Most of the security industry treats a stolen password as a problem of time. Rotate it faster, expire the token sooner, add one more prompt. Jasson Casey treats it as a problem of motion. A password, an API key or a session cookie leaves the browser, passes through a proxy, a content delivery network, a load balancer and a service mesh, and sits in memory at every stop along the way. On this episode he put the whole idea in three sentences.
“And so, our kind of take on all of this when we first got started was, it’s not necessarily that long-lived and rotation is the problem. Like, that’s also a heuristic to the problem. The problem is that it moves in the first place.”
Jasson Casey, at 7:52
Jasson is the co-founder and chief executive of Beyond Identity. His LinkedIn shows him as its chief technology officer from September 2019, interim chief executive from July 2023 and chief executive since October 2023 (LinkedIn). The company’s own page lists him as Jasson Casey, PhD, and its board includes chairman and co-founder Jim Clark, co-founder Tom Jermoluk, and Taher Elgamal, the father of SSL, who comes up on the tape (Beyond Identity). Before Beyond Identity, Jasson was chief technology officer of SecurityScorecard from 2016 to 2019, and before that vice president of engineering at IronNet Cybersecurity. He founded Flowgrammable, a nonprofit research organization for software-defined networking, in 2013, and Compiled Networks, which built network appliances for service providers and enterprises, in 2008. His first jobs were in telecom: business development at Level 3 Communications, where he developed its first SIP trunking product, product strategy at Alcatel, and building and running the voice over IP business unit at CenturyTel. He is a fellow in advanced cyber studies at the Center for Strategic and International Studies and a visiting fellow at the National Security Institute at George Mason University’s Antonin Scalia Law School (LinkedIn). His LinkedIn headline today reads “Retiring the credentials attackers exploit” (LinkedIn).
Where we sit, so you can weigh what follows: YSecurity helps startups pass the security reviews enterprise buyers run, and identity is where most of those reviews begin, with how employees sign in, which devices are trusted and who can approve what. Sasha Sinkevich hosted this one alone, and my voice is only in the sponsor reads. We recorded in the fall of 2025, around the time Battlefield 6 came out, and the episode aired on April 7, 2026. In the sign-off, Sasha says the show was brought to you by YSecurity and Beyond Identity. The video went up on our YouTube channel on June 17, 2026.
It surrounds us
Sasha started with Jasson’s years in telecom and the question underneath every call and every connection: how do you know you’re talking to a trusted party? Jasson answered with a line from Star Wars about the Force, then with the problem.
“There’s nothing that you do that shouldn’t involve secure communications in some way, shape, or form. And the reality is, you’re not actually getting it.”
Jasson Casey, at 2:45
His first example is the login form. Most websites still take a username and a password, and most developers send that password as clear text inside TLS, trusting the encrypted connection to keep it private. The catch is that the connection rarely runs end to end. A modern site sits behind a third-party load balancer and a content delivery network, many enterprises run proxies that open and re-terminate TLS to inspect it, and inside the cluster a service mesh may do it again.
“So, like, there are all of these scenarios where TLS, certainly not end-to-end, and generally being opened and re-terminated by third parties.”
Jasson Casey, at 3:32
Every one of those hops is a moment when the password exists in plain text, in memory that someone else operates. That’s the surface area Jasson is describing. It isn’t the login page. It’s every machine the secret passes through, and most of those machines belong to companies you’ll never audit.
Definitely double digits
Jasson gave three examples of what goes wrong in those hops, each a different kind of failure.
The first was a study he half-remembered from the academic world. Researchers created accounts on online services, never logged in again, and watched for the passwords to be used somewhere else. Anything above zero or 1% would have surprised him, he said, “And it was definitely double digits.” The study that matches his description is Tripwire, from a team at UC San Diego, presented at the Internet Measurement Conference in 2017. In a pilot that monitored more than 2,300 sites for a year, the team detected 19 site compromises (UC San Diego). Nineteen is a double-digit count of companies whose users’ passwords got out.
The second was a mistake at the most sophisticated company in the path.
“if you sent the right kind of crafted packet to Cloudflare, you got a 4096-byte dump of somebody else’s TLS connection.”
Jasson Casey, at 4:37
That was Cloudbleed, a buffer overflow in Cloudflare’s edge servers that Google’s Project Zero disclosed on February 17, 2017. The leaked memory held other customers’ HTTP cookies, authentication tokens and POST bodies, and by Cloudflare’s own count it happened more than 18 million times before the fix (Wikipedia). Jasson placed it five or six years before our recording. It was closer to eight and a half, which only strengthens his point that even old news is still the current design.
The third was happening as we recorded. Jasson couldn’t remember which research team had published it, and guessed at the Russian agency behind it.
“And the way they do it is their malware is dropping kind of a CA that they control into the target device, but then it deletes itself, just leaves the CA, or browser plug-in and some other versions, and kind of opens the connection and copies everything out.”
Jasson Casey, at 5:10
The report was Microsoft Threat Intelligence’s, from July 31, 2025. It describes Secret Blizzard targeting embassies in Moscow from an adversary-in-the-middle position at the internet service provider, with custom malware called ApolloShadow that installs a trusted root certificate so devices trust sites the attackers control. Microsoft notes that CISA attributes Secret Blizzard to Center 16 of Russia’s Federal Security Service (Microsoft), the same unit a 2023 joint advisory tied to the Snake malware (Wikipedia).
