrying that voice background with security, which is interesting, because a lot of folks in security focus on the digital side of things. Very few focus on the voice.”
Vijay Balasubramaniyan, at 0:48
That combination is the whole company. Most security tools look at packets, files and logins. Pindrop looks at the audio of a phone call, and for most of its life the question was whether the person on the line was the account holder. Now it is also whether the voice belongs to a person at all.
Reading an encrypted call
Vijay’s answer reminded me of the paper that first got me excited about this kind of work. I was in graduate school at the University of Minnesota, working on the Tor network with Roger Dingledine, and a paper came out showing that you could tell which language people were speaking on an encrypted internet call. Vijay knew it immediately: “That’s Fabian’s paper.”
It is. Charles Wright, Lucas Ballard, Fabian Monrose and Gerald Masson, all at Johns Hopkins, published it at USENIX Security in 2007 under the title “Language Identification of Encrypted VoIP Traffic: Alejandra y Roberto or Alice and Bob?” Their 21-way classifier named the language of an encrypted call with 66 percent accuracy, almost 14 times better than guessing (USENIX). Vijay explained why it works better than I could have.
“But you know, his paper was, again, super eye-opening, because they figured out, because of the variable bit rate that a lot of these audio systems use, the length of the encryption is revealing in what you’re saying.”
Vijay Balasubramaniyan, at 2:42
Encryption hides the content of each packet, not its size, and a codec that spends fewer bits on silence and simple sounds makes the sizes follow the speech. He also remembered the order the work went in: language first, “and then they started cracking words in that language.” The words came in 2011, in “Phonotactic Reconstruction of Encrypted VoIP Conversations: Hookt on fon-iks,” from Andrew White, Austin Matthews, Kevin Snow and Monrose at the University of North Carolina, with Matthews listed in linguistics as well as computer science (IEEE Symposium on Security and Privacy). That is the linguist Vijay meant when he said Monrose “combined with a researcher who was actually a linguist,” and it is why he loves the paper.
“And that’s why I love that paper, because we have a common kinship there. I come from a very different background than security, but then made security my own, much like Fabian [had] done it with his researchers.”
Vijay Balasubramaniyan, at 2:42
One small update to the tape. Vijay thought Monrose had been at Georgia Tech “for a brief moment.” He is there now, as a professor and the Julian T. Hightower Chair in Cybersecurity, welcomed in August 2022 after years as a Kenan Distinguished Professor at UNC Chapel Hill (Georgia Tech).
Where did this call come from?
The “Pindrop paper” Vijay mentioned is his own, and it is spelled with a zero. “PinDr0p: Using Single-Ended Audio Features To Determine Call Provenance,” by Vijay, Aamir Poonawalla, Mustaque Ahamad, Michael Hunter and Patrick Traynor, appeared at ACM’s Conference on Computer and Communications Security in 2010. It listened only to the audio at the receiving end, ignored the voice itself, and used the artifacts that each network leaves on a call to work out where it came from. Combined with machine learning, those features traced a call’s path through as many as three providers with 91.6 percent accuracy, and told apart traditional phone lines, cellular calls, Vonage, Skype and other phones with better than 90 percent accuracy. Ten years later the paper won the conference’s Test of Time award (Business Wire).
The company followed the paper closely. In September 2010 Vijay and Ahamad started a Georgia Tech VentureLab project called Telineage, and in 2011 Vijay and Paul Judge founded Pindrop, with Ahamad also listed as a founder, licensing the technology from the Georgia Tech Research Corporation (Wikipedia). His LinkedIn dates the company to April 2011 (LinkedIn).
The idea is worth holding onto, because it runs under everything else on this page. Caller ID is a claim the caller makes. The audio is evidence the network leaves. A fraudster can type any number into a spoofing tool, but a call routed through an internet phone service in another country still sounds like one, to software that knows what to listen for.
The phone call is dead
Thirteen years as co-founder and chief executive is a mark of success, and I said so. Vijay gave the credit to the team and to luck, and then told me what raising money for a voice company sounded like in 2011 and 2012.
“I remember every time I would go try to raise a round, the common refrain was, you’re doing this in voice, or does that mean a phone call? Oh, the phone call’s dead. Like, no one’s, no one cares about a phone call.”
Vijay Balasubramaniyan, at 4:18
The record lines up with the story. Pindrop’s first money, in 2012, was about $1 million from Andreessen Horowitz, other firms and angels, and Vijay said that round took about six months. Then came an $11 million Series A in June 2013, led by Andreessen Horowitz and Citi Ventures, $35 million in February 2015, led by Institutional Venture Partners, and $75 million in 2016, led by Google Capital, for a total of $122 million (Wikipedia). By then, in his telling, “came the Alexa,” and voice assistants changed what investors heard. The $90 million Series D in December 2018, led by Vitruvian Partners with Goldman Sachs, EDBI and others joining earlier investors, came with customers in eight of the ten largest U.S. banks and five of the seven largest U.S. insurers, more than 200 million consumer accounts protected, and revenue growing at a 137 percent compounded annual rate over three years (Pulse 2.0). “And then our Series D we raised in six weeks.”
