Cohesity let Mythos loose on its own code, and now wants your backups to feed your agent

Okay, folks. One thing that would be nice for all of you to also do as a takeaway after this: at the bottom of the signature of my email is a 10-minute overview of the company. Every six months we refresh it, and it'd be worth you watching it. It's in English, but you can translate if you know it's on YouTube. And I find for a lot of people who don't know — I mean, I think many of you have come here many years, but the story evolves. It's a very good introduction to the company's mission, what we're trying to get done. So Michael will send you a link to it. If you get an email from me, it's at the bottom of our signature. Many of my team also copy that same. It's a very quick way for everybody to get — if you don't know anything about Cohesity — what's an introduction to the... what do we do? Where are we taking the company in like a short 10 minutes? And we tend to update that. I feel it's important for the story of the company to be updated, you know, every six months. So you'll see that also refresh itself over the four years that I've been here. Okay. You're driving the slides. Yeah. Michael.

Okay. So if you think about our mission, it always starts with this. This probably won't be very different from what you saw if you came here last year or the last few years, but it's tweaked a little bit. We think of our company in three acts. The first aspect of it is protect. I mean, where this — the story, the category of the company is AI-powered data security. So when people ask me at a cocktail party, what do we do? We are responsible for protecting data and then getting insights from data. Right? So there's a security element to it and then there's an AI element to it. So you could think of it like a — you know, if you think of security companies, great security companies are like CrowdStrike and Palo Alto. Great, you know, kind of analytics and data companies would be companies like Databricks or Snowflake. The purpose of this company is sort of a join or cross between CrowdStrike and Databricks, or Palo Alto and Snowflake, that you should think about. So if you think about the amount of data that we protect today in the world, we have about 12,000 customers. Most of our focus is the Global 2000. We protect about 200 exabytes of the world's data. Just to put that in context: in the US there's this big building called the Library of Congress, right? It has all of the digital — sorry, now it's digital, but it's also the past paper documents of law and so forth. The United States — 200 exabytes is a million Libraries of Congress. One million Libraries of Congress. Okay, if we were to — so you look at all of our largest customers: they're banks, public sector, healthcare companies, telco companies, and tech. Those are the five verticals. Inside all of those companies, what's happening is that their data over time, their historical data gets sedimented in this thing called immutable backups, which we do. So we're just a historical copy of everything the bank may have if they have a retention-forever policy. So if they want to go back to 10-year, 5-year, even 30-day history, it's in our format sitting in there. It just gets sedimented. It's like soil with a time-series stamp on it. Right? So the purpose for us is to ensure that that data stays secure. We protect it. We secure it. But then if it's just sitting there in a sediment and you don't actually — it's gold to you, right? It's like sort of oil. If you knew that you had gold underneath your house, why would you just want to keep it secure? You'd want to use it. And that's in essence what we're doing. So this sort of three categories represents how we have put our story of the company. This represents how we categorize our products, which Vas will get into more into. And this sort of — the first place is where we started the company, right? Data protection across all of the key workloads. We think of that as making sure every workload in the enterprise is covered. What's a workload? A workload is something that sits underneath an application, right? An application could be something as heavy as SAP or as lightweight as an agent, right? Think of those as two ends of an application. And those applications sit on some infrastructure, on-premise or in the cloud. That's a workload. And our goal is to ensure that every workload is protected. Could be a virtual machine, it could be a database, it could be a file, it could be an identity, could be a SaaS application. Those are typically the five types of workloads we see in a company. We want to make sure that every one of those workloads are covered. So what's the reason that's important? Not covering a workload is like having a bank where one of the doors and windows are open. The bad guys get onto that application or workload, extract the data, exfiltrate the data, and then from there get to your entire network. So we kind of, you know, talk about like no workload left behind. And one of the things that has happened over the last 18 months since we completed the acquisition of Veritas is we now have the largest workload connectors. So the way you think about this — the problem statement in Act One is you collect the data from all these workloads. Those are — that's a pipe problem, right? You build lots of pipes to all of these connectors and then you store that data in an immutable location so it's extremely fast for cyber recovery when you have to plan for that type of event. And that really is a sort of — if you think of the core of our tech, it's a pipe and lake problem. You get the data from all these sources, you store it in a secure lake that's extremely safe, free of diseases if you want to think like a healthcare example, but you want to recover that data really fast. So the engine of how quickly you can recover. So the science and what we had built over the last 10, 12 years since the company was founded is the recovery of data is supersonic. Nobody in our competition peers can match our speed of recovery. That was the genius of what our founders built. And the reason they were able to do it — for those of you who know the history of us — was our founders had built this with the Google File System. That was the basis of a software-defined file system. Did a second time at Nutanix and the third time here at Cohesity. The tech is just perfected and we take great pride in ensuring that the speed of recovery of all of that data sourced from all these workloads is supersonic. Does that make sense? That is a key assumption. So if it's a VMware workload that's coming in, okay, or if it's a — now there's people who want non-VMware options too. Nutanix, Red Hat, OpenStack is getting a renaissance. For example, in European telcos, and many other sort of companies that are looking at this — AWS, Azure, Google, Oracle — there's typically four private cloud, four public clouds that we see. And in Europe, there are also now sovereign clouds being developed. All of those workloads, we extract the data, build pipes to all of them, extract the data extremely fast. That's all of what's involved in this sort of a small bubble here, but there's a significant amount of tech that's involved in that green bubble. Okay, data protection. Okay, once the data is in that secure immutable lake, you want to keep it free of diseases and you want — that's the way we run our security algorithms. We have a number of very sophisticated scanning algorithms. We have cyber vaulting capabilities and all of these you can either think of them in security. I also like to think of them in healthcare because my mom's a doctor. I'm on the board of Philips, as you know. And you can think of them in analogies — you know, when you keep some kind of thing sterile, right? You call that a clean room in a hospital. We call that the same thing in the way we run it — it's a clean room. And you want to make sure that the people who go into that, all of the materials that are there are extremely safe. If you're taking things out, you have got a way in which you do that. So that entire methodology...

