Co-Founder, Chief Executive Officer & Chairman of the Board, Zscaler
The AI Cybersecurity Crisis Nobody Is Ready For | With Jay Chaudhry |Live at Zenith'26 Live, Zscaler
AI Will Find Every Vulnerability: Zscaler Founder on Mythos, Zero Trust & the Future of Cybersecurity
Artificial Intelligence is transforming cybersecurity faster than any technology wave before it. In this exclusive conversation, Zscaler Founder and CEO Jay Chaudhry explains why AI is a "giga wave," how Project Mythos is uncovering hidden vulnerabilities, and why traditional security models are no longer enough.
Jay shares practical strategies for CIOs, CISOs, and Chief AI Officers to defend against AI-powered cyber threats, including Zero Trust architecture, attack surface reduction, decep...
About Jagtar Chaudhry
Jay Chaudhry, co-founder and CEO of Zscaler, has been discussing the cybersecurity implications of artificial intelligence, particularly the rise of AI agents and the impact of Anthropic's Mythos vulnerability-finding model. He described AI as a "giga wave" and argued that traditional security models based on firewalls and VPNs are outdated, advocating instead for a zero-trust architecture in which applications are hidden behind an exchange. Chaudhry stated that AI agents, which he said could number between 50 and 100 per user, represent a new security risk because they "act in milliseconds" and "take no break, no weekends, no sleep." He characterized Mythos as "powerful" but suggested its launch involved "a little bit of marketing mystique," and said that while it finds vulnerabilities at a record pace, the greater challenge is that organizations cannot patch all of them. On Zscaler's financial performance, Chaudhry noted that the company reported 25% revenue growth in its third fiscal quarter of 2026 and crossed $3.5 billion in market capitalization. He attributed a subsequent 28% stock decline to a "misunderstood" comment about new customer acquisition, and said the company expects total ARR and revenue growth of 16 to 17% for fiscal 2027. Chaudhry also announced the intent to acquire Symmetry Systems, a company that provides an access graph mapping identity and data connections, and said Zscaler is partnering with Microsoft, Google, and AWS for agent identity rather than building its own solution. He described AI and frontier models like Mythos as "one of the strongest tailwinds our business has ever seen."
Transcript (41 segments)
✨ AI-enhanced transcript with speaker attribution
J
Jagtar Chaudhry0:00
Firewalls are facing the internet and saying, 'I am here. Come and connect with me or attack me.' Every firewall is an attack surface. When I started Zscaler, we said firewalls have lived their useful life.
But the name Zscaler stands for zenith of scalability.
S
Sanjay0:33
Welcome to the CISO podcast.
J
Jagtar Chaudhry0:35
Sanjay, appreciate the opportunity. Thank you.
S
Sanjay0:38
Well, thank you so much. Jay, this is your fifth go-round in terms of Zscaler. You had many successful companies, but you also said that this moment feels a little different than before. For our audience, just briefly to kick things off, what is different about this moment?
J
Jagtar Chaudhry0:57
Technology incrementally changes all of the time. We expect technology to get better every year. Every decade, two decades or so, there have been a big change. I call them mega waves. Internet was a mega wave in the '90s. In the 2000s, we saw mobile and cloud came and changed many things. And now this AI is not even a mega wave, it's a giga wave. It's fundamentally changing everything. Cloud changed how applications are built and run. But AI is changing applications, all the use cases, security and all. So it is big. I compare it to the industrial revolution. The difference is the industrial revolution took place over a few hundred years. This is just happening every week, every day. There are some big innovations happening. So businesses need to learn how to move faster. AI is challenging every CEO, every CIO, and every AI officer to stay on top of it.
S
Sanjay2:13
Jay, when our audience knew I was coming here to talk to you, there were some questions from the chief AI officers. One of the questions was that Jay has had access to Project MyThos. What did Jay learn that would be helpful to these chief AI officers? What surprised Jay looking at that, and what would be helpful to them?
J
Jagtar Chaudhry2:36
So we have been part of the Glasswing program from day one, which was early March. We got access to it, and the mission was to go and check our own software infrastructure using MyThos to see what kind of vulnerabilities it can find, so we can fix them. Because over 45% of Fortune 500 companies depend upon us to keep them safe. So our platform has to be safe. We take cyber very seriously. I sometimes say we are always paranoid when it comes to cyber. We are never going to be complacent.