None of the three needed a password to be weak. They needed it to travel.
Almost like an iceberg
Sasha’s reply widened the frame to the parts of a system nobody sees.
“It’s almost like an iceberg. When you communicate with a platform or with a system, it may seem that on the surface things are [peachy], but you don’t really know what happens under the hood.”
Sasha Sinkevich, at 5:35
Sasha added the supply chain, the packages under the hood that the end user of a system never gets to control, and then brought up a past guest. Taher Elgamal, known as the father of SSL, told us on episode 61 that passwords are holding us back. Sasha asked what the solutions look like.
It was a coincidence, Jasson said. He had talked to Taher two hours earlier.
“He’s on our board, and he was an early advisor in the company.”
Jasson Casey, at 6:21
Beyond Identity’s site lists Taher on its board of directors alongside Jim Clark and Tom Jermoluk (Beyond Identity). Jim Clark founded Silicon Graphics and Netscape, and his Wikipedia entry records that in 2020 he announced a Series A round for Beyond Identity, for a phone-resident, certificate-based way to sign in that eliminates passwords (Wikipedia). The company was built around the idea Taher described on our show.
Static, it’s long-lived, you can rotate it
Jasson’s first move was to make the question bigger than passwords.
“So it goes deeper than passwords, would be my first point. Access tokens, API key, session cookies.”
Jasson Casey, at 6:24
All of them are shared secrets, a string that proves who you are to anyone who holds a copy. Tying a session to an IP address, he said, is “a very brittle and silly concept, and certainly a security heuristic.” And the attacks that end in session hijacking all come down to the same move.
“When you look at classes of attacks that ultimately result in session hijacking, what are they doing but copy and pasting some sort of symmetric secret, right, or shared secret, from one location into another?”
Jasson Casey, at 6:36
Sasha summed up the usual answer in six words, and Jasson pushed past it.
“Static, it’s long-lived, you can rotate it.”
Sasha Sinkevich, at 6:46
“Well, I would argue it’s even more fundamental than that. Imagine it wasn’t long-lived. Imagine it was short-term use. The fact that it moves is where the original surface area comes from.”
Jasson Casey, at 6:48
This is the center of the episode, and it’s worth slowing down on. Shorter lifetimes, rotation and IP pinning all assume the secret will be copied and try to make the copy less useful. Jasson wants to make copying impossible. We wrote the practical side of this up in Why Passwords Still Get Stolen: The Case for Device-Bound Credentials, the companion post to this episode, which covers how device-bound and synced passkeys differ and how a startup rolls them out.
A series of bits
To show how far a secret travels, Jasson borrowed the opening of Tron: Legacy, imagining himself as a series of bits moving through the microprocessor, and followed a session cookie or an access token from the browser to the database. It’s written and read from memory on the laptop. In an enterprise it probably passes through a Palo Alto or Zscaler proxy. Then a content delivery network such as Cloudflare or CloudFront, then the regional load balancer in the cloud account, maybe another load balancer in the zone, then a Kubernetes service mesh. Then he went back to how computers work.
“When I store and then load a piece of memory, the data’s still in the memory, unless I go back and write over that memory.”
Jasson Casey, at 7:46
That’s the part that’s easy to forget. Each hop doesn’t just see the secret go by. It holds it in memory until something overwrites it, which is exactly what Cloudbleed exposed. Rotation shortens how long a stolen copy is useful. It doesn’t reduce the number of copies.
What if we could guarantee it didn’t move?
The first answer, Jasson said, is trivial.
“Yeah, you could adopt some sort of asymmetric cryptography, use a private key for signing, and then the only thing that has to move is the public key, which gives away nothing.”
Jasson Casey, at 8:10
That’s how TLS and SSH already work on the server side, and it’s how passkeys work for people. The second answer is the one Beyond Identity was built on.
“But then one of our guys came up with this stronger idea, which is, what if we could guarantee it didn’t move?”
Jasson Casey, at 8:18
The engineer’s observation was that servers already have hardware security modules, which create key pairs whose signing key can’t leave the box. And from mobile payments he knew that something very much like an HSM, in Jasson’s words, “exists in your pocket and exists on your desktop.” Chipmakers prefer to build one product line rather than many, so almost every modern processor now ships with a secure cryptographic coprocessor.
“And if it never moves, it can never be copied, it can never be stolen. And so, like, a class of problems gets eliminated.”
Jasson Casey, at 9:07
The chips have names. A trusted platform module, or TPM, is a secure cryptoprocessor built to the ISO/IEC 11889 standard, first deployed in 2003, and TPM 2.0 is part of the system requirements for Windows 11 (Wikipedia). Microsoft’s documentation for Windows Hello for Business says the private key it creates is protected by the TPM and can’t be exported (Microsoft). Apple devices have the Secure Enclave (Apple), ARM chips have TrustZone, and Intel had Software Guard Extensions, the old-school instruction set Jasson mentions, which Intel deprecated on its 11th and 12th generation Core processors in 2021 while keeping it on Xeon (Wikipedia). Sasha put the shared idea in one sentence.
“And then you hold the secret inside that secure enclave, and then you have a guarantee that nothing moves out.”
Sasha Sinkevich, at 9:28
The standards have caught up with this. NIST’s current digital identity guidelines require that authentication at the highest assurance level use a phishing-resistant authenticator with a non-exportable key (NIST). What Jasson describes as a design choice is now the federal definition of strong.
