A minute into this conversation, before I had asked a real question, Deepak Dutt brought up a video call.
“I remember there was a breach not too long ago. It involved two hundred million dollars in a transfer and a fake Zoom call and a deepfake, and there it goes. We started off the journey a long time ago, but I think what’s happened now is, you know, the same tools are available to everybody. And the barriers to entry have been lowered so much that now you can have AI-powered attacks at scale.”
Deepak Dutt, founder and chief executive of Zighra, at 1:03
The call was real. In February 2024, Hong Kong police described a finance worker who had joined a video meeting with what looked like his chief financial officer and several colleagues, all of them deepfakes, and then sent HK$200 million, about US$25.6 million, in fifteen transfers (CNN). Three months later the engineering firm Arup confirmed it was the victim (CNN, via ABC 17). Deepak’s “two hundred million” is the Hong Kong dollar figure. The point he was making survives the currency conversion: the person on the call had every signal a human uses to decide who they are dealing with, and every one of those signals was fake.
That is the whole episode in one exchange. Deepak has spent more than a decade on a different question than “does this person look right?” His question is “does this person still behave like themselves?”, asked continuously, answered by the device in your hand, without the data ever leaving it. This page is what I wish I had known before we hit record: what his approach actually does, where it sits next to passkeys and device-bound credentials, what has happened in the thirteen months since we spoke, and the numbers behind every claim on the tape. There are four things to try along the way: a quiz on the episode, a scorecard for your own sign-in setup, a timeline you can filter, and a small calculator for the friction you are paying right now.
The call that cost HK$200 million
The Arup case is worth sitting with, because it is the cleanest example of the thing every identity product is now up against. The employee received an email asking for a secret transaction and suspected phishing, which is the right instinct. Then he joined a video call, saw faces and heard voices he recognized, and dropped the doubt (CNN). Every check that a reasonable person would run came back clean, because every check was about appearance.
It is no longer a one-off. The FBI’s Internet Crime Complaint Center counted $20.877 billion in reported losses for 2025 across more than a million complaints, with 22,364 complaints carrying an AI component and $893 million in adjusted losses, and it noted that “voice cloning is now being layered into these attacks” as follow-up calls that impersonate executives (SecureWorld, summarizing the IC3 report). Business email compromise alone was $3.046 billion. The Hong Kong call is the business-email-compromise playbook with a video layer on top.
Arup lost $25mn in Hong Kong deepfake video conference scam via @FT. UK engineering group Arup lost HK$200mn ($25mn) after fraudsters used a digitally cloned version of a senior manager to order financial transfers during a video conference
David Zaikin (@DavidZaikin) · May 17, 2024
When the story hit Hacker News, the thread ran to 339 comments, and most of the useful ones were about process rather than technology: agree a challenge phrase with your executives, require more than one channel for a sum that size, double-check every order (Hacker News). Those are good rules. They are also rules that depend on a stressed person following them on a Thursday afternoon with a convincing boss on the screen. We had Ben Colman of Reality Defender on episode 71 on detecting the fakes, and Jasson Casey of Beyond Identity on episode 92 arguing that detection is a losing race and the fix is cryptographic proof of who is on the line. I wrote up that argument in Deepfake attacks keep working because we keep detecting instead of proving. Deepak’s answer sits between the two: don’t ask whether the face is real, ask whether the hands on the device behave like the person who owns it.
A gate or a flow
Here is how he described the founding idea, when I asked what started the company.
“Starting early on, our biggest thing was how do you detect who’s the right person behind a device or a transaction, in a very transparent fashion, without having to go through hoops to prove who you are. And we said, can we do that in a very invisible fashion, without bringing in all the friction involved with security?”
Deepak Dutt at 1:40
Almost everything we call authentication is a gate. You arrive, you prove something once, the gate opens, and from that moment the session is trusted on the strength of that single moment. A password is a gate. A one-time code is a gate. A passkey is a much better gate. None of them have anything to say about who is holding the device twenty minutes later, which is exactly the moment a stolen session cookie, a borrowed laptop, or a help-desk reset gets used.
I tried to summarize it back to him on tape as “passive, adaptive, dynamic… That’s continuous. It’s not like a point in time where you do an action, you have a proof,” and he took it one step further: the score is tied to a specific application. Your banking app, your payments app, or, in the work Zighra does now, a military one. “How do you know it’s still the right user continuously using it, without them having to go through all kinds of hoops?”
The first public test of the idea is the detail I keep repeating to people. Zighra built an AI-powered lock screen, put it on the Google Play Store when Android still let an app replace the lock screen, and watched.
“So we created an AI-powered lock screen, put it on the Google Play Store, and within five weeks we had around twenty-five thousand downloads across seventy different countries, on seven hundred different device types.”