His explanation for the change is the reason the company exists: “voice has become so important for people, because that’s what makes us human.” The example was at home. His son had started talking with ChatGPT-4o, OpenAI’s voice-capable model, and asked to do it almost every day, which Vijay greeted the way a voice security person would: “okay, I don’t know where the security of this thing is, so I’m really not going to have too much of your voice out there.” He also has a number for why voice wins as an interface, “you speak about four times faster than you type,” and I added the part I care about, that voice carries far more than the words.
“Our brains are almost hardwired to detect sentiment that’s underneath the words, that’s in between the words.”
Jon McLachlan, at 6:57
Vijay’s version came from Pindrop’s own chat channels: lots of miscommunication on Slack, fixed by a simple phone call. It is a sweet point, and also the threat model in one sentence. We trust voices because they carry so much. A clone that carries the same things gets the same trust.
A cat and mouse game
How he got into security is a family story. His older brother worked in security, still does, and loved it.
“And he would tell me about what an incredible cat and mouse game it is. Something that’s as dynamic as security. Like, you can never say, okay, you’ve solved a problem, right? Like, it’s always, whatever is your best solution is always going to get beaten, because you have a very smart adversary on the other end.”
Vijay Balasubramaniyan, at 7:58
Georgia Tech, he found once he arrived, was one of the strongest security schools in the country, and then he met the professor he did his PhD under, Mustaque Ahamad, a co-author on the PinDr0p paper and a co-founder of the company. The brother never joined Pindrop. “So yeah, we are a strict, no family involved in the company.” I told him about my two brothers, one I grew up with in Minnesota and one I chose in California, and we agreed that a lot of looking up goes both ways.
The cat and mouse line matters for what comes later. A detector that wins today gets beaten tomorrow, and Pindrop’s answers to that, a consortium of known bad actors and a warranty that pays out when it loses, are both ways of living with an adversary that adapts.
Better than anyone else on the voice channel
I asked the question I ask every founder: what do you do better than everyone else in the world?
“So, you know, there’s several things that we do better, but at the core of what we do, multi-factor authentication on voice channels, really well, better than anyone else. And we do deepfake detection, or fraud detection, on voice channels better than anyone else.”
Vijay Balasubramaniyan, at 10:42
The deepfake part started early. “Like, our first patents on deepfakes are back in 2015,” he said, and Pindrop had been at it for eight years, “way before it became hot.” His example from the party-trick era was a documentary: “We are the ones who discovered the deepfakes in Anthony Bourdain’s documentary two years back.” The film was Roadrunner, released in July 2021, three summers before we recorded, and its director said three lines in Bourdain’s voice had been generated with AI (NPR). The coverage we could reach confirms the three lines, not who found them, so Pindrop’s part is his account.
“So at that point in time, it was a cool party trick.” Three years later it was, in his words, one of the biggest tailwinds the company has had, and the reason is the line at the top of this page: generative AI has brought into question what is human and what is machine.
Welcome back, John
Pindrop had just published its yearly report on fraud and deepfakes, so I asked what surprised him. He had two findings. The first was volume.
Pindrop’s 2024 Voice Intelligence and Security Report came out on May 22, 2024. It found that fraud against contact centers had risen 60 percent in two years, to its highest level since 2019, projected that one call in every 730 to a contact center would be fraudulent by the end of the year, put the deepfake fraud risk to U.S. contact centers at about $5 billion, and found that retail fraud, refund abuse included, had quadrupled among Pindrop’s call center customers through 2023 (MarTech Series). On the tape Vijay rounded the first number down, “So over the last couple of years, it’s grown by 50%,” and the report says 60. His explanation for the rise is about money that stopped flowing: “a lot of fraudsters who used to get easy money back in the day from PPP loans and unemployment now don’t have the government giving you free money.”
His second ingredient was the breaches. “Data breaches have gone up by 78%,” he said, with 3,000 incidents in 2023 against 1,800 in 2022, and that one checks out almost exactly. The Identity Theft Resource Center counted 3,205 data compromises in the U.S. in 2023, a record and a 78 percent jump over the year before (Cybersecurity Dive). What the two have to do with each other is the best explanation of phone fraud I’ve heard.
“And so being able to use the call center to say, punch it in into all of the top ten banks, and if one of them say, welcome back, John, your account balance is so on and so forth, I can then go ahead and know John banks with bank number one, and then take over his account.”
Vijay Balasubramaniyan, at 11:57
A stolen Social Security number doesn’t say where its owner banks. The phone line can. An automated phone menu that greets a caller by name, or reads out a balance before anyone has been verified, answers the question the breach left open, and the fraudster calls back later with the rest of the stolen answers. The fix is unglamorous and very effective: confirm nothing, not even that an account exists, until the caller is verified with something they have. The scorecard below starts there.
One a month, then one a day
The second finding was the one that made the episode’s title feel modest.
“Back in 2020, we used to see one deepfake attack a month at our customers. So it used to be one a month, and we’d look at it, we’d thwart it. In this year, every day, there is a deepfake attack per customer.”
Vijay Balasubramaniyan, at 11:57
He put a number on it, a 450 percent increase, and he was careful about what it measured: “The 450% is if you compare just the first four months with all of last year.” Those customers are some of the largest banks and insurers in the country, so one a day per customer is a lot of synthetic voices reaching agents who were hired to be helpful. On October 30, 2024, the day this episode came out, Pindrop posted the same comparison from the New York Stock Exchange.