Battlefield 6
His surprising example of the same hardware in the wild was a video game. The observation of one of Beyond Identity’s early engineers, he said, was that the chips existed and almost no one used them for anything but mobile payments. Then he named the game that had just come out.
“So what they’re doing is they’re making the kernel prove at load time that there’s no unauthorized drivers that some of the guys who come up with mods make”
Jasson Casey, at 9:56
Battlefield 6 came out on October 10, 2025, and it uses EA’s kernel-level anti-cheat, Javelin, which requires Secure Boot to be turned on in the computer’s firmware (Wikipedia). The requirement caused a fuss among players, which, Jasson admitted, is the only reason he knew about it. Sasha turned it into a business lesson.
“This is actually super cool, because this is a perfect example of security solving a very specific and painful issue for a business.”
Sasha Sinkevich, at 10:16
Cheaters bypass the controls that make people buy things, and that shows up in revenue. Jasson agreed, and gave the line that runs under the rest of the episode.
“I mean, at the end of the day, security is there to serve the business.”
Jasson Casey, at 10:50
Every so often, he said, security is the business, as it is for Beyond Identity. Most of the time it exists so that everyone else has a good experience, whether that’s a multiplayer game or a customer portal. That framing comes back at the end of the episode with a number attached.
Five seconds, ten seconds
Then Sasha turned to the question the episode title promises.
“It takes about five seconds, 10 seconds for a really accurate representation of your voice and your image generated by AI.”
Sasha Sinkevich, at 11:33
Jasson is a public figure who has been on a lot of shows, Sasha pointed out, which means there is plenty of his voice and face to train on. At that point it’s hard to know whether you’re talking to the actual person on the other end of the TLS connection the conversation had just spent ten minutes on. So how do you know?
We’ve had the detection side of this question on the show too. Ben Colman of Reality Defender talked about how startups can fight deepfake fraud on episode 71, and Vijay Balasubramaniyan of Pindrop on voice on episode 57. Jasson came at it from the other direction.
Knife fights and phone booths
He started, as engineers do, by asking whether the problem was framed right.
“They think that’s the problem. That always kind of didn’t sit well with me. And I was trying to figure out why.”
Jasson Casey, at 12:17
The problem most of the industry had picked was “is this a deepfake?” He found two reasons it bothered him. The first was engineering. Detection felt like another heuristic.
“I probably can detect the artifacts of how you generate deepfakes, but how long will it last? Is it stable? Could you use me to train another generator?”
Jasson Casey, at 12:31
A detector learns the artifacts of the generators that exist today, and every detector can become training data for the next generator. Choosing that fight, he said, felt like choosing a career in knife fights in phone booths. The outside evidence points the same way. When Australia’s national science agency, CSIRO, tested 16 leading deepfake detectors in 2025, it found that many struggled with deepfakes outside their training data (CSIRO). Jasson’s alternative is a rule for business as much as for engineering.
“You want unfair fights. You only want to show up where it’s almost impossible for you not to win, or solve the problem, or whatnot.”
Jasson Casey, at 12:51
A realistic mustache
The second reason was stranger, and to me it’s the more convincing one. If everyone is right that AI is about to be in everything, then a lot of synthetic video will be legitimate. Jasson listed his own examples: trying out a realistic mustache before growing one, translating himself into Spanish in real time because he doesn’t speak it, or covering an injury he’d rather not discuss on a call.
“So if AI is really around the corner, we know there’s valid uses for deepfakes, right?”
Jasson Casey, at 13:27
“And if all these tools have decided they’re going to work on deepfake detection, then how does it work against valid use?”
Jasson Casey, at 13:36
A detector that works perfectly would still flag the colleague using a translation filter. Whether a video is synthetic stops being the same question as whether it’s an attack the moment synthetic media becomes normal, and that moment is close.
Who’s on the other end?
So Jasson changed the question.
“The right question is, who’s producing this content right now, from what computer, with what level of authenticity and trust?”
Jasson Casey, at 13:45
The same chips that keep keys from moving can answer it, through a mechanism called attestation. In the moment, the device produces a mathematical proof about itself.
“under the assumption that the foundry that made this coprocessor hasn’t been violated, I can prove to you I’m this device, made in this location, with this configuration, and this is the programmed user that’s running”
Jasson Casey, at 14:10
The proof can cover more than the device. It can say the user unlocked it with a biometric, that endpoint protection is running, that the system has been hardened. How to show all that to a person, he said, is a security UX question, small emblems, an audit log, the details on demand. The core of it is that nothing in the answer depends on what a face looks like.
Beyond Identity sells this today as RealityCheck. Its product page pitches it as making deepfakes irrelevant rather than detecting them, by verifying both the user and the device behind the camera at login and during the meeting (Beyond Identity). Our companion post, Deepfake Attacks Keep Working Because We Keep Detecting Instead of Proving, walks through what a proof step proves and what a startup should change first.
I mean commit the code
Beyond Identity built the answer into the tools people already use, Jasson said: plug-ins for Zoom, Teams and Outlook, and for GitHub and GitLab. Each answers the same question at a different moment. Who is the person, on what device, with what posture, on this call? Who sent this email? And for code, something more specific than a login.
“I don’t mean log in to the repo’s website, and I don’t mean synchronize the repo’s database. I mean commit the code.”