Deepak Dutt at 1:40
Seven hundred device types is the number that matters for a founder. It means the model had to work on sensors it had never seen, on phones with wildly different accelerometers and touchscreens, with no central dataset to fall back on. That constraint, chosen in the first product, is the reason the same software could later run in places with no network at all. We will get to that.
The field has a name, behavioral biometrics, and it got its biggest validation ten months after we recorded. On August 3, 2026, Visa agreed to buy BioCatch, the best-known company in the category, for $2.4 billion in cash. BioCatch analyzes about 18 billion user sessions a month for more than 350 financial institutions, protecting some 760 million users, and says it prevented $4 billion in fraud in 2025 (Finovate). Visa’s rationale, in the words of its value-added services president Andrew Torre, is to stop fraud “before it reaches the point of payment” (bne IntelliNews). That is Deepak’s thesis, bought by a card network: the signal that matters is upstream of the transaction, in how the session is being handled.
🚨 BREAKING: Visa is buying BioCatch for $2.4bn in cash. BioCatch says it watches 19 billion banking sessions a month. As a behavioral biometrics company, it scores how you type, swipe, and hold your phone. It catches account takeover, fake applications, mule accounts, and people being coached through a transfer by a stranger on the line. When fraud and scams are exploding, the ability to see upstream of the payment is key. What were the keystrokes? What was happening?
Simon Taylor (@sytaylor) · August 3, 2026
Two caveats, so this reads as analysis and not a brochure. First, behavioral biometrics has had a privacy problem since banks started quietly deploying it. When The New York Times reported in 2018 that banks and retailers were tracking how people type, swipe and tap, the Hacker News thread was mostly about consent, and one commenter raised the attack Deepak’s design has to beat: malware on a compromised device could record a person’s behavioral data and replay it (Hacker News). Deepak’s answer to the first concern is in the privacy section below. His answer to the second is that the model keeps changing, so a replay goes stale, which is a good answer and not a complete one. Second, the big deployments are at banks with hundreds of millions of users. A thirty-person startup is not going to license BioCatch next quarter. The useful question for you is which of these ideas you can borrow with the tools you already have, and that is what the scorecard further down is for.
What the phone already knows
Deepak listed the signals on tape without being asked, which tells you how long he has been explaining this.
“So we looked at how you deal with your devices: mobile devices, laptop, desktop, or any sensor-based device. And we took a look at how you hold the device, the angle you hold it, the pressure you apply on the touchscreen, and the acceleration with which you move, because all these sensors are used for different purposes by applications, be it for gaming or be it for locating where you are. And we take a look at that and look at unique patterns that people have.”
Deepak Dutt at 1:40
On Forcepoint’s To The Point podcast in March 2025, six months before ours, he went a level deeper: Zighra runs a fusion engine that combines a device engine, a contextual engine, a network engine, a bio-behavioral engine and a social engine, so that an attacker has to defeat several layers at once, and the on-device model can run continuously for a couple of days on about 1% of battery, producing a score in milliseconds (Forcepoint). The 1% figure is his claim, not a measurement I can check, but it answers the first objection every mobile engineer raises.
The part I find most useful for founders is what the application receives. Not a face, not a fingerprint, not the raw motion data. A score, and the factors behind the score. That shape, a continuous confidence number that your own code can act on, is something you can build toward with ordinary tools: device posture, session binding, a step-up check when the risk changes. The phone already knows. Most products just never ask.
The credential pile got bigger while nobody looked
Deepak dated the shift to the moment the password breaches stopped being countable.
“So back in 2016, when we looked at this site called haveibeenpwned.com, I think there was maybe 10 million passwords or something like that that was breached. Today, when you look at it, billions and billions of passwords. So that entire space has kind of changed.”
Deepak Dutt at 5:38
His 2016 number is a memory, and Have I Been Pwned was already well past ten million by then, so treat it as a feeling rather than a statistic. The feeling is right. In November 2025 alone, Troy Hunt loaded about two billion email addresses and 1.3 billion passwords from one dataset supplied by the firm Synthient (GIGAZINE). A month later he was helping the FBI process another 630 million passwords seized from a suspect.
We're very happy to help the @FBI with processing another 630M passwords recently seized from a suspect. This includes 46M we'd never seen before, which are now all in @haveibeenpwned's Pwned Passwords, where they're queried 18 *billion* times a month:
Troy Hunt (@troyhunt) · December 12, 2025
The number on that chart that should change how you build is not the passwords. It is the 8.6 billion stolen cookies and session artifacts SpyCloud recaptured in 2025, alongside 642 million credentials from 13.2 million infostealer infections and 18.1 million exposed API keys and tokens (SpyCloud). We had SpyCloud’s Trevor Hilligoss on episode 86 explaining why almost none of those infostealer infections was aimed at anyone. A stolen session cookie walks straight past your login gate, however good the gate is, because the gate already opened for the real user an hour ago. That is the dashed square on the first figure, and it is the case continuous authentication was built for.