Michael, do you have a picture of the five-step cyber resilience? Maybe you can bring that up. Follows a pattern. Just anything that shows the five steps. There they go. All right. Third, just put that up. I'm only going to use a few slides. Okay. Michael produced a lot of slides for me, but I'm going to use three or four. But put this up. It's coming. Okay. So, if you can go to that one slide and just like put it up so I can go and talk to it right now. 30. Number 30. There you go. 39. Okay. Okay. So look at this slide. These represent the five steps of how we think about keeping the data secure. Okay. That's that second bubble. We protect all the workloads. That's what I covered. We do cyber vaulting. This is making sure there's a clean room aspect of it. We do deep amount of scanning threat protection. There's a whole set of application recovery capabilities. And then this actually — step five could actually become step one. And Vasu will go into more detail about it. We pioneered this concept because many of our largest banks came and said to us, if the entire bank goes down, okay, how do we set up a minimum viable bank? We're working with some of the largest US banks, including banks that were the foundation to some of the other banks. Some of them have talked, for example, at our Catalyst event — Bank of New York Mellon spoke last in March, JP Morgan spoke at some of these. Big customers have spoke about this notion of a minimum viable bank and cyber resilience. We learned from them and then pioneered this concept. Now that has become a standard practice that we're able to do now. This is a combination of product, practice, and people. I call this sort of the three Ps, right? The product are all the things that come alongside our Cohesity Data Cloud. And secondly, you know, the process is these steps, how we outline it. It's not intended to be rigid, but it's a little bit like, you know, a doctor providing you both the medication — that's the product — but then a set of steps that you should do, right? What are the steps of cyber hygiene? Might be a good idea for you to brush your teeth in the morning, the night. And then the people are the people we bring to bear that can advise you on cyber resilience. We have an entire security event response team called cyber event response team. They act like the mini Mandiant. If you've heard this company Mandiant that got acquired by Google, we have a small team here that is able to run that practice in a way they can advise customers and constantly when they look out. And one of the things they do is if there's a major cyber event — we've had many of them in the United States — and if you're a NATO-friendly country, a democratic country, it's a matter of time. I tell everybody when I travel in Europe and Asia, you're also going to get hit because the bad guys are looking for any opening to find a way by which it's either driven by a nation-state actor or just gangs sometimes. And now in an agentic world, there's many things that can happen much faster. Okay. So remember this diagram. Okay. Now let's go back, if you don't mind, Michael, go back to the picture that sort of talks about the agentic process. Yeah. Let's go here. Let's come back to — yeah — site and the next one after that. Let's go. Thank you. Okay. So, what's changed since we last met? When were you last in? When did we do one of these meetings before, Michael? Year ago. Two years ago. June. June. June. Last year. Okay. So, it's been a year. So, around February or March this year, you know, Anthropic announced this project called Glasswing and they had an unbridled model called Mythos. So you know all the different models from Anthropic — Opus, Sonnet, now Fable. You also know the models from OpenAI. Okay. Leading up to Astra today and ChatGPT. And these two are rivaling each other here. Okay. They're all up in San Francisco. We've been working very closely with them. So I went to Dario, who's the CEO, and said we need to be included in Glasswing because many of our customers — there were two security companies that were involved in Glasswing: CrowdStrike and Palo Alto. There was a very small group initially who had access to Mythos. Okay. 10 companies or so. Now what ended up happening was that several of our customers were, you know, the early customers getting access to it — some in the US government and some among banks — and them also said, listen, our last line of defense is Cohesity. So if Cohesity has a vulnerability in its code, you know, we're likely to get affected too. So between my asking them and our customers asking them, we were the first company in that space to get access to Glasswing. And we basically took the approach — we took was to give it unfettered access to our codebase. We just said, listen, you run in our codebase and find vulnerabilities. Why? Because we want to protect our customers from those vulnerabilities showing up. Okay. Now what our — a couple of our customers could also do — I can't tell you which one — but a couple of them basically took Mythos, they had access to it, and ran it on our binaries. So we can run it on our source code, they can run our binaries. That makes sense because they have our code running in appliances, but they don't have our source code. We have our source code and binaries. Okay. And we wanted to compare what our customers were finding in binaries versus what we found in source code. Obviously you'll find more in source code than you find in binaries. And the goal was to quickly find those and become the first company in a space to address them, fix them, so we could go back to our customers and say we are Mythos-scanned. It's like basically saying if you knew a hospital in Santa Clara had the best possible cancer scanner, okay, and you were suspicious that a small little thing on your skin might be cancer, you want to go to that hospital, whatever the cost might be, right? And then you want to compare to another hospital. And what's the name of that hospital? OpenAI. So we went to both these companies. Go to the next chart. I think I show a picture of how we approach this. I mean, this just tells you — yeah, go to this slide — was there were multiple companies starting to do these adversarial approaches. So what's an adversarial approach? An adversarial approach means that you take an unbridled model and you attack the target. You attack the target really hard to find a vulnerability. And obviously you want to keep that agent or that model fettered so it doesn't go jump the fence, so to speak, right? Unfortunately, one of them jumped the fence and attacked Hugging Face, as you heard in the last few minutes. So those things can happen. But this unbridled model, Mythos, was very helpful to us. And this was all like in May and June that we began working it. The moment we got access — we were the first company in our space to get access to it. We used that and we've been announcing to our customers the timelines of where all those — those tremendous amount of privilege we feel being inside the, you know, kind of circle of Glasswing. And we've also now started to build product. So the first thing was having it run inside our codebase. And I was very keen not to have the whole company have access to it, but our security researchers — we have some very special security researchers — they had access to it to allow our entire codebase to be fixed. And now we're well on that way. We're also now actively building product on top of this model. So you'll see in the next few weeks and months some products that were announced that use that engine — unbridled adversarial security — to build out new products. We're not ready to announce it today, but you'll see that coming out very soon in close concert with many of these companies. Now I was keen to make sure we can't do this only with Anthropic because I don't know who's going to be the winner in these frontier models that are unbridled or adversarial over time. So let's go and do the same thing with OpenAI. Okay, at the time it was GPT-5.5, now it's GPT-5.6 Cyber. Okay. Our finding initially was it wasn't as strong as Mythos, but it was a close second. And everybody was saying, who knows what's going to happen. No one could tell me in 12 months who was going to be ahead, Anthropic or OpenAI. So I said, let's work with both. Huh? Or