Jasson Casey, at 15:16
Developers type git commit, git pull and git push without thinking, but each one has its own authentication and authorization model.
“And just because you log into the website using something secure doesn’t mean you really know where that code came from.”
Jasson Casey, at 15:32
Sasha put the mechanism in plain words, and Jasson added the part that makes it flexible.
“In essence, it’s a signature.”
Sasha Sinkevich, at 15:37
“It’s a signature over a series of properties that you can kind of decide what level of trust you want to assign.”
Jasson Casey, at 15:47
This is the idea I’d hand to a startup preparing for its first enterprise security review. Buyers increasingly ask who can change production and how you know. A login log answers who signed in. A signature from a registered device on the commit, the deploy or the wire approval answers who did the thing.
Some dude in the room
Sasha guessed that this would fit neatly into a company that already runs mobile device management. Jasson agreed, with a story that anyone who has run an enterprise deployment will recognize.
“And then there’s some dude in the room that won’t make eye contact anymore.”
Jasson Casey, at 16:27
Every organization starts the same way, he said: a hardened environment, everything managed, endpoint protection everywhere, no exceptions. Then, around the second or third meeting, the exceptions come out. Some devices don’t have device management. There are PR partners, legal partners, outsourced developers, an outsourced call center.
“And so the reality of, like, an actual enterprise is always heterogeneous.”
Jasson Casey, at 16:44
So Beyond Identity built for both worlds from the start. With device management, the best part is distribution. Their integration can package and ship the authenticator and also enroll the users automatically, “to where they don’t even have to be part of the rollout,” and it can feed what device management sees into each decision. Without it, the same checks still have to work.
The identity bridge
The reason it can work without device management is the line I’d put on a poster.
“Because identity is the thing that everything crosses.”
Jasson Casey, at 17:19
Employee or contractor, business partner, managed device or personal laptop, everyone crosses the same bridge to reach an application.
“So the identity bridge is really the best point of enforcement to understand who, on what device, is asking for what, and for how long.”
Jasson Casey, at 17:29
That shaped one more design decision. The authenticator had to run without elevated privilege, because nobody wants to install a privileged work tool on a personal device, and Beyond Identity didn’t want the power to remote-wipe anyone’s phone. Jasson expected pushback on how much you can see without privilege, and he had an answer ready: think about how malware works, because it starts without privilege too.
AI in our toothpaste
Then Jasson tied it back to AI, with the line our social team turned into a quote card.
“We now have AI in our bread, we have AI in our toothpaste, everybody’s AIing all the things.”
Jasson Casey, at 18:10
Some of those things will become businesses, he said. In all of them, developers paste secrets and credentials where they shouldn’t. “It’s a new game, but all the same injuries.” Sasha said the classical security principles still apply, and Jasson named the injuries: credential sprawl, devices and services that don’t meet policy, data sprawl, and data classes that mix. A company may have a part of the business under CMMC, a part under FedRAMP, and services that clearly aren’t approved for either.
That list is the one we see in startups adopting AI quickly, and it’s why we wrote Your Agents Share One API Key and the AI agent credential inventory. The fix starts with knowing where each secret lives.
What user, on what device, with what posture
Jasson’s next answer is, to me, the most useful paragraph in the episode for anyone deploying agents. Start from first principles, he said. Even in an AI system there’s a user on a device and an agent that runs either on that device or on a cloud workload, and the agent has access to services, vector databases, MCP servers and the rest.
“I want to know what user, on what device, with what posture, is using what agent, running what model, on what device, with what posture, and has been given permissions to what services, for how long.”
Jasson Casey, at 19:03
“These are the classic identity questions.”
Jasson Casey, at 19:15
He wants the answer every time, tamper-proof and attested, so he can audit and enforce, and so his team can keep putting AI in toothpaste without putting the company’s CMMC compliance at risk. He comes back to these questions near the end of the episode and turns them into a vector.
If your parents can’t do it
Sasha asked how much seamlessness matters in security products, which are often complex to deploy and maintain. Jasson split the audience in two. The IT administrator will live, breathe and sweat your product, and there’s complexity you can’t hide from them. The end user is different.
“If you only have a soundbite, the end user is your parents. So, if your parents can’t do it, you’re going to fail.”
Jasson Casey, at 20:16
The lesson Beyond Identity learned the hard way was to keep those two people apart in the design.
“We merged those concepts a little bit too much up front.”
Jasson Casey, at 20:42
They oversimplified things for administrators and didn’t simplify enough for end users. Treating end users like your parents means remembering that their job isn’t technical. They work in finance, accounting or medicine, and use technology to do it. Every time a product asks them to take part in something that doesn’t help them do that job, he said, it’s a small self-inflicted wound, and at a company with 80,000 employees that adds up fast.
Asking a user to change their password
The biggest example is how identity products roll out.
“the standard operating procedure in the identity world is to involve the entire workforce in rollouts. That’s horrible.”
Jasson Casey, at 21:21
It’s also, he said, why identity products rarely get replaced. Sasha added the practical side: you depend on people who are busy, unavailable or on vacation, and who come back to a full inbox. Jasson’s example is so familiar that nobody questions it.
“asking a user to change their password is asking a user to be part of a deployment update for a technical control.”
Jasson Casey, at 21:53
“But I put the failure more on the system designers, like us.”