Microsoft’s view of the same year: identity-based attacks rose 32% in the first half of 2025, and more than 97% of them were password attacks, mostly bulk attempts with leaked credentials, with help-desk impersonation scams rising alongside (The Record). The help-desk path is the one that worries me most for startups, because it defeats multi-factor authentication without touching it.
ALERT—The FBI has recently observed the cybercriminal group Scattered Spider expanding its targeting to include the airline sector. These actors rely on social engineering techniques, often impersonating employees or contractors to deceive IT help desks into granting access. These techniques frequently involve methods to bypass multi-factor authentication (MFA), such as convincing help desk services to add unauthorized MFA devices to compromised accounts.
FBI (@FBI) · June 27, 2025
Read that alert next to Deepak’s design and the point lands. A newly enrolled MFA device is a new gate. A behavioral model that has watched the real employee for months would see the new device being used by somebody who holds it differently, and the score would drop. I am not saying that would have stopped every Scattered Spider intrusion. I am saying it is the only layer in that story that looks at the person rather than the credential. If your own questionnaire answers say “MFA enabled” and you have never checked what can still sign in without it, read Enabled isn’t enforced before your next deal, and read the case for device-bound credentials for why the secret moving at all is the root problem.
”We are not passwords,” a year later
This is the line from the episode that people quote back to me, and for the record, it is mine, not Deepak’s.
“I believe that we are not passwords. We can do better. I would love a piece of technology that matched all of the great advancements that we’ve had along the way. And so I think, does this need to exist? This absolutely needs to exist, because we are not passwords.”
Jon McLachlan at 15:46
Deepak’s reply was more careful: “I think we are constantly hearing the days of passwords are over. I think people are on the journey. We’ll see. But every year we are hearing more and more passwords getting compromised.” A year later the data says he was right to hedge, and right about the direction.
The FIDO Alliance’s Passkey Index for 2026, published October 15, reports that 36% of successful sign-ins at fifteen participating consumer services used a passkey, up from 26% a year earlier, that passkey sign-ins succeed 93% of the time against 78% for passwords and one-time codes, and that they are nearly three times faster, saving about 27 seconds each (FIDO Alliance). FIDO’s estimate of passkeys in use went from 5 billion in May to 7 billion in October (FIDO Alliance). On the workforce side, 68% of organizations surveyed in April 2026 had deployed or were deploying passkeys for employees, and 57% still relied on a phishable method for day-to-day sign-in (FIDO Alliance). Those two numbers are both true of the same companies.
The part FIDO’s press releases don’t dwell on is that a lot of people hate the experience, and they said so in public this summer.
Passkeys were invented by security engineers with zero understanding of the consumer brain. We had no way to argue its merits since no one understood what they are. As a result, we are now all asked to produce the magic fairy dust to login to apps. It may be on your phone, in your browser, in your operating system, in your body, who knows. But now you're safe.
Nikita Bier (@nikitabier) · July 22, 2026
"The entire setup/enrollment/add a passkey to your account processing is DMV inspired."
IceHegel, on Hacker News, July 22, 2026
"Passkeys are a vector for locking your logins to Big Tech ecosystems."
coldpie, on Hacker News, July 22, 2026
Both complaints are about friction and control, which is exactly the ground Deepak was fighting on in 2013. His whole pitch is that the best authentication is the kind you never notice. I think passkeys and continuous behavioral scoring are allies rather than rivals, and the honest way to put it is this: a passkey fixes the moment of login, which is where phishing lives. It does nothing for the hours after login, which is where stolen sessions, shared laptops and help-desk resets live. A behavioral score covers the hours. Neither one asks you to remember anything.
Taher Elgamal, who designed SSL, made the longer version of this argument on episode 61: passwords were a shortcut taken in the 1990s to keep the internet moving. I said something similar to Deepak near the end of this episode, that a shared secret was the right trade at the time and that “now we have a different problem on our hands, because everyone’s on the internet.” His own version of the regret is sharper, and it comes up in the founder notes below.
Explainable, small data, and the bias question
I asked the question I ask every AI founder now: how do you secure the model, and where does its data come from? Deepak’s answer was that Zighra wrote its own stack rather than fine-tuning somebody else’s, for one reason.
“See, when we started up with the AI space, we kind of wanted to build our entire AI stack from the ground up, because the core tenet was: how do you make it explainable, day one? Because we didn’t like the fact that it was a black box, and we didn’t really understand what was going on inside.”
Deepak Dutt at 11:50
The reason is commercial as much as ethical. His customers are banks and defense departments, and “they don’t want to make a decision that says, it’s a black box, they made a decision. How do you stand behind that decision? So if you get challenged, you need to be able to explain, tell this to your citizens, what you’re actually doing.” If you sell into regulated buyers, that sentence is your next security questionnaire. The buyer will not ask whether your model is accurate. They will ask whether you can explain a single decision it made about one of their customers, and whether you can show it treats people fairly.