somebody else. Or somebody else. Exactly. Point taken. So let's work with both. And who was the somebody else initially? Potentially Google. So we went to Google's Gemini 3.5 Cyber and started working with them. And every place we could take an adversarial model, we unleashed it in our code to see what we would find. That's within us, okay, to be able to do. And we wanted to be — it's a little bit like a human being saying, I went to the best three hospitals, I looked at myself and I'm cancer-free. It's the same way here. We want to make sure our entire codebase went through that. We're the only company in our space, I believe, that's taken this approach to work very close. Now we have a sort of an advantage with Google because they're an investor in our company. We're able to work closely with them. But then a fourth company appeared that started doing some interesting work on that: Nvidia. And as you know, Nvidia is also an investor in our company. So we went to Jensen and his team, and he's been a good friend of mine for many years. And he said, listen, we support the open — sorry, the closed frontier models. These are closed frontier models, but we're really excited about open weights. Okay. Because there's many companies that want open weights. Many of the European companies want sovereign clouds which might be based on open weights. By the way, in France, Mistral is doing good work there. So we'll have to see how this plays out. So we became the first company in our space — I think the only company in our space — to support the open-weights AI model and the Open Secure AI Alliance. Okay. And you're going to see us very soon. They've extended this now to some new work in open shell. There's a lot that Nvidia is doing, but it's all premised around the notion of open weights. Now, what does that mean? The same unbridled approach might exist with GLM. That's the name of an open-weights frontier model. And we'll test that ourselves to see if it is as good as Mythos and GPT-5.6. That's on us. And we may build product on those too. We need to be flexible to ensure that we work with the best frontier model that could be used in an offensive way to ensure that our code is secure and the products we build are built on top of these models to help our customers. And that's a game-changer which we couldn't talk about last year. It's a major set of progress. And in our space, I track all of the key companies. I mean, certainly CrowdStrike and Palo Alto. I don't think any company in our space is doing as progressive work on security with the frontier labs as we are. And they are, to me at least at this point in time, four major companies. So I tell my team there are four consequential companies in AI that we have to stay close to. They're in this order: Nvidia, Anthropic, OpenAI, Google. Number one and number four are investors in us. So we have special closeness to them. They have a vested interest in ensuring we are successful. Not exclusively. They'll support others too. Nothing is exclusive in life, but we work with them in a very close way. And two and three are also very important. But we'll have to see — if I asked all of you, you're smart, you know, reporters and journalists — who's going to be the winner of these four? We don't know. And in the first example, the reason why I think Nvidia is very important is if they're supporting open weights and the open weights become responsible models. Responsible models. That's important. We don't want models that are being used for bad purposes. They're responsible models produced by any country, right? It shouldn't be just the United States that is able to play this — of a European company. So Mistral is a company example we're, you know, tracking very closely in France. Black Forest Labs in Germany. We encourage all the European — all of you from the European press — we encourage all the European companies to be at the frontier of this great research. And if it is, there will be what will emerge running likely on Nvidia right now — they are the leading hardware provider — a certified open-weights model that we will use. So question: you mentioned the four — Nvidia, Anthropic, OpenAI, and Google — but where is Microsoft? Microsoft will be next on the fifth. There never — never is the list always only four. If you ask me 10 years ago who are the cloud players I cared about, I would have said Amazon and Azure. Then they became a Google. Then there's an Oracle. Oracle. So the four is never a finite list. Yes, Microsoft will be next. Which is — and we're actually working with Microsoft's security model, but I haven't as yet seen it to be overtaking Gemini. They don't have a frontier model themselves. Yeah. Right. So they're dependent on OpenAI, but they could — they're working. We know — I was at the CEO summit with Microsoft. They've got some talented people there. So as they develop — now, by the way, this list is not just my list. This is what I'm hearing from customers. Okay. So the customers tell me, here's who we see. Obviously everyone's very interested in this topic because they don't want to be just tethered to a closed model. So open weights — that's why I put Nvidia on the top of the list. But in the enterprise it appears right now Anthropic is slightly ahead of OpenAI. But if I look at the recent Astra model versus Fable, Astra seems stronger. So we'll be watching this race very closely. We will build our products internally. We will also assess the strength of those models as they find vulnerabilities in our code and we fix — and we fix and are fixing anything we find. So it's a very, very positive statement when we do these types of briefings for our customers. They're very, very excited. What I think — when I just — I was talking to a European customer last week, a key bank, they were very impressed with the type of work we're doing in this area with the leading frontier models. So our goal with our product teams and our business development teams is to stay close to these companies. Fortunately, all four of these companies, you know, are here in Silicon Valley, two within a mile of our company. Okay, Google, I think, is that way and Nvidia is that way. Okay, so we're very close. Microsoft is in business in Seattle, but they also have an office here. But we have no problem. I mean, like everything I said, it's not to say Microsoft isn't a player, but if you ask me to add Microsoft to the list, they'll be number five. Okay? So, you know, as we do more, if you come here next year, I'll probably have to tell you there's five that I'm working with. Thank God the number isn't 50 or 500. Right now, we as a smaller company have to focus. Those are the four. So what else can I share with you? Maybe we can have — are you going to talk about agent resilience? Yes. Okay. Then you'll cover that. So one of the things we've done on sort of this topic of agentic is to ensure that we can not just, you know, use these models but build things that can protect these mini applications. I think of an agent like a mini application. It's like, you know, have you seen these cars that are the big cars like the 7 Series from a BMW and then you have a Mini Cooper? Okay, an agent's like a Mini Cooper, right? Doesn't put as many people, but it's an important car. You got to protect the car. Why would you have Mini Coopers without brakes? You have a 7 Series with brakes. You have a bus with a brake. SAP is like a bus or a tank. A Mini Cooper is a small agent. It has state. So the team here, Vasu and his team, they went behind this little agent and decomponentized it into all of its subcomponents. And you'll see this, and we just built backup of all of them, built connects to all of them. And now it's becoming one of the most popular things when we did a survey of our customers to say what is the area — we put out all the areas we're investing in — they were most interested in agent resilience. So you're going to see us put — we're shipping that shortly. That's a very exciting aspect of our R&D. And think of that as when you use an agent, we think there's a world where there will be as many agents as employees in the company. Okay. I was at a major conference where the CEO said we are at this stage. It's a major tech company. We have as many agents as employees. In their case they have a few hundred thousand employees. So I