Jasson Casey, at 22:24
A password lives in someone’s brain, so the person has to be part of every change to it. A credential that lives in the device doesn’t need them. The guidance has moved the same way. NIST’s current rules say verifiers “SHALL NOT require subscribers to change passwords periodically,” and force a change only when there’s evidence of compromise (NIST).
Beyond Identity’s rule after that, Jasson said, was to find the absolutely necessary things end users have to do, and nothing else.
The champion, the budget holder, the administrator and the user
Sasha connected it to a startup basic, knowing your ideal customer profile and its specific pain, and asked whether Beyond Identity embeds engineers with customers. Their most successful deployments work exactly that way, Jasson said, and then he added a warning for founders trained on consumer software.
“The champion, the budget holder, the administrator, and the user. These are all very different people, and they can all shut you down for different reasons.”
Jasson Casey, at 23:47
What sells the product to one of them can work against it with another, which is why he says UX, product management and product marketing have to move in lockstep in enterprise infrastructure. And enterprise administrators expect a vendor that’s critical to their business to be in the trenches with them, however new and exciting the product is. Not every customer wants that, though. Some, in his words, want you to deliver the equipment to a parking lot on a given day and leave the rest to them.
Flowgrammable
Sasha asked about the proudest day of Jasson’s career. He said he likes building things from scratch, and teams from scratch, and that many of the people he works with now he has worked with on and off for half his career. The project he picked came from graduate school.
“We did a thing back when I was in grad school called Flowgrammable that I was very proud of.”
Jasson Casey, at 25:11
Software-defined networking was the big idea of the moment. His LinkedIn describes Flowgrammable as a nonprofit research organization that maintained a knowledge base, open source tools and research on the protocols, which he founded in May 2013 and recruited with industry engineers and university researchers (LinkedIn). On the tape he describes a team of PhD, master’s and undergraduate students who built kernel drivers and user space applications and entered reference design competitions against open source software supplied by the industry, and beat all of them but one. It took everything you know about how a computer works: the CPU, the cache, what the compiler won’t figure out for you, and six versions of a protocol that nobody wants to be the developer’s problem.
“Only so many developers are going to read a thousand-page spec.”
Jasson Casey, at 25:56
Everyone from that team, he said, now works with him at Beyond Identity, worked with him at SecurityScorecard or IronNet, or landed at big tech companies on the West Coast with his help. He knows the counterculture against calling work a family. He doesn’t share it, and said so: “startups aren’t families, but they kind of are.”
Run-ins with state actors
Sasha said a company is just a collection of people, and that the best part of a startup is iterating on one problem quickly and protecting organizations that serve millions of people. Jasson took the thread further.
“But a little further in that story is, we’ve had a couple of run-ins with state actors, and our product performed, and our team performed exceptionally.”
Jasson Casey, at 27:23
He didn’t name the actors or the customers, and neither will we. What he described is what it feels like when a design meets the adversary it was built for.
“It’s one thing to do something and know intellectually, like, why you do it and what you’re doing it, but it’s a whole nother thing to actually see it do its job”
Jasson Casey, at 27:47
Two incident response teams and red teams credited the product afterward, he said, and the company came out of it with relationships it didn’t have before.
Shift left on the spectrum
Sasha brought the conversation back to the start. Trying to identify deepfakes is chasing the problem, Sasha said, a reactive function, when the systems we all want to build are proactive and prevent the event instead of analyzing it afterward. Jasson put that in historical terms.
“And it was, hey, we’re all compromised. We now live in a post-compromised world. How do we find the bad guy and evict them?”
Jasson Casey, at 28:43
That was the strategic choice of the 2010s, he said, the rallying cry that endpoint detection and response was built around. This decade is a chance to shift left. He reached for the NIST Cybersecurity Framework and listed its functions as identify, protect, detect, respond and repair, then corrected himself that the last one is something else. It’s recover. Version 2.0, released in 2024, added a sixth function, govern, so the framework now reads govern, identify, protect, detect, respond and recover (Wikipedia). His point stands either way: most of the money went to detect and respond, and the opportunity now is in identify and protect.
“If I can guarantee credentials don’t move, then credential theft goes away, stuffing, spraying, all of that goes away.”
Jasson Casey, at 29:28
If the authenticator runs on the device where the work happens, it can spot an adversary in the middle and prevent session hijacking. If it carries a small sensor and asks whether the device is safe enough for what it’s asking for, it can stop the pre-staging tactics he described at the start, the root certificates and browser plug-ins Secret Blizzard used.
“if you pull up your MITRE chart, there’s a ton of squares now that have X’s through them”
Jasson Casey, at 30:02
The part I’d underline for a small team is his closing note on it: “Simpler things are easier to maintain. Fewer moving parts are easier to maintain.” Removing a class of attack is also removing work.
Single digits in production
Sasha then raised the AI adoption gap. Companies have invested billions in AI and built workflows meant to make everything better, but the share that reaches production is small.
“From 100%, only single digits are deployed in production.”
Sasha Sinkevich, at 30:50
One of the core reasons, in Sasha’s telling, is that companies don’t trust the system enough and can’t audit whether it did what it was meant to do. Jasson didn’t claim to know all the reasons, but he hears one version of it from customers again and again. A defense contractor has to keep its CMMC accreditation. A federal civilian business has to keep FedRAMP. A retailer has PCI DSS.
“How do I make sure, if I’m going to turn on this agentic system and give it access to some of my controlled data, that the controls will flow with the data?”