So what did we find? It is able to find a large number of software vulnerabilities that have been hidden for a long, long time in our applications, in the OS, in some of the open source pieces we are using. Now, I think MyThos is powerful, but the way it got launched, probably there's a little bit of marketing mystique around it. When you make it available only to a small group, it's goodness, but others wonder what it is. The notion of making it available to a small group is not new. When Microsoft used to announce those Tuesday vulnerabilities, they'd give us a week's notice so we could fix things. But this model is significantly better at finding vulnerabilities than previous versions. For example, take Opus 4.6, 4.7 — pretty good, but this is a few notches above. So it needs to be taken seriously.
But here's the challenge. Every CIO and CISO is saying, 'I need to get hold of Metasploit so I could work on it.' But here's the problem. Every organization has thousands of software vulnerabilities that are not remediated because there are not enough resources. Software always has bugs. Now Metasploit is going to add 5x more to the list. What are you going to do? You're never going to be able to just patch and patch and get out of it. Resources are limited. Change control windows and upgrades are hard. Vulnerabilities will be found quickly. Patching will be behind. So more breaches will happen. The question is not how do I patch, patch, patch. The question is how do I minimize breaches? Number one, what else can I do besides patching so I don't get breached? And number two, if I get breached, how do I minimize its impact so it doesn't bring my entire business down? Those are the discussions I'm having with many CEOs, CIOs, and CISOs. We have a very focused program to help our customers.
In fact, as I just got off the plane, I had a call with a CIO of one of the top 10 global system integrators. Same topic — what do I do? We have a pretty good game plan to help them. The actions can be taken in weeks, not in months. That's not to say everything gets fixed in weeks. That means if I can take low-hanging, high-impact items sooner, I'm safer. In cyber, nothing is 100% safe. But if everyone has a 4-foot fence around the house, and I can build a 10-foot fence, I'm a lot safer.
S
Sanjay6:28
So just very briefly, you're saying it'll detect a lot of vulnerabilities. But chief AI officers or CISOs don't have the resources. Is ignorance bliss, or should they still find out their vulnerabilities? What do you recommend?
J
Jagtar Chaudhry6:43
So they need to. These models, they don't need to wait for MyThos. Opus 4.7 is pretty powerful. Opus 4.8 just got released and is becoming publicly available. They can use it. Then ChatGPT or OpenAI is not sitting behind — they want to be in front of this thing. So they announced a daybreak program. Their model is GPT 5.5. These numbers start getting mixed up after a while — they change so frequently. So they should discover and try to work on patching. But that's only one action.
Here is the second most important thing they should do. How do attacks happen? Your applications, your resources, your firewalls, your VPN — they're all facing the internet. Anything facing the internet can be scanned and discovered by anyone. And now with AI tools, they can scan and find that this VPN can be compromised, this firewall can be compromised. Then they compromise those devices or applications. Once they get on the application, that application is connected to the rest of the company's network. Now they move laterally on the network. They get to your crown jewel applications, encrypt them, break them down. That's where the bad thing happens.
So our number one recommendation is hide your applications behind a zero trust exchange like Zscaler. The model we pioneered is the opposite of firewalls. Firewalls are facing the internet and saying, 'I am here. Come and connect with me or attack me.' Every firewall is an attack surface. When I started Zscaler, we said firewalls have lived their useful life. We need a new architecture. In this new architecture, you don't build a moat behind a firewall. We built an exchange, a switchboard, which assumes everyone is untrusted. Your applications are hidden behind us. When people try to come to us, we check them — are you allowed to access? If they get validated, we make the connection. Otherwise, they don't even know where the applications are. That is the biggest innovation Zscaler brought to the market. We call it hide your attack surface. If they can't reach you, they can't breach you. Our customers are already doing a lot of it.
That's number one. Number two is assume some user got breached or some firewall in a branch office got compromised. How do you make sure the malware doesn't spread from that machine everywhere? For that, zero trust architecture becomes very important. Zero trust says trust no one. Every employee is untrusted. Every branch is untrusted. So an employee, whether sitting in the office or at home, connects to our exchange, our switchboard. Just like the old school phone switchboard — you'd call the switchboard and say, 'I want to talk to John,' and you got connected to John and John only. Then you'd say, 'I want to talk to the president of the US or the prime minister of India,' and they'd say, 'Sorry, go away. You're not allowed to.' Policy-based access after checking who you are and what access you have. That's what we do with our model.
With that, even if a user got compromised, the worst that can happen is he can reach one application and do some damage — can't go anywhere else. If a branch network got compromised, the blast radius is confined to the branch. Doing that type of zero trust architecture becomes extremely important, and that's what we have been doing from day one. That's why we feel Zscaler was designed and built for this kind of moment. Maybe I should say it is built for two moments. Even though I started the company in 2008, knowing that breaches were happening and customers were spending more money on firewalls and VPNs, it wasn't working. The security model wasn't working, so we wanted to come up with a new model of zero trust. It took a fair amount of time to evangelize.