Our CEO Vijay Balasubramaniyan is LIVE from NYSE. Last year, we saw one attack per month for Pindrop customers, which includes the top banks and insurance companies. Now, we see one attack per day per customer – that’s a 1400% increase in just one year.
Pindrop (@pindrop) · October 30, 2024
The windows differ, and the shape doesn’t. Whatever the exact multiple, a deepfake went from something a fraud team studied to something it saw every day, in about the time it takes to raise a Series A.
Giraffe Man
My follow-up was whether Pindrop could tell which model made a voice, if all the fakes from one place shared a fingerprint. That, it turns out, is the fun part.
“In fact, if you come to Pindrop, every conference room is named after a bad actor that we’ve caught. So we have, yeah, we have Pepe and Chicken Man and Giraffe Man.”
Vijay Balasubramaniyan, at 15:11
Giraffe Man uses generative AI, and the name is a diagnosis. “The only person who could have produced this audio is someone with a seven-foot-long neck, and their vocal cords are thrashing between that seven feet.” A generator doesn’t have a throat, so it can produce sounds no human vocal tract could, and software that models the anatomy can hear the impossibility. It is the same move as the 2010 paper, one level down. The network leaves artifacts on a call, and so does the machine that made the voice.
The consortium is what turns one catch into a pattern. Pindrop sees the same bad actors across many customers, so it can say where calls come from, West Africa, Eastern Europe or the Philippines in his examples, and which tools they used. The tool count is the part that should worry a security team. When Pindrop first ran into deepfakes in occasional events, “there used to be one voice cloning software. It used to be the software called Lyrebird.” By the end of 2023 there were 120 apps that could clone a voice, and by the end of March 2024, by his research team’s count, “There are 358 AI apps that you can use to clone your voice.” Each one, he said, “leaves behind a telltale sign of how inhuman they are,” which is how Pindrop tells them apart.
The first case of election interference
Then he told the story that made Pindrop famous outside banking.
“And we were the ones who discovered that it was ElevenLabs that was used to create Biden’s deepfake. And then ElevenLabs saw our blog and then shut down the actor who actually was perpetrating this attack.”
Vijay Balasubramaniyan, at 15:11
The calls went out the weekend before New Hampshire’s presidential primary on January 23, 2024, in an imitation of President Biden’s voice telling voters to “save your vote” for November (NBC News). One correction to the tape: Vijay said the Republican primary and Republican voters, and the record says the calls targeted Democratic voters (NBC News, The Boston Globe). The state’s attorney general, John Formella, put the count at between 5,000 and 25,000 calls.
Pindrop published its analysis on January 25, under Vijay’s name. The system broke the 39-second clip into 155 segments of 250 milliseconds each, compared it against 122 text-to-speech systems, and found “with a 99% likelihood” that the voice came from ElevenLabs or a system using similar components, a result its validation step put at 84 percent. Its liveness score, from 0 for synthetic to 1.0 for authentic, dropped below 0.3 two seconds into the call and stayed there (Pindrop). Two days later Fortune reported that the account had been suspended once the deepfake was traced to its creator, and that ElevenLabs couldn’t comment on specific incidents (Fortune). On February 6 the attorney general traced the calls to two Texas companies, Life Corporation and Lingo Telecom (Engadget). On February 25 a political consultant who had worked for a rival campaign said he was behind the call (Axios), and the next day a New Orleans magician said he had made the audio with an ElevenLabs voice generator (The Boston Globe). That is the “a month later” on the tape.
“That’s a great public service as well,” I told him, and he explained why they did it anyway: it was an election year, and “it didn’t even take the first month to happen.” The work brought Pindrop into the room with the people who run elections.
“But one of the things that the New Hampshire AG said is, how does it feel to protect the nation’s democracy?”
Vijay Balasubramaniyan, at 18:09
Pindrop went on to brief attorneys general on how to prepare and was on its way to the secretaries of state conference when we spoke. Vijay’s takeaway for his own team was that the mission keeps growing, from protecting bank accounts to this.
Two out of three times, they get it wrong
The skeptic’s question, which I asked for the skeptics: surely a person can just hear that something is off? Vijay answered with an experiment. At an attorneys general symposium, with New Hampshire’s attorney general’s permission, Pindrop played two versions of him, one real and one a deepfake. “But 40% of the people in the audience got it wrong.” Then he cited the research.
“And one of the things is, when studies have been done on humans and their ability to detect deepfakes, in particular, it depends on the medium, but on voice, they have an accuracy of 30%. So, really poor. Two out of three times, they get it wrong.”
Vijay Balasubramaniyan, at 19:53
The record puts people a little higher than his number and makes the same point. A meta-analysis of 56 studies published in December 2024 found that people correctly identify deepfakes 55.5 percent of the time overall, with a confidence interval that crosses 50 percent (Diel et al., Computers in Human Behavior Reports). A coin does about as well.