Jasson Casey, at 31:26
That fear slows those organizations down, he said, while their developers are already pulling in whatever AI tool helps. Beyond Identity knows the compliance side from its own work. It announced FedRAMP Moderate authorization through the SMX Elevate platform on April 8, 2025 (Beyond Identity). For startups selling into those markets, our CMMC Level 2 and FedRAMP readiness work starts with the same question Jasson’s customers ask: which identities touch the controlled data, and can you prove it?
Identity sees everything
Jasson’s answer to those customers starts with the fundamentals of computer science.
“every process, whether it’s a daemon, a scheduled thing, or run by a human, has a user ID.”
Jasson Casey, at 32:03
The user ID may or may not belong to a person, but there’s always an identity. Connect processes across a distributed system and identity shows up again, either implicitly, as the cross product of all those identities, or explicitly, because someone designed it into the data model.
“So if identity is involved in everything, that means identity sees everything.”
Jasson Casey, at 32:24
AI systems, he said, mostly look alike: an orchestration layer that calls open source or public models, MCP servers and AI services, data stores old and new, all running on machines that have a posture. So ask the basic questions about every one of them. The only new thing is the shape. It’s no longer client and server, but client, an AI party, and server, sometimes with several servers and several agents.
The vector
Then came the most precise definition of agent identity I’ve heard on the show.
“Well, an agent, fundamentally, it’s a program.”
Jasson Casey, at 33:27
An agent senses its surroundings, builds a model, decides, acts and measures the result, or it’s simply a workflow that gathers information and writes a report. Either way it’s a program, and we already know how to identify programs. That’s what a software bill of materials is for.
“And so, to us, agent identity is really just tying in the SBOM to the device identity that’s executing it, to the provider identities that we’re attaching to it.”
Jasson Casey, at 33:51
Then he laid it out as a vector, since everyone in AI likes vectors. The top is the user: user ID, the user’s device ID, its posture and the user’s geography. The middle is the agent: agent ID, the agent’s device ID, posture and location, sometimes the same machine as the user’s. The bottom is the providers: their identities, and the permissions you’re actually allowing.
“Well, the top of the vector is user ID, user’s device ID, user’s device posture, user’s geography.”
Jasson Casey, at 34:04
Make it tamper-proof and you can audit it. Audit it and you can add a policy layer that decides who can do what. Collect it and you have an intelligence source for everything happening in your AI infrastructure. There’s endpoint detection and response, and identity threat detection and response, and, he guessed, there’s probably an AIDR coming. Sasha finished the thought: the acronyms change, and the core principles of security don’t. For a startup putting agents into production, this vector is a good checklist for the questionnaire that’s coming, and it’s the same chain we argue for in Who Signs for the Agent?.
They price in fraud
Security is always a scale, Sasha said. You can make something perfectly secure and nobody will use it. Jasson put a price on the trade.
“What’s the cost of being wrong versus the benefit of being right?”
Jasson Casey, at 35:13
A scaled business measures that precisely, he said. Banks set their investment from their balance sheets.
“They price in fraud. And only when fraud passes a certain level do they increase their investments in fraud detection and prevention.”
Jasson Casey, at 35:21
From 80% to 95%
Sasha asked for his words of wisdom, and Jasson’s first two were the episode in miniature.
“Fundamentals matter. You still kind of need to know how a computer works.”
Jasson Casey, at 35:37
“I think there’s this implicit assumption the industry has, you can only run detection and response. I think it’s time to actually challenge that.”
Jasson Casey, at 35:40
Move to credentials that can’t move, with Beyond Identity or anyone else, he said, and you’ll see fewer help desk tickets and fewer security incidents. Done right, it isn’t a productivity loss. It’s a gain. Then he gave the number.
“One of our customers is using our technology for their customer side of the business, and they’ve seen the success login rate of their customers go from 80% to 95%.”
Jasson Casey, at 36:16
The customer did it to stop fraud and account takeover, and was running reports to work out what the extra successful logins meant for revenue.
“So, like, this is one of those rare times where doing the right thing from a security perspective is actually a better experience, and, depending on your business, can actually contribute to the top line.”
Jasson Casey, at 36:36
Sasha closed on the same note: plenty of security problems can be lined up with revenue growth or a better user experience, which leads to revenue anyway. You can try the arithmetic on your own sign-in numbers in the login calculator further down.
What happened after the tape
The months after the recording read like the episode turned into a product plan. In February 2026 Jasson went on Legitimate Cybersecurity, hosted by Frank Downs and Dustin Brewer, to talk about what breaks in identity and access when software can reason, plan and act.
Two days later Beyond Identity opened a public preview of Ceros, which Jasson called its agentic AI trust layer. His post makes the argument from this episode about agents.
Trying to detect every bad behavior is a losing strategy. The leverage point is identity.
Jasson Casey · February 17, 2026
In March, during RSAC week, he wrote that Beyond Identity was working with his alma mater on identity for drones, using the same primitives it uses for AI systems: hardware-rooted identity, attestation and policy.
Proud to be working with my alma mater, Texas A&M University, on bringing identity to drones.
Jasson Casey · March 25, 2026
This episode aired on April 7. In June, at Identiverse, Ceros launched.
Ceros is live! The first trust layer built for the agents your IAM stack never planned for.