Then COVID came. Suddenly people had to work from home. That was the first moment that actually realized the value of zero trust widely. Now with AI in the mix for the whole world, people are saying you'll get compromised unless you can do zero trust. It's an important moment for us, and we are busy helping our customers.
So patching is one thing to do. Hiding your applications is number two. Embracing zero trust is number three. There's a number four actually — a very simple thing they can do. It doesn't take a lot of time and it's not very expensive. It's decoy technology — honeypots. If a bad actor gets into your data center somehow, they look for mission critical applications. They look for Active Directory. If I get into Active Directory, I can escalate privileges and do all kinds of bad stuff. Or I find your SAP deployment and do bad things.
So we have a technology called Zscaler Deception. Those deploys can be turned on pretty quickly. Even I can have a deploy for an AI model. When they scan and say, 'Oh, I found this company's AI model,' as soon as they touch it, the alarm bell goes on. We pick them up and they can be taken out. You can do the same thing in your hyperscalers too. If somebody broke into your data center, catch them quickly. These are pragmatic things customers should be doing.
S
Sanjay13:16
So I think Jay has given four great examples. If they can't reach you, they can't breach you. He's also talked about decoys, setting up honeypots. This is fantastic. Jay, another chief AI officer had a question for you. You said that an average employee is going to have between 50 to 100 AI agents across the board. According to this chief AI officer, there are two concerns. One is organizational: where do these agents reside in an organization? Second is what is the governance framework? How do you protect? You were just talking about employees. Now you raise the scale by 100 to 150.
J
Jagtar Chaudhry13:59
Today a user is the weakest link. Tomorrow an agent will be the weakest link. We talked about how they can act at machine speed. The numbers will grow. They need no coffee break, no weekends, no evening, no sleep time. If they get hijacked or hacked, they can be very dangerous. So what should we do?
First of all, let's understand the scope and size of agents. All agents aren't created equal. For example, you may have an agent that works in Snowflake database and spawns five more agents and fetches some information for you. I don't need to get in the middle of everything. There may be an agent pulling some information from your desktop. The key areas to focus on is inter-application or inter-company communication of the agent — agent going from application A to application B, maybe trying to reach a data source it shouldn't be reaching, or company A to company B.
For that, the solution is fairly simple and elegant for Zscaler. We built zero trust exchange for users, then we took it to branches, then to workloads. Now we are taking it to agents. In fact, at this conference, we are announcing our exchange for agents. Agents are somewhat like people — they're digital workers. We built an exchange for users. Agents are also code — they're like workloads. We have built an exchange for that too. We're extending that exchange where we understand the language, prompt translation, we built MCP and A2A gateways, we built prompt inspection engine.
Most agents are actually coming from user endpoints right now. We already have endpoint technology to make sure the traffic hits our exchange. Then we look at the policy and can say, 'This group of agents can only talk to this group of agents. This group of agents can only talk to this group of applications.' The underlying foundation for policy, for logging, and all is already in place. We need to make sure it can scale. But the name Zscaler stands for zenith of scalability — that was the big dream early on. Today, every day we process about 750 billion transactions per day. The number of stars in the Milky Way is 300 billion. The number of searches on Google per day is 14 billion. So the amount of scale we can handle is pretty good. We'll still have to do more work to scale for agents, but we believe we can.
Now here's another interesting challenge. The policy and governance for who has access to what for users is hard for a large company, but not impossible. There are a finite set of employees. Lots of companies sit between 5 to 20,000 employees. Those employees don't change every day. Marketing department has 2,000 people. They need a set of applications. Engineering does, and so forth. But these agents — they'll come, they'll go, here, there — and technology is moving so fast. How do you even figure out the policy? Who should have access to what? Once they go through Zscaler, we can tell you. But even before they come to Zscaler, they're already being used out there.
So we needed to solve the problem of who is talking to who in the enterprise. That's a very hard problem to solve. We looked around and found a company called Symmetry Systems — a number of PhDs from University of Texas in Austin started that company to solve this problem. You've got 50 applications sitting out there. You've got Databricks and Snowflake and AWS databases, GCP, you've got users, identity entries, all kinds of stuff. This solution pulls in metadata and telemetry from different sources. Using AI, it figures out who's talking to who — we're talking about billions of transactions. That's where it creates what we call the access graph. It shows you patterns of groups talking to what groups and why. That can tell us if unwanted entities are having access. It also helps us create and apply policy enforcement for which we have built zero trust exchange. First part understands it. Second part, by integrating Symmetry Systems with Zscaler, allows you to enforce it.