Which led to the question I was most curious about. Military user bases skew heavily male. Does a behavioral model trained in that world work as well for women?
“Absolutely. See, the key for us is to work with small amounts of data. So when you, let’s say, take the phone and start interacting with it, we’re building the model on the fly. It’s not like we have trained it on…”
Deepak Dutt at 12:54
I finished the sentence for him, “because you’re looking at doing anomaly detection,” and he agreed three times. Each model is built for one person from their own use, within about two minutes, and then keeps adapting. There is no population model to inherit a skew from, because there is no population. “Whatever they bring to the table, whatever their experiences, however they hold the phone, whatever their tics are, can be baselined, essentially.”
That is a cleaner answer to the bias question than most AI companies can give, and it is worth understanding why it works. A classifier trained on a crowd learns what the crowd looks like and flags people who don’t fit. An anomaly detector trained on you learns what you look like and flags moments that don’t fit. The second one cannot be biased against a group, because it has never met a group. It can still be wrong about an individual, on a day they have a broken wrist or a new phone, which is why the score has to be a question and not a verdict, and why Deepak keeps saying the word “adaptive.” If you are building any product that scores people, the authorization post is about where those scores should and should not be allowed to act.
Privacy that was designed in
The sentence I would put on the wall of every startup that touches sensor data:
“So I think it’s about thinking through the technology, and being very empathetic about your approach to building things. So then privacy is not something that you bolt on.”
Deepak Dutt at 16:35
He was specific about what the data would give away: “the sensor data is highly, highly sensitive, because it tells you if you’re sitting, standing, walking, where are you” at a given point in time. I made the joke that accelerometer data might end up treated as personal information under European privacy law, and it is less of a joke than it sounds. Under the General Data Protection Regulation, biometric data processed to uniquely identify a person is a special category that needs an explicit legal basis, and a behavioral profile used to identify you is biometric data by definition. Zighra’s answer is architectural rather than legal: keep the raw data on the device, learn there, infer there, and let only a score move.
The engineering cost of that choice is real, and Deepak did not pretend otherwise. “It has to be highly thermally efficient to be able to run entirely at the edge, on the device. So how do you learn at the edge? How do you infer at the edge, without having any of that sensor data coming back onto the cloud?” The failure mode he described is the one I see in most AI products: build in the cloud first, then try to push the model to the device later and find it is too heavy. It has to run on “a $50 phone to a thousand-dollar phone plus,” and the sensors on those phones are not the same. When I suggested the mobile device would probably catch fire, he did not disagree.
Here is the part founders should underline. The privacy decision was made in 2013 for reasons of trust, and it became a sales requirement a decade later for reasons of war.
“Can you work in highly contested environments, GPS-denied environments, cloud-denied environments? Can you work in these kinds of scenarios? And you can say, yeah, absolutely.”
Deepak Dutt at 20:35
A product that never needed the cloud to make a decision could be dropped into a defense environment where the cloud is not available on purpose. Nobody planned that in 2013. Designing for the user’s privacy produced a product that survives an adversary’s jamming. I cannot think of a better argument for doing the harder thing early, and it is the same argument I make to AI startups about where their models run and whose data they can see: a cheap decision now is an expensive migration later, and the buyers who matter will ask.
From phones to the electromagnetic spectrum
The turn in the conversation I did not expect was from touchscreens to satellites.
“But in the defense context, now the electromagnetic spectrum is a huge space, because these are all invisible spaces where there could be drones, there could be jamming happening, there could be spoofing happening, but these are all driven by sensors. It’s a world driven by sensors.”
Deepak Dutt at 7:11
The logic is the same as the phone. Model what normal looks like, then flag the deviation: “Somebody’s jamming me. Am I losing positioning because of that jamming? Or somebody’s spoofing my location, and suddenly it shows me like I’m somewhere else.” And then the line from this episode that I have quoted most since:
“Today’s wars are initially started off with cybersecurity and electronic warfare, because you can do this from afar.”
Deepak Dutt at 8:38
The scale of the jamming problem is easy to underestimate from California. According to a filing Baltic states made to the International Civil Aviation Organization and reported by Sweden’s SVT, nearly 123,000 flights by 365 airlines were affected by GPS disruption over the Baltic Sea region between January and April 2025, with 27.4% of flights in the region disrupted in April alone (ERR, citing SVT). Deepak’s version of the barrier to entry: “Tools are available today for $50. You could buy, and you could start jamming the satellite signals.”
Spoofing is the subtler cousin. In January 2026, the researcher Nariman Gharib published telemetry from a Starlink terminal inside Iran during an internet shutdown that showed the dish detecting eighteen GPS satellites with a valid lock and still switching on its anti-spoofing countermeasures, because the signals were fake.