For Pindrop’s own accuracy he pointed to an outside test. “We had this NPR reporter who took different clips, and each of them are like five seconds long,” he said, and the vendors who claimed better than 90 percent were run through them. NPR’s version, published April 5, 2024: 84 clips of 5 to 8 seconds each, about half real radio from three NPR reporters and half clones of them made with PlayHT. Pindrop got all but three right. AI or Not got about half wrong, missing most of the clones, and AI Voice Detector misidentified 20 of the 84 at its original threshold (NPR). Eighty-one of 84 is 96.4 percent, and Pindrop said its model caught every deepfake in the set (Unite.AI). “We ended up in that study at 96[%] accuracy,” Vijay said, and that holds. The one thing the record moves is the runner-up. He said the next closest vendor got half of the deepfakes wrong, and by NPR’s count the second-best was AI Voice Detector, while the one that got about half wrong was AI or Not.
His explanation for the gap is the data and the customers who attack it.
“and we now have to protect eight of the top ten banks, five of the top seven insurances. Their red teams constantly are trying to beat our systems.”
Vijay Balasubramaniyan, at 21:29
The deepfake data set was 20 million clips by then, with fakeprints, his word, of more than 350 apps, and the bank and insurer counts match what Pindrop said in 2018. His conclusion about people: “It is damn near impossible.” That is the strongest argument for putting a detector on a phone line, and also the reason a detector shouldn’t be the only thing standing between a caller and a wire transfer.
A warranty that works like a bug bounty
The report came with something I hadn’t seen a detection vendor do: a warranty. Pindrop called its Pulse Deepfake Warranty “the first-of-its-kind.” It reimburses customers for certain losses from synthetic voice fraud that its products miss on eligible calls, at no extra cost for customers on a standard three-year subscription to Pulse, its deepfake detection product. Vijay’s line in the announcement: “We need good AI to beat bad AI” (MarTech Series). On the tape he explained why a company would sign up to pay for its own mistakes.
“But the point is, any security vendor who says they’re 100% sure of their capabilities is lying, right? And so we know there are going to be situations where we’re going to get beaten. We’re going to pay out a warranty as a result of that.”
Vijay Balasubramaniyan, at 22:40
Every payout is a missed fake with a dollar figure attached, which is exactly the sample a detection team most wants to study, and “the warranty serves almost like a bug bounty program for us.” I put it the way a customer would: “you’re going to put your money where your mouth is, and you have your customers’ back.” It’s a question worth asking any vendor whose product is a score. What do you pay when the score is wrong? A vendor that has priced its own misses has measured them, and a vendor that hasn’t is asking you to carry all of the risk.
Being human again
I asked what success looks like, and Vijay’s answer was not about Pindrop at first.
“Yeah, so you know, one of the things that we’ve always been excited [about] is a future where we get to being human again.”
Vijay Balasubramaniyan, at 23:53
Pindrop’s offices are close to Georgia Tech, and walking those streets now, he sees people buried in their phones, on collision paths, surprised at the last second by “the real world.” His hope is that the interfaces we invented because computers couldn’t handle our natural abilities, the keyboard and the mouse, give way to talking. The moment that convinced him came from a customer who wanted everyone at his company to have a vocal password. Vijay asked why, since nobody talks to their computer, and the customer pointed out that he spends eight hours a day on conference calls.
“So all it requires is a paradigm shift where, as I’m speaking, it naturally authenticates me and identifies me.”
Vijay Balasubramaniyan, at 23:53
He wants Pindrop to be “that security solution that makes us human,” which means two checks at once: that the voice belongs to you, and that it isn’t a machine pretending to be you. I pointed out that the talking computers are already here, ChatGPT and its peers answering at a depth Alexa and Siri never reached, and asked the obvious product question.
“So is there an app in the Zoom store that I can download, and it’s going to continuously do a background verification on all of the participants in the video?”
Jon McLachlan, at 26:37
“Yeah, that is the next frontier, right?” It arrived. Pindrop describes its own path as contact centers first, then audio deepfake detection in 2024, then meeting deepfake detection in 2025 (Pindrop). Pulse for Meetings was named one of TIME’s Best Inventions of 2025, and in March 2026 Pindrop put real-time deepfake detection and identity verification into Zoom Contact Center (Pindrop newsroom). Not quite the app in the Zoom store I asked for, but close enough that I’m counting it.
Five seconds of a TikTok
Pindrop has always sold to enterprises, banks, insurers and health care companies. The future Vijay wants includes the people those companies serve, and the scam he described is the reason.
“And they’ve become so programmatic at this right now that attackers are focusing on a county, attacking all the elders in that county, and then they’re attacking them over Friday, Saturday, Sunday.”
Vijay Balasubramaniyan, at 27:06
It is the grandparent scam, a call from a grandchild in trouble asking for a wire, and by his account local police start Monday with seven elders in one county out anywhere from $20,000 to $200,000 each. The grandchild’s voice is the easy part. A teenager has a TikTok video that’s 10 seconds long, he said, “a lot of these voice cloning apps require five seconds of your audio to completely impersonate you.” Protecting those calls means working with carriers and handset makers, “and that’s not an easy linear path, man.”
AI scams are up 900% in 2 years. Our CEO Vijay Balasubramaniyan told CBS News why deepfakes are now easier than ever to make – some take less than 10 minutes. After the Marco Rubio impersonation last week, it’s clear that anyone’s at risk.