S
Sanjay19:05
Is this exchange — just for our audience — is it like an app store, or is it just an exchange that will have your governance, your rules, and everything? And is that individual exchange for each company separate?
J
Jagtar Chaudhry19:21
A marketplace is a destination you go to and bring applications down. We're not a marketplace. That's bringing software down. Think of the phone switchboard — it had to be everywhere. Or think of us almost like an international airport. When you deploy Zscaler, the traffic of an agent or user from your laptop gets sent to our nearest data center. Your traffic comes almost like to an airport. We check your boarding pass, your passport, your visa, your luggage — all to make sure that matches policy and governance. Can you fly to this country? Does your passport allow you to do that? What are you taking in your luggage? Is there dangerous stuff, weapons? All that policy enforcement gets defined. We are the policy engine for enforcement. That's what we do — we tell you what got enforced, what didn't, who tried to do what. It's inline policy to make sure nothing bad happens.
S
Sanjay20:38
And coordination with standards like MCP and A2A and other stuff is already there?
J
Jagtar Chaudhry20:44
So we use and support those standards. MCP is essentially a translation gateway — we support that. Now A2A is there as well, and we'll support that too. It's A and B. We are the Switzerland. That's why we need to be there — because we will support wherever these agents are coming from, where they're going to. They could be coming from Microsoft, AWS, Google, Databricks, or whoever. We are not biased towards applications.
S
Sanjay21:16
You're like Switzerland.
J
Jagtar Chaudhry21:17
We are like Switzerland.
S
Sanjay21:17
Thanks for taking the time. It's just been fabulous, and I hope you have a fantastic conference ahead, Jay.
J
Jagtar Chaudhry21:24
Wonderful. If I may make a closing comment — the role of chief AI officer is becoming extremely important. But they have to work with CIOs, digital officers, and sometimes AI and data officers are the same, sometimes they're different.
I think most enterprises have a lot of inertia. It used to be okay to write an application in 2 or 3 years. The world is moving faster. Overcoming inertia is not a technology issue — it's a people, mindset change issue. To be successful in your job, you need to bring teams along. You need to find some change agents in your organization to drive change, otherwise there may be five reasons to do something and somebody will find one reason not to do it. So drive it. Don't let security get in the middle of stopping you. Challenge because in AI you're doing things differently. Expect security to be done differently. Expect networking to be done differently. You can't be sitting still because the network side hasn't created a special network for this. Those things have to happen at a faster pace for all of us to be successful. Otherwise companies will be left behind.
Second quick comment. I talked to lots of chief AI officers. Some try to do these projects across the whole organization. Gets very hard. I tried to do that in my company — work with engineering, let's go use AI. They're all using AI and saying it's helping. Of course it's helping, they say. So how much? They'll say, 'Oh, I'm 20% more productive.' Then I say, 'If that's the case, are you going to deliver something you were going to deliver in 6 months — will you deliver in 4 months?' The answer is, 'Not really.' So where is the productivity? It becomes a little mushy.
A few months ago, I wanted some tangible projects with start dates and end dates. I gave them a project and said, 'This specific code they were migrating to a new platform — start on a clean slate. Tell me how long it will take.' Traditionally, they had come up with an estimate of about 5 months with 10 engineers. Then they started using a tool — Cloud or anything else out there. This was about 200,000 lines of code. In about 2 days, it moved that code which was written in C, C++ to a language called Rust. In about 2 days the system came up — it wasn't cooked, but it was working. In about 5 days or a week, it was operational. We were very excited.
Then I say, 'So when would it go into production?' I expected they would say in a week or two weeks. They said, 'We're not sure.' Why aren't you sure? They say, 'I haven't automated all the tests to make sure this thing is going to work in the new environment.' Creating the test infrastructure for all the stuff you're writing in AI becomes very important. So we're spending some time fixing that. Rather than doing it in 5 months with a team of 10 engineers, we think we'll finish in 5 to 6 weeks with a team of about 3 people. That's pretty good. I'll take that.
Putting some boundaries around showcasing small, well-defined projects become examples to showcase. And then other teams start looking at it and say, 'Let's do it.' If you are asking a thousand-person team working on a given project to change, it's not going to happen. If you create a ten-person team, give them this thing to go and do, things are happening faster. That's what I have been learning at Zscaler as we are embracing AI and developing more and more code in AI. In some of my new projects, over 80% of the code is written by agents. And the 20% is tweaking and changing and assigning the stuff. It is powerful, but it needs to be used safely and securely, and we're looking forward to working with our customers to help them embrace it at a pretty good speed.
S
Sanjay26:11
Sanjay, thanks again. Thanks for taking the time. This is fabulous and we'll follow up on this. Wonderful, Jay.