asked him, are these agents like ephemeral? Ephemeral means they use and they die. They're not permanent. No. No. They have state. Well, they have state. They've got prompts. They've got vector databases. They've got MCP connections. They've got all of this state. We should protect it. We should save it. We should sediment it. We should store it. I mean, that's the things that Cohesity does well. Why not, right? So we were the first to announce that. If they're sitting on a laptop, you know, there's probably state on there. There's a lot of things we could do to protect an agent that you'll see us doing more and more, which we're very excited about. The final thing I should share — do you have a slide or two on Gaia? Maybe I can finish on Gaia and then take some questions. Whilst he's looking at that slide, I've got two questions. One: did you know OpenAI Cyber found something that Glasswing — you know, that Mythos didn't? And also what about — I mean, open weights is not open source. You've got the Chinese models. Are you doing anything with those models as well? So the way we think about Kimi or GLM — these are all good examples of them, you know. Yeah, we want to make sure that they are typically running on a US cloud because I rely on AWS to do the testing. And so if Kimi is available on Bedrock, that's usually a sign it's ready to go. And I rely on the — I mean, by the way, AWS is also an investor in our company, so we have very good friendship with Andy and Matt there. So my first sniff test is if this is a legitimate model, has Bedrock certified it? Does it run on Bedrock? That's a very easy test. It's a little bit like, you know, when you have a new drug that's been produced by a company, I want to make sure it's FDA certified, right, before I use it. I'm not going to use it just — I mean, maybe if I'm dying of cancer there might be an experimental drug I'll be willing to take, but otherwise I want to have FDA certify. Right. So all Chinese models may not be good models, but if it's gone through that Bedrock and their equivalent harnesses in Azure and Google — Bedrock is the one we see most commonly used by our customers — we'll use it. So it's possible Kimi — I think Kimi's been certified in Bedrock. I think GLM is too. And then the other company we work very closely with is Nvidia. They would — between those two companies. Now the reason you may not get an objective answer from Anthropic, OpenAI, and Google is they have their own models that compete with them. Amazon and Nvidia have nothing to lose. They don't have frontier models. So okay, is that — and then did one frontier model find something the other didn't? I mean, Mythos is the gold standard, but did you find things in — what happens if you go to a hospital and find something the other hospital didn't find? You address it. Well, I mean, it justifies you using multiple models. Yeah, absolutely. So far, our initial finding at this point is Mythos is superior to GPT-5.6 Cyber in cyber. Now, remember these are cyber unbridled models. They're equivalent models in the world of like, you know, Astra and Fable are not unbridled models. They're ones that you can use in Copilot or whatever have you. Variants of those get used in cyber. Okay. So we're very focused in this dialogue now inside the company. We're actively using Anthropic and OpenAI and Cohesity. We're like, you know, I can talk to you about how we've changed the company. Everybody has access if you're a consumer to either Copilot or Anthropic. Copilot is obviously powered by OpenAI and I've got equal bets on OpenAI and Anthropic internally for my business users. By the way, everything I'm describing is for my engineers. They're using — and my researchers for security. But our coding now is done heavily on Codex, which is OpenAI's coding, or Anthropic Claude Code. We use both equally. And I'm constantly asking the engineers which one's better. If you ask me last year, they said Claude Code. Now they started to say Codex is getting a little better. We were using this harness called Cursor that's got an IDE and all that. But people have started saying, hey, maybe I want to just go raw on a command line CLI, command line interface without having an IDE. So we're watching those three. And if you track the spend of Cohesity on these AI tools, I'm closely watching how that spend is getting allocated between these — today for the casual user, OpenAI, Anthropic, and Cursor. Cursor's really for the engineers. So that's another metric of how — and it's transformed. I mean, our CFO, for example, has transformed the way in which we run the operations of the company to AI. So does it mean it impacts and changes the pace of the roadmap, meaning that you shrink? I'll give you a couple of examples, right. So first off, let's talk with the business user because every employee of the company — maybe we can end on this because it's a good topic to kind of talk about how AI is transforming our company. First off, I want to make sure all 5,500-plus employees of Cohesity are AI-empowered. There should be nobody who feels like — so if you're a consumer, you get Copilot and some sophisticated consumers may get Claude, okay, which is Anthropic's model. And now increasingly we're also going to go to OpenAI so you have both the ChatGPT and sort of core chat of Claude available. Okay. Now what I want to make sure is that I'm tracking their usage. I track how much they use. And remember, using one of these sophisticated models is like coming into a room with a very expensive light on. You turn — don't turn — turn the light off, I pay the bill. Okay? So I want to make sure people are using it. They're not. And by the way, I hope they're using it for their work, not for, you know, their cooking recipes and all at home. There are consumer models you can use at home. No problem. I mean, in my window, if you watch me work, I have two windows open. One's a, you know, Claude usage for internal stuff and I'm feeding it — I generally my adviser to how to run the company. I feed it documents. I'm asking, should I do this? Should I do that? It's just one more whisper in my ear. On my other tab, I have Gemini. Because my consumer life is there. It's free. I don't even pay for it. I want to see what the consumer — because I want to see what the common person uses without a payment. Okay? And I have both of these. Now, of course, I'm not going to feed into Gemini the company secrets, but inside my firewall use of Claude in the future. I'm open to either of these. ChatGPT — it knows me, strategic decisions of what we're making. But it doesn't mean that what they tell me is what we're going to do as a company. It's just one more adviser. When I'm sitting in my E-staff, you know, Vasu is one of the people at the table, but I have an empty seat. That empty seat is the AI, right? It also tells me what — why not. So if I have 12 people who report to me and a 13th person who's virtual, that's how I look at it, right? So I want everybody in the company to think of it that way. Trust it to give you a point. But don't take that advice blindly. I mean, I have 12 people who sit around my E-staff. The 13th person is one more adviser. It's my job to then collate what — but it's a very useful advice. Often it can give me a perspective of what's happening in the world, what my competitors are doing. That's all very helpful, right? This is a general, let's say, behavior from companies in your domain? I talk to CEOs a lot. I mean, now different people — I'm a techie at heart, so all CEOs may not do that, but I can tell you Satya Nadella, all the tech-savvy people are doing it. I'm sure Thomas Kurian, Satya Nadella, Andy Jassy — the people who are tech-savvy. You know, some people aren't. They may not be doing it as much. But I hope that every — it's not very difficult to do it. You just need to use prompts. What about Europe? When you travel to Europe and meet some, let's say, IT CIOs or let's say IT vendors, it's a software company, etc. Do they also — they may be using Mistral, they might be using Black Forest Labs, they may use different models. But they start in the same behavior. I had a group of Korean people who came here last week from Korea, a major bank. And I asked them — all of them were, you know, white-collar professionals from this major bank in Korea. They're always...