Pindrop (@pindrop) · July 16, 2025
For the founders listening, I pointed out that a lot of Pindrop’s business comes from rules banks already live under, like know your customer, and Vijay agreed and said the same expectations will have to reach communications, because “I have seen really strong, even security professionals, [fall victim] to social engineering.” The reason is the best thing about people. When someone we love calls in trouble, we want to show up for them. A technology that quietly answers human or machine, and green lights the human, lets us keep doing that. We also took a short detour into Blade Runner and killer robots, which Vijay, who once shared a panel with the chief executive of Boston Dynamics, waved off: “And so this notion of killer robots is like super far-fetched.” Software is moving fast, and hardware is still hard.
Writing code, fighting crime
I asked for the best day he’s had at Pindrop. It was a customer day, when a customer came in and described one of Pindrop’s alerts, the red or green an agent sees on a call. On a Saturday, an account came up red. The agent sent the call to a fraud analyst, who shut it down and called the customer, whose identity had been compromised so thoroughly that the fraudster knew all the answers. She had been born in [1930], and the account held her retirement savings, about $90,000. The analyst reassured her that even at zero she’d have been covered up to $250,000.
“And she was like, you’re saying that, but if it had gone down to zero, I’d have had a heart attack, because my husband’s no more, and I wouldn’t know what to do.”
Vijay Balasubramaniyan, at 32:03
Pindrop keeps a Slack channel called Customer Anecdotes for stories like that, a retirement, someone’s school tuition, a 529 college savings account, an insurance agent’s $32 million book of business. That day produced a company t-shirt: “writing code, fighting crime, hell yeah.” My confession on the tape was that I’d make a horrible chief executive of a company like that, “because I would just want to give it away for free to everyone, immediately, all of the time, everywhere.” His answer was the practical one: hiring the people who build it takes revenue.
One in four customers, dollar weighted
Then I asked for the hardest day, and he told me about the year he thought he’d lose the company. Pindrop’s first customers were big banks, insurers and health care providers, and in 2011 and 2012 they wanted everything on their own premises, so until 2018 Pindrop was largely an on-premises company. When Amazon launched Amazon Connect, its cloud contact center, Vijay saw the cloud arriving and moved. The problem was the audio. Getting sensitive call recordings from a customer’s phone system into Pindrop’s cloud depended on carriers and on mechanisms that were still new, and customers didn’t trust the security of their calls in transit.
“So we had a churn rate of one in four customers, dollar weighted, leaving Pindrop.”
Vijay Balasubramaniyan, at 34:40
That was 2020, with the pandemic on top of it, and he did the math out loud: “Like, I thought we were going to die, right? Because if you think about it, right? If one in four customers that you sign up are leaving, you’ve essentially cut off any new thing coming.” I said it the way an investor would, “25% growth rate just to maintain flat, flat line,” and he finished the thought.
“Even if you’re growing at 25%, you’re growing at zero then, because you’ve lost all of those customers to churn.”
Vijay Balasubramaniyan, at 36:35
The way out was a huge investment in security. The audio had to be encrypted end to end, and the carriers had to build encrypted links between the customer and the cloud, because otherwise “I don’t want some third-party box sniffing sensitive traffic on my networks, right?” Every call carries passwords, Social Security numbers and secret answers, so the encryption had to come from the right parties, the carriers and the contact center platforms. “And that required the entire industry to mature up.” When we spoke, Pindrop’s churn was 4 percent, against what he called an industry best practice of under 10.
This is the part of the episode I’d send to founders who think of security as a cost. Pindrop’s product worked. What customers left over was whether they could trust how their data got to it. Security was the churn, and security was the fix, and the calculator below puts your own numbers on the same math.
A more measured CEO
“But it was baptism by fire.” That was when Pindrop went through layoffs, and he described the worst day, before the pandemic had quite hit, when about 15 percent of the company had to go. He told his head of human resources that he would walk everyone out himself, and he did, over two days, every hour. Some people shouted and some cried.
“And the worst were people who didn’t say a word, because at Pindrop, we have one in four employees as PhDs.”
Vijay Balasubramaniyan, at 38:00
They were people who had never failed at anything, told they were out of a job for no fault of their own, and he kept telling them it wasn’t their fault. “It was the worst experience, but it, it taught me how to be a more measured CEO.” Several of them came back a year to 18 months later, he said, because of the promise he made on the way out: “I’m going to do everything to find you a job.” I told him the thing I believe about running a company, “a loyalty to people, not to companies,” and he agreed that a company is the people in it.
Out-hire, out-innovate, or out-execute
I asked what he’d tell a younger Vijay. He had a mental model ready.
“Like, one of the things I’ve realized is, most problems, you either out-hire, out-innovate, or out-execute your way out of a problem. And you have to decide which is it, right?”
Vijay Balasubramaniyan, at 41:49
His examples are the company’s history. Engineering buried in debt, without the right cloud skills, is a hiring problem, so hire the right head of engineering. A crazy competitor is an execution problem. A brand-new cloud where nobody has figured out how to move sensitive call audio in real time is an innovation problem, which is the hardest day above, solved. Early on, he said, he had no models at all, and would “let’s throw whatever on the wall and see what sticks.” The three-way question is a good one to keep on a sticky note. Most of the time, a problem only looks unsolvable until you’ve decided which kind it is.