Exclusive: Obtained Starlink terminal debug data from Iran during the ongoing internet shutdown. The telemetry shows direct evidence of GPS spoofing: the dish detected 18 GPS satellites with valid signal lock, but activated its anti-spoofing countermeasures (inhibitGps: true). This isn't simple jamming; the government appears to be broadcasting fake GPS signals to confuse terminals. The impact: 20%+ sustained packet loss, connection never stabilized after 24 minutes, bandwidth restricted, and beam pointing ~1° off target.
Nariman Gharib (@NarimanGharib) · January 14, 2026
When we recorded, Zighra’s defense work was already its main business. “More recently, over the last few years, it’s been primarily in the defense space.” The public record since then: in December 2022 the company had an Innovative Solutions Canada contract to test a sensor identity network with Royal Canadian Navy partners (PR Newswire). In May 2024 Shared Services Canada contracted its DeepSense product for continuous identity validation testing (Biometric Update). On June 2, 2026, the Department of National Defence announced it would test IntentHunter, Zighra’s intent-inference platform for counter-drone work, which fuses radio-frequency, optical and radar sensors and infers what a drone is trying to do from its behavior rather than matching it against a library of known signatures (GlobeNewswire via BNN Bloomberg). Deepak’s line in that release: “Drone warfare has changed the problem.” Then on July 23, 2026, Canada launched the Defence Drone Initiative, a supply arrangement meant to speed up uncrewed and counter-drone systems for the Canadian Armed Forces and Coast Guard (National Defence), and by September Zighra said it had qualified across all five streams of the marketplace.
Zighra has qualified across all five streams of Canada's Defence Drone Initiative (DDI) Marketplace.
Zighra · Company post · September 2026
The five streams span uncrewed and Counter-UAS systems, C2/communications/data/EW/cyber, systems integration, testing and evaluation, and innovation and experimentation.
For us, the breadth matters. Because the future of uncrewed and Counter-UAS capability will not be solved by any one sensor, platform or system.
If the defense market interests you, we have covered it from the founder’s side: Garrett Smith of Reveal Technology on episode 90, the guide to building a defense tech startup, and how counter-drone technology works and where the openings are. The compliance side, ITAR and CMMC Level 2, is where most founders lose a year, and it is work we do.
Protect the human endpoints
I asked Deepak to pick a point in the future and name the hardest security problem waiting there. He did not pick a technology.
“But at the end of the day, I mean, how do you protect every single person in an invisible fashion, right? I think otherwise humans become that core weak point where the attacks can permeate from. So I think in the future, I think we need to have these agents that are protecting every single person, avoiding them to become the starting points of new kinds of cyber attacks.”
Deepak Dutt at 24:30
He framed the shift as back end to front end: “we currently think about protecting the back end, data and the servers and stuff. But I think the challenge would be to protect the endpoints, not the… the human endpoints.” The data agrees with him. Verizon’s 2026 Data Breach Investigations Report put the human element in 62% of breaches, with credential abuse appearing somewhere in 39% of them even as vulnerability exploitation became the most common way in (Abnormal AI, summarizing the 2026 DBIR). The Hong Kong call, the Scattered Spider help-desk calls, the infostealer on a developer’s laptop: each one goes through a person, and each one looks legitimate to every system that only checks credentials.
What I did not expect is where Deepak took the idea a year later. His recent writing is about agents rather than people: a legitimate AI agent, with valid credentials and real permission, following an instruction planted in content it read, and buying something its customer never asked for. His framing is four questions, identity, authority, behavior and intent, and the observation that “every individual check can pass while the activity serves someone else’s objective” (Deepak Dutt on LinkedIn). That is the exact failure Jesse Anglen described on episode 106, seen from the fraud team’s side of the glass, and it is why an inventory of what your agents can touch and a named human who signs for each one are not paperwork. Behavior is the only signal left when the credentials are real and the instruction is not.
The founder notes
This is a founder show, and Deepak gave the most useful answers on the business in a while. In order:
Profitable for four years. “We’ve raised a couple of rounds of funding early on, and over the last four years, I think we’ve been profitable.” The public trail: a FinovateFall debut in 2013, a Barclays Techstars accelerator in 2016 that led to a Barclays pilot, a $1 million seed round in December 2016 led by the Investment Accelerator Fund, Export Development Canada and the Business Development Bank of Canada, with six employees at the time (Finovate). Then, in his words, “we were lucky enough to get to a stage where we could self-fund the company from that point onwards.” His explanation for how a small team stays profitable in a market with hundreds of millions of users is that they threaded AI into their own operations early: “Everything that we do is highly efficient.” On tape I called it “the old-fashioned building-a-business thing, where you operate in the green.”