In ChatGPT or some variant of that. So I think the idea — many of them are young people. Some of it's age related. I think some of the older people are struggling to sort of find out how to use it. But if you're a young person, you're using AI in your daily life. So you need to bring that usage into your professional life and then realize and hope the company — right, most progressive companies are unleashing AI on their employees so that they can use it. Some companies are guarded, don't want it, but if you trust that within the firewall, those documents you load into it are not being sent — none of our documents that are in our private cloud uses of Claude or whatever is being sent to the cloud. We're very careful about it. And if we are ever at risk, we'll build a private cloud on our own AI GPUs here and run an old model on prem. We could do that. We want that to be the next step if we — it's not difficult.

We have a very efficient AI leader of the company in it who's watching this. We can tokenize and we can — it's a little bit like arbitrage, right? I mean, think of anything in the world as arbitrage, like oil. Somebody decides to raise the price somewhere, go buy it from somewhere else. So there's lots of ways in which you can do it. And the ultimate aspect, if you don't want to use a cloud, you can build your own GPU servers and AI and run the models on prem — may not be as powerful. Not every one of our 5,500 people need the bright light. Okay, they may be okay with a dimmer light, but it's still a light, you know. Now, the most powerful people are engineers. Now, among those engineers, we're watching very carefully the use of Cursor, Claude, and Codex because they are the people — probably they have the brightest light because they are the people who are using it at the highest level. And also our security researchers who use Mythos — they have the brightest light there. You think of those as the highest wattage because the intensity. Now, we're very fortunate — the early days of doing this, we were able to get some complimentary usage of these high-intense frontier models by the labs because they wanted to see — see, one of the benefits we can give the labs is we can tell them how good this model is on our code. We can also tell them — because it's in their interest to help each other get better — we can also tell them, hey, here's things we found with your rival's tool that you may also think about. How are you not finding these things?