Security for my kids
My last question was the one for the entrepreneurs listening: what would you pay for that nobody has built? Vijay didn’t hesitate. “Yeah, it’s security for my kids.” Social media and a browser in everything have opened horizons for his son in good ways and bad, and he wants a security company to help with the bad ones.
“I want him to be a kid first, and not have to grow up to be an adult too quickly.”
Vijay Balasubramaniyan, at 44:22
He’d looked at what exists, and his review was polite: “nothing’s great yet. Nothing hits the spot quite yet.” Two years later he brought parenting to a different problem. At TEDxAtlanta in October 2026, he talked about building trust between people and AI agents, and how it can borrow from “the oldest trust building mechanism that humans have: Parenting” (LinkedIn). The father who wanted his son to stay a kid a little longer now argues that raising agents with values is how we’ll learn to trust them.
What happened after the tape
Pindrop’s public record from the summer of 2024 to the start of November 2026, in order.
- July 17, 2024: $100 million in debt. Financing from Hercules Capital, with Pindrop saying it had analyzed 5.3 billion calls and detected 104 million spoofed ones (Help Net Security).
- October 30, 2024: this episode, the same day as the post from the New York Stock Exchange above.
- 2025: meetings. Meeting deepfake detection joined the contact center and audio products (Pindrop), and on November 19, 2025, Pulse for Meetings was named one of TIME’s Best Inventions of 2025 (Pindrop newsroom).
- November 26, 2025: retail. Vijay told Axios that three in ten attacks on major retailers were AI-driven.
AI-generated fraud is accelerating → In a new Axios piece, Pindrop CEO Vijay Balasubramaniyan reveals that 3 in 10 attacks on major retailers are now AI-driven, some seeing 1,000+ bot calls a day.
Pindrop (@pindrop) · November 26, 2025
Our team is building defenses to fight back.
- February 2026: more platforms. Fraud and deepfake defense inside NiCE CXone, and deepfake detection for health care (Pindrop newsroom).
- March 2026: Zoom, and investigations. Real-time deepfake detection and identity verification in Zoom Contact Center on March 12, and what Pindrop called the first agentic fraud investigation product on March 17 (Pindrop newsroom).
- April 2026: TIME. Pindrop made TIME’s list of the 10 most influential software companies of 2026, and TIME’s reason was that Pindrop’s technology “was used to authenticate footage of a real-world news event.” Pindrop’s own data in the same piece: AI attacks up 1,210 percent in 2025 (Pindrop).
- May 21, 2026: FICO. Pindrop’s fraud intelligence on the FICO Marketplace, with an 80 percent fraud detection rate and under 0.5 percent false positives across large institutions, and about $3.5 billion in fraud losses prevented for customers through December 2025, by Pindrop’s estimate (GlobeNewswire).
- August 19, 2026: a privacy ruling. Pindrop announced that the Third Circuit had ruled it qualifies as a financial institution under the Illinois Biometric Information Privacy Act (Pindrop newsroom).
- September 2026: agents on the phone. Pindrop says it detected calls from Meta’s Muse agent voices on September 9, a week before Meta announced its outbound calling beta.
Pindrop detected Meta's Muse AI agent before Meta even announced it.
Vijay Balasubramaniyan · October 10, 2026
On September 9, Pindrop detected the first calls from Muse's AI voices, Brett and Hailey, across our customers' contact centers. Meta didn't publicly announce its outbound calling beta until September 16.
This was effectively a zero-day detection. A new AI agent entered the ecosystem, and Pindrop was already identifying it without waiting for a public launch or a known threat signature.
And it wasn't an isolated event. Within four weeks, weekly Muse call volume had grown 8x. These agents were calling banks, healthcare organizations, and retailers, navigating IVRs and reaching human operators.
- September 16, 2026: BotStopper. A product that tells within about two seconds whether a caller is human or machine, and recognizes known AI agents through a registry of more than 5,000 AI voices. Pindrop said one Fortune 500 health care deployment cut bot activity by 94.3 percent in under four months, and that AI-powered attacks had risen nearly 1,680 percent since late 2024 across more than 700 million customer interactions (GlobeNewswire).
- September 28, 2026: a survey. Of 250 U.S. security leaders at companies with 1,000 or more employees, 74 percent had encountered or suspected a deepfake attack in the past year, and 10 percent had purpose-built tools for it, in a Wakefield Research poll for Pindrop (GlobeNewswire).
- October 2, 2026: TEDxAtlanta, on trust between people and AI agents (LinkedIn).
- October 28, 2026: CFX. Vijay’s keynote at Pindrop’s customer conference, and the same day, a post about two new voice models, Meta’s Muse and Block’s Dessa, both detected with “Zero changes to our detection model” (LinkedIn).
- October 29, 2026: the World Economic Forum. Pindrop joined as a Global Innovator, working with the Forum’s Centre for Cybersecurity (LinkedIn).
- November 2, 2026: hiring. Vijay shared what Pindrop sees in job applications across its customers, with Axios.
$6.3 billion. Stolen in just 20 minutes.
Vijay Balasubramaniyan · November 2, 2026
North Korean hackers are operating at an extraordinary speed and scale. Social engineering, automation, and increasingly accessible AI tools are fundamentally changing the economics of cybercrime.