Beyond Identity (@beyondidentity) · June 16, 2026
That list is close to the vector Jasson described on the tape. The same day he shared an article from Cyber Insurance News that asks whether a company can account for its AI agents, in which he’s quoted saying underwriters “can’t insure what they can’t measure.”
How do you handle agents for cyber insurance? Here is a conversation around just that topic.
Jasson Casey · June 16, 2026
The episode’s video went up on our YouTube channel the next day. Beyond Identity’s AI trust layer page now describes keys for agents that never leave hardware, discovery of every MCP server, continuous posture checks, and a trace of the full identity chain for every agent action, which agent, which device, which tool and which user (Beyond Identity). And on November 3, 2026, Jasson reposted a keynote on drone identity that Ligeia Zeruto of Beyond Identity gave at DRON8, with a first line that could stand as this episode’s summary.
Securing identity against human and AI adversaries starts with first principles. Device-bound identity is foundational to getting it right.
Jasson Casey · November 3, 2026
Read together, the arc is consistent. On the tape, Jasson’s answer to stolen passwords, deepfaked executives and over-permissioned agents was the same one: stop asking whether something looks right, and ask the hardware to prove who did it, from where, in what state. Within a year his company had applied it to agents, to drones and to insurance. For a startup the practical version is smaller and just as useful. Find the secrets that travel, and make the important ones stop.
Did you catch it? A six-question quiz
Interactive · six questions
Did you catch it?
Six questions on what Jasson and Sasha actually said, with the timestamp for each answer. The answers are under each question for anyone reading without JavaScript.
-
What does Jasson say is the real problem with passwords, tokens and session cookies?
- They live too long
- They aren't rotated often enough
- They move in the first place
c. "The problem is that it moves in the first place." Transcript, 8:03.
-
Which game does Jasson use as an example of a company using the chip in your computer?
- Battlefield 6
- Fortnite
- Valorant
a. "What's the game that just got released? Battlefield 6." Transcript, 9:46.
-
What does Jasson say is the right question to ask instead of "is this a deepfake?"
- Does this face match the one on file?
- Who's producing this content right now, from what computer, with what level of authenticity and trust?
- Has this video been edited?
b. "The right question is, who's producing this content right now, from what computer, with what level of authenticity and trust?" Transcript, 13:45.
-
Why does Jasson call identity the best point of enforcement?
- It's the cheapest control to buy
- Device management covers every device
- It's the thing everything crosses
c. "Because identity is the thing that everything crosses." Transcript, 17:19.
-
Who does Jasson say the end user is, if you only have a soundbite?
- Your parents
- An engineer
- The IT administrator
a. "If you only have a soundbite, the end user is your parents." Transcript, 20:16.
-
What happened to one customer's login success rate?
- It held steady at 95%
- It went from 80% to 95%
- It dropped while fraud fell
b. "One of our customers is using our technology for their customer side of the business, and they've seen the success login rate of their customers go from 80% to 95%." Transcript, 36:16.
How far do your secrets travel?
Jasson’s test for any credential is whether it moves. This scorecard asks eight questions about how your company signs in, approves and deploys, from passwords to AI agents. It takes about two minutes.
Scorecard · eight questions
How far do your secrets travel?
Answer for the systems that hold customer data and the people and agents who can change them. If two answers fit, pick the lower one.
-
How do employees sign in to the tools that hold customer data?
- Passwords, with codes or push approvals on top
- Passkeys for some people and some tools
- Phishing-resistant, device-bound credentials for everyone, with no password to type
"The problem is that it moves in the first place." See Static, it's long-lived, you can rotate it.
-
If someone copied a session cookie or an access token from a laptop, could they use it from another machine?
- Yes, until it expires
- For a short time, because we rotate and expire them quickly
- No, sessions and tokens are bound to a key held in the device
"Imagine it wasn't long-lived. Imagine it was short-term use." See A series of bits.
-
When someone asks over a call or a chat to move money, change bank details or reset access, what proves who asked?
- A familiar face, voice or name
- A call back to a number we already have
- An approval signed from a registered device, with a second approver for large amounts
"The right question is, who's producing this content right now, from what computer, with what level of authenticity and trust?" See Who's on the other end?.
-
Can you tell who committed a change to production code, and from what device?
- We know which account pushed it
- Some commits are signed
- Commits and deploys are signed with keys held in registered devices
"I mean commit the code." See I mean commit the code.
-
Do contractors, partners and unmanaged devices go through the same identity checks as employees?
- They use shared accounts or sit outside the policy
- They have their own policy, and it covers some of them
- Everyone crosses the same checks, with or without device management
"Because identity is the thing that everything crosses." See The identity bridge.
-
How much does an employee have to do when you roll out a new way to sign in?
- Everyone installs and enrolls by hand
- Device management ships the software, and people still enroll themselves
- They're enrolled automatically and aren't part of the rollout
"the standard operating procedure in the identity world is to involve the entire workforce in rollouts. That's horrible." See Asking a user to change their password.
-
For each AI agent you run, can you name the user it acts for, the device it runs on, the services it can reach and for how long?
- We don't have a list of our agents
- We have a list, but not the chain from user to device to permission
- Every agent session ties to a user, a device and permissions that expire
"These are the classic identity questions." See The vector.
-
How often do you make people change their passwords?
- On a schedule
- Only when there's evidence of compromise
- There's no password left to change
"a password lives in someone's brain." NIST's current rules say not to force periodic changes. See Asking a user to change their password.