Yeah. So each of these companies, if they're going to play in Europe, you know, if they want to follow what I just did, have to be comfortable selling inside France and Germany, who are EU, and UK, that's not EU, and then Singapore and India, a model where the data doesn't leave their country. That's not on us to solve. That's on OpenAI and Anthropic to solve it for those countries, as also Google. Right? In parallel, what I'm hearing from many of the European countries — I spent a lot of time in the Netherlands — is they're all building sovereign AI superfactories that may run a European model, whether it's Mistral or Black Forest Labs, and they may not be at the exact same level as yet as OpenAI and Anthropic, but they're close.

I mean, it's a little bit like — yeah, it's a little bit like this light wasn't exactly the same wattage. It'll get there. It's not like the Black Forest Labs, Mistral guys, and Kimi — and they may even use the Chinese models, provided some US company or an independent company make sure — see, the biggest worry about Chinese models is it's running in a China cloud, that when you host the data there it goes to China. That's what they're worried about. It's the same thing as sovereignty. But Kimi itself might be a very powerful model, as also GLM. It's when you use the open source, it doesn't send code, of course, if it's been scanned and you've classed it and it's been completely — and it's not — it's running inside your own containerized country or sovereign. And the risk is low. So which will be the actual model that every country uses in the world of the top 170 countries? I don't know. But I'm very sure Anthropic and OpenAI — I don't want to be a US-only company. So they have to solve sovereignty for every country where they can assure the people of France or Germany or UK or Switzerland or Norway, or then go to Asia. Africa, I think, they're not yet as much putting focus on, but I think that'll only come. And by the way, this is the exact same pattern of what the public clouds did. 20 years ago, you know, I was one of the first speakers of the first conference of AWS re:Invent in 2012 when I was a president at SAP. We were announcing HANA running in AWS. It was the first year of re:Invent. 2,000 people were in the audience, like 20. Now there's like 100,000 people there. At that time, nobody was worried about sovereignty because it was mostly a US phenomenon. Within 5, 10 years, I watched the dynamic at VMware. Everyone worries about AWS cloud. Do you have a sovereign cloud running in France? Otherwise, we're not going to deploy in AWS. They solved it. So I think the AI guys will also have to follow that. Cloud guys are doing this for generic workloads. These guys are having to solve it for AI.