But what concerns me most isn't just the size of this heist. It's how North Korea is industrializing cybercrime, and how the same tactics are being used to infiltrate organizations from within.
I spoke with Sam Sabin at Axios about this growing threat and what we're seeing firsthand at Pindrop.
🚨 Last year, 1 in every 343 job applications we analyzed across our customer base was linked to North Korea. This year, that number has jumped to 1 in 47. A whopping 7x increase!
Read in order, the record is the 2010 paper’s question asked of every new kind of caller. First it was which network a call crossed, then whether the voice was the account holder, then whether it was a person at all, and now which agent is calling, for whom, and what it should be allowed to do. Vijay’s posts this fall put it in three phrases, the Right Identity, the Right Intent and the Right Action (LinkedIn). That is a good checklist for any company whose phone rings.
Did you catch it? A six-question quiz
Interactive · six questions
Did you catch it?
Pick an answer for each, then check. Every answer is on the tape.
-
Besides a security person, what does Vijay call himself?
- A voice person
- A math person
- A product person
a. "I think of myself as a security person and a voice person." Transcript, 0:48.
-
What are the conference rooms at Pindrop named after?
- Pindrop's first customers
- Bad actors Pindrop has caught
- Famous phone hackers
b. "every conference room is named after a bad actor that we've caught." Transcript, 15:11.
-
Why is one fraudster called Giraffe Man?
- He only called at night
- He used a giraffe as his profile picture
- His audio could only have come from someone with a seven-foot neck
c. "the only person who could have produced this audio is someone with a seven-foot-long neck" Transcript, 15:11.
-
Whose tool did Pindrop find behind the Biden robocall?
- ElevenLabs
- PlayHT
- Lyrebird
a. "And we were the ones who discovered that it was ElevenLabs that was used to create Biden's deepfake." Transcript, 15:11.
-
What does Vijay say Pindrop's deepfake warranty works like?
- Cyber insurance
- A bug bounty program
- A money-back guarantee
b. "the warranty serves almost like a bug bounty program for us." Transcript, 22:40.
-
Vijay's mental model: most problems, you out-hire, out-innovate, or what?
- Outlast them
- Outspend them
- Out-execute your way out of them
c. "most problems, you either out-hire, out-innovate, or out-execute your way out of a problem." Transcript, 41:49.
Would a cloned voice get through your phone lines?
Most startups don’t run a contact center, and most of them still have a phone number that resets passwords, changes bank details or moves money for someone who sounds right. This scorecard asks eight questions about yours, from what the line reveals before anyone is verified to what happens when an AI agent calls for a customer.
Scorecard · eight questions
Would a cloned voice get through your phone lines?
Answer for every number a customer, an employee or a vendor can call, including your own help desk. If two answers fit, pick the lower one.
-
Before a caller is verified, what will your phone line confirm?
- Whether an account exists, and sometimes more
- Nothing by policy, though agents sometimes slip
- Nothing, and the system enforces it
This is how a stolen Social Security number finds its bank. Welcome back, John has Vijay's version.
-
How does someone on your line verify a caller?
- Personal details or security questions
- A one-time code by text message
- A prompt in your app, or a key on a registered device
There were 3,205 data compromises in the U.S. in 2023, and the answers to those questions are in them.
-
Can one phone call change where money or codes go: a new phone number, email or bank account, a password reset, a new sign-in device?
- Yes, on the same call
- Yes, after a callback or a waiting period
- No, those changes need the registered device or the app
That change is what a cloned voice is calling for.
-
Do you notice when one caller asks about many different accounts?
- No
- Only in a later review
- Yes, it's flagged while the calls are happening
Pindrop names its conference rooms after callers like that. Giraffe Man is one.
-
Does a matching caller ID skip any step?
- Yes
- It's one signal among several
- No, caller ID is a claim we check
Where a call really came from was the question in Pindrop's first paper.
-
Is anything listening for a synthetic or replayed voice?
- No
- Agents are told to listen for it
- Calls are scored for it, and the agent sees the score
Across 56 studies, people spot deepfakes about 55 percent of the time. Listening harder isn't a plan.
-
What happens when an AI agent calls on a customer's behalf?
- We wouldn't know it was an agent
- We treat it like any other caller
- We can tell, and a written rule says what an agent may do, and for whom
Vijay's three questions for agents: which one is calling, who it represents, and what it should be allowed to do.
-
When someone on your line flags a call, what happens next?
- They finish the call and move on
- They note it in the ticket
- Someone investigates the same day, and the customer gets a call from a number they know
That is the Saturday in Vijay's best day.
A line that confirms an account exists, verifies with personal details and lets one call change a phone number is the standard setup, and every part of it is fixable. Start with the first two: confirm nothing until the caller is verified, and verify with something they have, a prompt in your app or a registered device. Then move phone number, email, bank and password changes behind that same step.
The basics are in place. The next layer is noticing patterns: one caller asking about many accounts, a synthetic voice, an AI agent calling for a customer. Score calls for liveness, write the rule for agents, and send every flagged call to a person who looks the same day.
Your line verifies with something the caller has, keeps account changes off the phone, notices patterns and has a rule for agents. Rehearse it once with a cloned voice, with permission, the way Pindrop did with New Hampshire's attorney general, and write down what got through.