Patents took five years. “It took us five years plus to get our first patents.” They came with a co-founder he met at school who is “a co-author on every single patent that I’ve put together, except one,” an academic who brought the mathematics and a philosophy of applying how biological systems self-heal to security. By 2021 Zighra had sixteen patents issued or pending, more than 100 forward citations, and its citations appearing in applications by Google, Mastercard, PayPal, IBM and Microsoft (Biometric Update). Those citations were the signal that pushed him from research to commercialization: “We have all these large companies that are looking at our patents and making it part of their forward citations. So we said, must be something very, very, very interesting.”
The proudest days were people days. First customer, first patent, first defense customer. “But at the end of the day, it’s about bringing the right people around… It’s very hard problems that we are trying to solve, and the technological risks are extremely high.”
The culture test is one question.
“So when we build something, the first question we ask is: does it need to exist? Sure, if it does need to exist, great. If not, then maybe not us.”
Deepak Dutt at 15:04
And the daily version, for every task on the board: “are there tasks there that you’re working on that don’t have any impact? If it doesn’t, maybe it shouldn’t be there.”
The hardest problem is distribution, not the model. Hiring is the hardest day-to-day problem. Running on very lightweight devices is the hardest technical one. But the one he says he is still solving is where a behavioral biometric belongs: “it can come in at the device level, it could go into the chip level, it could go into the operating system level, it can go into the application level.” Applications are fastest, through an SDK, and every application is turning into a payments application, so the surface keeps growing. “That would be our biggest challenge: get this in the hands of as many people as possible.” If you are a deep tech founder nodding at this, what is deep tech and building a cybersecurity startup are the two posts we point people to.
Advice to a younger self: secure the inflection points.
“Or you go back and say, from a security cap on, you say, shouldn’t have used passwords day one. So that’s why I said you could take each inflection point of technology and say, this is the way you should do it. Well, when they put together the GPS satellites, we should have put more security into them, so that these signals could be secured.”
Deepak Dutt at 27:54
“Now we’re solving the symptoms,” he said. The civilian GPS signal was designed in the 1970s with no authentication, which is why a $50 transmitter can lie to a $100 million aircraft. Kevin Kane made a parallel argument about encryption on episode 96: a post-quantum algorithm doesn’t make a product quantum safe unless the construction around it is done correctly.
The problem he would pay someone else to solve: traffic. “It’s a problem that is omnipresent everywhere in the world… So the journey is all about the user experience and making things frictionless. Traffic is one of those friction points.” It is the same instinct as the lock screen.
Did you catch it? A six-question quiz
Interactive · six questions
Did you catch it?
Six questions on what Deepak actually said, with the timestamp for each answer. The answers are under each question for anyone reading without JavaScript.
-
How long does Deepak say Zighra's on-device model needs before it has learned a new person?
- About two weeks of normal use
- One full working day
- About two minutes
c. "We learn extremely fast, within two minutes. That's how fast we can learn. And then we adaptively learn beyond that." Transcript, 13:08.
-
Who said "we are not passwords" on this episode?
- Deepak Dutt
- Jon McLachlan
- Both of them, in the same breath
b. It was Jon: "I believe that we are not passwords. We can do better." Deepak's reply was more careful: "the days of passwords are over. I think people are on the journey. We'll see." Transcript, 15:46.
-
Where does Zighra keep the raw sensor data it learns from?
- On the device. Only a score leaves it.
- In Zighra's cloud, encrypted
- In the customer's cloud, under their keys
a. "How do you infer at the edge, without having any of that sensor data coming back onto the cloud? That's a thought process. That's the philosophy that you've got to implement day one." Transcript, 18:27.
-
Deepak says a $50 purchase can do what?
- Clone a voice well enough to fool a bank
- Buy a month of a rented botnet
- Jam satellite signals
c. "Tools are available today for $50. You could buy, and you could start jamming the satellite signals." Transcript, 24:30.
-
What was the actual sum in the deepfake video-call case Deepak opens with?
- US$200 million
- HK$200 million, about US$25.6 million
- HK$25 million
b. Deepak remembered "two hundred million." Hong Kong police put it at HK$200 million, about US$25.6 million, in fifteen transfers. The company was Arup. See the first section.
-
Which of these did Deepak name as the hardest ongoing challenge for Zighra?
- Distribution: chip, operating system, device or app
- Raising the next round
- Explaining the model to regulators
a. "Then distribution becomes an extremely challenging kind of prospect... That would be our biggest challenge: get this in the hands of as many people as possible." Transcript, 21:22.
Would your sign-in survive this episode?
Most of what Deepak described is out of reach for a thirty-person company this quarter. The principles are not. This scorecard asks eight questions about your own setup, each one tied to a point on the tape, and tells you which of the three gaps an attacker would use first.
Scorecard · eight questions
Would your sign-in survive this episode?
Answer for the company, not the ideal. Weighted toward the attacks in this episode: the stolen session, the help-desk reset, the convincing video call.
-
Can anyone sign in to your main systems with a password plus a code sent by text or email?