I'm sorry, they what?

Oh, are they — how do they compare? No, they're not the same level right now.

Yeah, could be. I don't think in this world you should count any country or any model being a laggard. It's a function of their research and the function of the data that models it. Right? I mean, many of the reasons is these folks are running enormous amount of data sets on it. Chinese are running massive model sets. So Mistral does that. Black Forest Labs — there are two examples of companies, one in Germany, one in France. We're watching a lot of these. And then we ask Nvidia. Nvidia wants to support every company because they are just the hardware underneath all of them. Hey, which one are you seeing popular? Jensen goes around the world. He's talking to various different countries. So they have a very good model, and most all those companies come to them for GPUs. So they have a good sense as to how those models are working. But I'm optimistic in France. We work with Mistral. You know, I think there will be many winners in the space. What the ranking will be, I don't know. It's the same way that you asked me 20 years ago which is going to be the top cloud — said likely AWS. Today they're still number one, but I didn't predict that Google would show up at a very strong number three. Now they're now a strong number three, and Oracle's now starting. Guess what? Last week, you know who announced that they want to be a major player in the enterprise cloud 20 years after the cloud started? Facebook, Meta.

20 years after the cloud, Mark Zuckerberg thinks he can participate in that because he thinks he's got an AI workload setup that's better.

Yeah. To be the sales guy. He's the sales guy. CJ, I know him very well. So I would just say, like, I don't know. I mean, if I'm thinking Facebook, I don't know that a major bank that's been putting all their effort native on Google is suddenly going to go to Facebook Meta. But mid-market companies, all these companies, these small companies who do Facebook Marketplace, they may be willing to move some of their workloads for their applications also to Facebook. And you know, the world is made up of 10,000 big companies and 100 million small companies. And those 100 million small companies might be okay with it. So they're obviously seeing some opportunity in the business, the B2B side of cloud 20 years after AWS started. I mean, we're talking AWS started in 2006, circa 2006. 20 years later, somebody wants to be the fifth cloud. We'll see. So I can't tell you what's going to happen in the AI world. It's very — I mean, the AI world of these frontier models, we're four years into this game. 2022 was when this started to take off, right? So if you told me now who's going to be the 1, 2, 3, 4, 5, I can tell you right now we're very close to these four companies. I think Microsoft is a strong player to potentially be a fifth, but because Microsoft doesn't own their own frontier model in some sense, as does Amazon. They're also — you could argue — whereas Amazon's building their own, I forget the name of it. They're trying to build their own frontier model, but they would be also a fifth and a sixth on their own that could play in the space. So we'll watch it very closely.