Your most important accounts still depend on secrets that can be copied, phished or replayed, and approvals that rest on a face or a voice. Start with the accounts that could end the company: move their sign-in to device-bound passkeys, and put money movement behind signed approvals from registered devices.
You've anchored part of the picture. Close the gaps attackers reach next: sessions and tokens that can be replayed elsewhere, contractors and unmanaged devices outside the checks, and agents without a clear chain from user to device to permission.
Credentials stay in hardware and actions carry proof of who did them. Keep it that way as you add agents: give each one an identity tied to a user and a device, and make its permissions expire.
What is a sign-in worth?
Jasson’s customer saw its login success rate rise from 80% to 95% after a change made for fraud reasons. Put in your own numbers to see what that kind of change is worth in completed sign-ins and revenue.
Calculator
What is a sign-in worth?
Four estimates for your customer-facing product: how many sign-ins people attempt in a month, how many succeed today and with phishing-resistant passkeys, and what a completed sign-in is worth to you on average.
Going from 80% to 95% completes 15,000 more sign-ins a month and removes 75% of the failed ones. At $3 a sign-in, that's about $45,000 a month, or $540,000 a year, before you count the help desk tickets and fraud you no longer have.
The 80% to 95% change is one Beyond Identity customer's result, as Jasson describes it on the tape. Your numbers will differ. The point is that a sign-in that fails is a customer who didn't get in.
From voice over IP to agent identity
Everything on this page sits on one timeline, from the first TPM to drone identity. Filter it by thread.
Timeline · 2003 to 2026
From voice over IP to agent identity
Filter by thread. Dates are from the public record except where marked as part of this show.
- The first TPM
Version 1.1b of the trusted platform module ships, conceived by the Trusted Computing Group (Wikipedia).
- Level 3 Communications
Director of business development, developing the company's first SIP trunking product (LinkedIn).
- CenturyTel
Builds and runs the voice over IP business unit, for consumer, enterprise and carrier markets (LinkedIn).
- Compiled Networks
Founder and chief executive of a company building network appliances for service providers and enterprises (LinkedIn).
- Flowgrammable
Founds a nonprofit research organization for software-defined networking, the project he calls his proudest (LinkedIn).
- IronNet Cybersecurity
Vice president of engineering, later on the advisory board (LinkedIn).
- SecurityScorecard
Chief technology officer until August 2019 (LinkedIn).
- Cloudbleed
Google's Project Zero discloses a buffer overflow that leaked other customers' cookies and tokens from Cloudflare's memory (Wikipedia).
- Tripwire
A UC San Diego pilot monitors more than 2,300 sites for a year and detects 19 compromises (UC San Diego).
- Beyond Identity
Jasson joins as chief technology officer (LinkedIn).
- A Series A
Jim Clark announces Beyond Identity's Series A, for certificate-based sign-in that eliminates passwords (Wikipedia).
- SGX steps back
Intel deprecates Software Guard Extensions on 11th and 12th generation Core processors (Wikipedia).
- Chief executive
After four months as interim chief executive (LinkedIn).
- NIST CSF 2.0
The Cybersecurity Framework adds a sixth function, govern (Wikipedia).
- Episode 61
Taher Elgamal, the father of SSL and a Beyond Identity board member, on why passwords are holding us back (episode 61).
- FedRAMP Moderate
Beyond Identity announces its authorization through the SMX Elevate platform (Beyond Identity).
- Secret Blizzard
Microsoft reports a campaign against embassies in Moscow that installs a trusted root certificate from inside the internet provider (Microsoft).
- Battlefield 6
Released with EA's Javelin anti-cheat, which requires Secure Boot (Wikipedia).
- We record episode 92
"The problem is that it moves in the first place."
- Legitimate Cybersecurity
Jasson on what breaks in identity and access when software can act (YouTube).
- Ceros preview
Beyond Identity opens a public preview of its agentic AI trust layer (LinkedIn).
- Identity for drones
Work with Texas A&M on hardware-rooted identity for autonomous systems (LinkedIn).
- Episode 92 airs
This conversation, on Podbean, Apple Podcasts and Spotify (Podbean).
- Ceros is live
Launched at Identiverse, with identity checked on every agent session (X).
- On YouTube
The video of this episode goes up on our channel (YouTube).
- First principles
Jasson on device-bound identity for people, machines and agents, sharing a DRON8 keynote (LinkedIn).
What to do this week
Five things, in the order Jasson’s argument suggests.
- Find the secrets that travel. List where passwords, API keys, tokens and session cookies live and every system they pass through on the way, starting with the accounts that could end the company. The scorecard above is a quick first pass.
- Anchor the important sign-ins. Move administrators, finance and anyone with production access to phishing-resistant, device-bound passkeys first, with a second authenticator for recovery. Our companion post on device-bound credentials has the rollout plan.
- Prove the action, not just the login. Put wire approvals, bank detail changes, access resets and production deploys behind a signature from a registered device. A face on a call doesn’t count as proof.
- Keep users out of the rollout. Use device management to ship and enroll automatically, and stop scheduled password changes, which NIST’s current guidelines no longer ask for.
- Give every agent an identity. For each agent, write down the user it acts for, the device it runs on, the services it can reach and when that access ends. If you want a second pair of eyes on any of this, our product security and AI access control audit teams do this work every week, and you can book 15 minutes with us.