- Yes, that's our normal path
- Only for a few accounts or fallbacks
- No, every path is phishing-resistant
Microsoft counts more than 97% of identity attacks as password attacks. A code by text is a gate a phishing page can hold open. Enabled isn't enforced has the ten-minute check.
-
If someone steals a session cookie from a laptop, does it work from another machine?
- Yes, until it expires
- For some apps, not others
- No, sessions are bound to the device
SpyCloud recaptured 8.6 billion stolen cookies and session artifacts in 2025. The gate already opened. Device-bound credentials explains the fix.
-
Can your help desk reset a login or add a new MFA device after a phone call?
- Yes, a convincing caller gets it done
- Only with a manager's say-so
- Only after a verified in-person or video step with a known device
The FBI's June 2025 alert on Scattered Spider: impersonate an employee, talk the help desk into adding an MFA device. No malware required.
-
Does anything notice when a logged-in session starts behaving differently: new device, new location, new hours, an unusual download?
- Nothing, a session is a session
- Some alerts, nobody owns them
- Yes, with a step-up check or a page to a person
This is the whole episode: the score after login. You do not need behavioral biometrics to start. You need one signal, one threshold and one owner. Monitoring and response is how we staff it.
-
For a wire, a payroll change or a vendor bank-detail change, is there a verification step that a video call cannot satisfy?
- No, a call from the boss is enough
- A call-back to a known number
- A second channel plus a second person, written down
The Hong Kong employee saw faces and heard voices he knew. The only defense that survived was a rule that did not depend on recognizing anyone.
-
Do your AI agents and service accounts have their own credentials, with a named human who owns each one?
- They share keys and nobody owns them
- Separate keys, unclear owners
- Separate, scoped, owned and inventoried
Deepak's 2026 framing: a legitimate agent with valid credentials following somebody else's instruction. The agent credential inventory is the starting point.
-
If a regulator or a customer asked why your system blocked or allowed one specific login, could you show them?
- No, the vendor's model is a black box
- Partly, with logs and a vendor ticket
- Yes, per decision, with the factors
"How do you stand behind that decision?" Deepak built his own stack to answer it. Your buyers will ask you the same thing about yours.
-
Does any raw behavioral, location or sensor data about your users leave their device for a model you run?
- Yes, and we have not mapped where it goes
- Yes, mapped and minimized
- No, or only derived scores
"Privacy is not something that you bolt on." Behavioral data used to identify someone is biometric data under European privacy law.
A phishing page, a stolen cookie or a help-desk call gets in, and nothing after login would notice. Fix the top two questions first: phishing-resistant sign-in everywhere, and sessions that die when they leave the device. Then write the wire rule down. Those three close most of this episode.
Login is in decent shape. The hours after it are not watched, and the human paths, help desk and finance, still trust a familiar voice. Pick one post-login signal, give it a threshold and an owner, and make the help desk's reset flow require something a caller cannot fake.
Phishing-resistant sign-in, bound sessions, a wire rule that does not depend on faces, and someone watching behavior after login. What is left is the explainability question and the data-residency question, which are exactly what your enterprise buyers will ask next.
The friction you are paying today
Deepak’s lock screen was built on a bet that nobody would accept more friction for more security. The FIDO numbers now put a figure on the friction: a passkey sign-in saves about 27 seconds against a password plus a code, and succeeds 93% of the time against 78% (FIDO Alliance). Put your own team into that arithmetic.
Calculator
The friction you are paying today
Hours a year your company spends at the gate, and what a sign-in that just works would give back. Rough by design. Change the numbers to yours.
At these numbers your company spends hours a year proving who it is. Deepak's point is that the better version is also the safer one.
Working days are counted as 250. The 27 seconds saved is the FIDO Passkey Index 2026 figure for consumer services, and the 45-second default follows from its finding that passkey sign-ins are nearly three times faster. Your workforce numbers will differ, which is the point of the sliders. Failed sign-ins and resets are not counted, so the real number is higher.
Thirteen years on one line
Everything on this page sits on one timeline, and filtering it is the fastest way to see how a privacy decision in 2013 became a defense contract in 2026 while the attacks caught up with everybody else.
Timeline · 2013 to 2026
Thirteen years on one line
Filter by thread. Dates are from the public record except where marked as Deepak's own account on the tape.
- KineticID at FinovateFall
Zighra demonstrates its first product, a mobile authentication system that evaluates how a person physically interacts with the device (Finovate).
- "Twenty fourteen."
Asked how long he has been doing this, Deepak dates the continuous authentication work to 2014, with research partners at the University of Waterloo and Carleton.
- The lock screen experiment
An AI-powered lock screen on the Google Play Store reaches about 25,000 downloads in five weeks across 70 countries and 700 device types, by Deepak's account.