What does a trust problem cost your growth?
Vijay’s hardest day was arithmetic: when one in four dollars of customers leaves every year, 25 percent growth only keeps you where you are. This calculator runs the same math on your numbers, and shows what bringing churn down is worth.
Calculator
What does a trust problem cost your growth?
Your recurring revenue today, the new revenue you'll book this year, the share of revenue that leaves in a year, and the churn you could reach if customers trusted how you handle their data.
With 25 percent new revenue and 25 percent churn, you grow 0% this year: $1,250,000 of revenue leaves, and you finish the year at $5,000,000. At 4 percent churn, you'd keep $1,050,000 more every year.
The 25 percent default is Pindrop's dollar-weighted churn in 2020 by Vijay's account on the tape, and 4 percent is where he said it was when we recorded. His under-10-percent best practice is his figure, not a benchmark we checked, and the $5 million is a placeholder. By his account, what moved Pindrop's number was security its customers could trust: encryption from end to end, carriers and contact center platforms that matured with it, and a huge investment in getting it right.
From call provenance to agent calls
Everything on this page sits on one timeline, from the first paper to the AI agents on the phone. Filter it by thread.
Timeline · 2002 to 2026
From call provenance to agent calls
Filter by thread. Dates are from the public record except where marked as Vijay's own account on the tape.
- Intel
Software engineer, on voice for landline switches by his account (LinkedIn).
- The language of an encrypted call
Wright, Ballard, Monrose and Masson, 66 percent across 21 languages (USENIX).
- Telineage
A Georgia Tech VentureLab project with Mustaque Ahamad (Wikipedia).
- PinDr0p
Call provenance at ACM CCS, 91.6 percent through as many as three providers (Business Wire).
- Pindrop
Co-founded in Atlanta with Paul Judge and Mustaque Ahamad (LinkedIn).
- Words from an encrypted call
Hookt on fon-iks, with a linguist among the authors (IEEE).
- About $1 million
The first round, from Andreessen Horowitz and others (Wikipedia). Six months to raise, by his account.
- An $11 million Series A
Led by Andreessen Horowitz and Citi Ventures (Wikipedia).
- $35 million
A Series B led by Institutional Venture Partners (Wikipedia).
- The first deepfake patents
Years of deepfake work before it was hot.
- $75 million
A Series C led by Google Capital (Wikipedia).
- $90 million in six weeks
A Series D led by Vitruvian Partners, with eight of the ten largest U.S. banks as customers (Pulse 2.0).
- One in four customers leave
Dollar weighted, until call audio could reach the cloud securely.
- Test of Time
ACM CCS honors the PinDr0p paper (Business Wire).
- Roadrunner
Three lines in Anthony Bourdain's voice were generated with AI (NPR).
- Aaron Painter on this show
Nametag, on redefining authentication (episode 19).
- A robocall in Biden's voice
To New Hampshire Democrats before the January 23 primary (NBC News).
- ElevenLabs, 99 percent likely
Pindrop's analysis against 122 text-to-speech systems (Pindrop).
- 3,205 data compromises
A 78 percent jump in 2023, a record (Cybersecurity Dive).
- Traced to Texas
Life Corporation and Lingo Telecom (Engadget).
- 81 of 84
NPR's test of three deepfake detectors (NPR).
- The report, and a warranty
60 percent more contact center fraud in two years, and the Pulse Deepfake Warranty (MarTech Series).
- $100 million in debt
From Hercules Capital (Help Net Security).
- Episode 57 airs
This conversation, the same day as the post from the New York Stock Exchange.
- Ben Colman on this show
Reality Defender, on fighting deepfake fraud (episode 71).
- A TIME best invention
Pulse for Meetings (Pindrop newsroom).
- Zoom
Deepfake detection in Zoom Contact Center (Pindrop newsroom).
- Jasson Casey on this show
Beyond Identity, on why the real problem is identity (episode 92).
- One of TIME's ten
The most influential software companies of 2026 (Pindrop).
- Muse, a week early
Calls from Meta's agent voices before the announcement (LinkedIn).
- BotStopper
Human or machine in about two seconds (GlobeNewswire).
- 74 percent
Of security leaders encountered or suspected a deepfake attack in a year (GlobeNewswire).
- TEDxAtlanta
Trust between people and AI agents, borrowed from parenting (LinkedIn).
- 1 in 47
Job applications linked to North Korea across Pindrop's customers, up from 1 in 343 (LinkedIn).
What to do this week
Five things, in the order the episode suggests.
- List every decision a voice can make at your company. Password and sign-in resets at the help desk, wires, vendor bank changes, payroll changes, access granted on a call. That list is what a cloned voice is shopping for.
- Confirm nothing until the caller is verified. Not that an account exists, not a balance, not the last four digits of anything. That’s how a stolen Social Security number becomes a taken-over account.
- Verify with something the caller has. A prompt in your app or a key on a registered device, instead of questions whose answers are sitting in someone else’s breach.
- Write the rule for AI agents that call you. Which agent, on whose behalf, allowed to do what, and which requests still need a person.
- Rehearse it once. With permission, clone a willing executive’s voice and call your own line, then fix what got through. When Pindrop tried it with New Hampshire’s attorney general, 40 percent of the room guessed wrong.