- Barclays Techstars, then a $1 million seed
Zighra graduates from the Barclays accelerator, lands a Barclays pilot, and closes a $1 million seed round in December led by the Investment Accelerator Fund, Export Development Canada and the Business Development Bank of Canada, with six employees (Ottawa Business Journal).
- "Banks and retailers are tracking how you type, swipe and tap"
The New York Times reports on behavioral biometrics in banking, and the Hacker News thread is about consent and about replaying a stolen behavioral profile (Hacker News).
- FIDO certification
Zighra's on-device behavioral biometrics are certified by the FIDO Alliance, which the company calls a first for the category (Biometric Update).
- U.S. patent 11,057,413
A patent for removing passwords through continuous biometric authentication across applications, with sixteen patents issued or pending and more than 100 forward citations (Biometric Update).
- Royal Canadian Navy test contract
An Innovative Solutions Canada contract to test Zighra's sensor identity network for the Department of National Defence (PR Newswire).
- The Hong Kong deepfake call
Police describe a finance worker sending HK$200 million after a video call in which every participant was a deepfake. Arup confirms it was the victim in May (CNN).
- "Passkeys: a shattered dream"
William Brown's post on platform lock-in tops Hacker News with 773 comments, the argument the passkey debate is still having two years later (Hacker News).
- Shared Services Canada
A contract to test Zighra DeepSense for continuous identity validation across government networks (Biometric Update).
- 123,000 flights
Baltic states tell ICAO that nearly 123,000 flights by 365 airlines were affected by GPS disruption in four months, 27.4% of flights in the region in April (ERR, citing SVT).
- FBI alert on Scattered Spider
Help desks talked into adding MFA devices to compromised accounts, now aimed at airlines (FBI on X).
- Episode 78
This conversation. Profitable four years, primarily in defense, "today's wars are initially started off with cybersecurity and electronic warfare."
- Two billion addresses in one batch
Have I Been Pwned loads about 2 billion email addresses and 1.3 billion passwords from a single Synthient dataset (GIGAZINE).
- Spoofing a Starlink dish in Iran
Terminal telemetry shows eighteen satellites locked and anti-spoofing countermeasures engaged: the signals were fake (Nariman Gharib on X).
- 8.6 billion stolen cookies
SpyCloud's 2026 report: 8.6 billion session artifacts, 5.3 billion credential pairs, 13.2 million infostealer infections in 2025 (SpyCloud).
- Five billion passkeys
World Passkey Day: 5 billion passkeys in use, 75% of consumers have enabled one, 57% of organizations still rely on a phishable sign-in (FIDO Alliance).
- IntentHunter
The Department of National Defence will test Zighra's intent-inference platform for counter-drone work (BNN Bloomberg).
- "Magic fairy dust"
Nikita Bier's post on passkeys and the consumer brain draws 4,200 likes and a 785-comment Hacker News thread (Hacker News).
- Canada's Defence Drone Initiative
A supply arrangement to speed up uncrewed and counter-drone systems for the Armed Forces and Coast Guard (National Defence).
- Visa buys BioCatch for $2.4 billion
The largest behavioral biometrics company, 18 billion sessions a month, bought to stop fraud before the point of payment (Finovate).
- All five streams
Zighra qualifies across all five streams of the Defence Drone Initiative marketplace and shows "From Spectrum to Intent" at GCXpo in Ottawa (Zighra on LinkedIn).
- 36% of sign-ins, 7 billion passkeys
The FIDO Passkey Index 2026 (FIDO Alliance).
What to do this week
Four things, in the order an attacker who has read this page would try them.
- Close the phishable gate. Find every path into your main systems that still accepts a password plus a code by text or email, including the fallback paths. Make the primary path a passkey or a device-bound credential. Enabled isn’t enforced has the five questionnaire answers that turn on this.
- Make a stolen session worthless. Bind sessions to the device where your identity provider allows it, shorten the ones you cannot bind, and alert on a session that changes device, location or hours. That alert is your first behavioral signal. Give it an owner.
- Write the rule the video call cannot satisfy. For wires, payroll changes and vendor bank-detail changes: a second channel and a second person, written down, no exceptions for the boss. Deepak’s phrase for the goal is “invisible.” The rule should be boring enough that nobody thinks about it until it saves twenty-five million dollars.
- Count the agents. Every AI agent and service account gets its own scoped credential and a named human. Then ask Deepak’s four questions about each one: who, what authority, what behavior, what intent. The agent credential inventory is the worksheet.
Deepak’s last answer on the tape was about traffic, and his first was about a fake video call, and they are the same answer. Both are friction problems. Security that gets in the way gets switched off, and then a convincing stranger walks through. Security that is invisible, continuous, and built on the device in your hand is the kind people keep. Twelve years after a lock screen on the Play Store, a card network just paid $2.4 billion for that idea, and a navy is testing the same math against drones. Does it need to exist? It does. The full transcript is below, with timestamps, and the guest and source notes follow it.