What do you mean by target?

Oh, I don't know the European countries will let that happen. I think the European countries want their own — but I'm not here to speculate on — I don't know how M&A will work out. I think the European countries will want independence, and I think it may also be good for the US countries to just have their own — that's speaking just as a technologist. I think the mindset of the LLMs isn't to necessarily conquer the world with all the LLMs being acquired, but to allow flourishing aspects of this. But time will tell.

Yes.

Europe is almost naked except Mistral.

Not all — not all the Indian — the Indian population in the US is only — Yeah. It may seem like there's a lot of Indians here, but the Indian population of the US, just to set the record, is about 3 million. 1% of this country is Indian. 1%. Okay.

And 1.4 billion Indians there. So it's a small piece, but yes, in tech there's a number of Indians. That's true.

No, I would like to go back on my — how was Anthropic the time before you got involved in the Glasswing project?

Good. Anthropic is a great company. They're more enterprise and B2B focused. Initially, OpenAI was more consumer. Now, OpenAI is hiring many people. You've heard they've hired Dali, who was the former head of sales at Wiz. So both companies aspire to be B2B.

They were focused, as were all of us, on what would help joint customers. What benefited us was the US government wanted Cohesity to be in the list because we have some good strong customers in the US government, and many banks wanted that. And that's how we — and we're a major security company. If you think about — I think three important security companies, many more than three. CrowdStrike protects endpoints, Palo Alto protects firewalls, Cohesity protects data. We're the leader in that area, and our largest customers are some of the largest customers also of CrowdStrike and Palo Alto. And because our customer base is so heavily — right, 200 exabytes in the world, the biggest brands of the world rely on Cohesity more than any of our competitors. The biggest banks, biggest hospitals, the biggest — we have the biggest logos. That's the benefit of being in the enterprise. We're not a mid-market company. Some of our competitors are mid-market companies. We're focused on the enterprise. So those customers are also saying — see, imagine we are protecting the backup data, which is all of the historical data at a major bank, and we have a vulnerability of those that was Mythos exploited. If the bank gets hit, they won't get their data because Cohesity doesn't work. So we are literally the last line of defense for that bank. So the bank says, listen, if we are running CrowdStrike and Palo Alto through Mythos — they were in the Mythos, in the Glasswing program already, slightly ahead of us, and they're okay, but Cohesity isn't, and we're in problem. So as much as we may not have the public profile as a public company of CrowdStrike and Palo Alto, our customers view us just as important. One's an endpoint, one's a firewall, the other's data. So that's, I think, what we do. And all these companies — you make a strong case for why a joint customer wants it, and we find them very listening. They're up the street in Anthropic. We have a very good board. Some of them are very strong cyber people in the world like Kevin Mandiant. So we can use a number of connections. Does it mean at enterprise level when a company picks one of these three —

The other is immediately associated because you have a strong connection, I mean, between —

Cohesity.

And CrowdStrike.

I'm sorry, what's the question?

Yeah, we have a security alliance with the top companies. You know, we're — by the way, we're not competitors with CrowdStrike and Palo Alto. They're complementary companies. One's endpoint, one's firewall, one's data. But we go to our largest customers, we ask them — this is the question I ask all of our largest customers. Remember, we are, you know, 70% of the Global 500 are our customers, okay? So we go to them and say, who other than Cohesity are your important security companies? And it's usually one of these two. There are others too, important. Okta is a good example in identity. But so if you ask me to extend the list, there's a few more, but typically those are the two. Sometimes it may not be Palo Alto, it might be Check Point, but Palo Alto's market share in firewalls and network security is very high. Same — sometimes it may be CrowdStrike, it might be SentinelOne, maybe Microsoft, but their market share in endpoint security is so — similarly in data security and data protection, we are the highest market share and the most prevalent with the big companies. That's the key thing, right? Cohesity is now associated with the biggest companies relying on us for cyber resilience. So we have a lot of clout with those customers to make sure that they are successful.

Yes.

Yes.

Michael, put up that slide. Put up that slide that shows the NIST framework with all the attacks coming out of that picture shows. Yeah. Yeah. They have no locals.

The way to think about it is now that when he brings up the slide, the attacks that used to come in would come in at some sequence, right? They were nation-state actors. They were coming in, you know, and typically NIST is this very popular security framework where the left side is you detect and you protect from a — and then on the right side of it you recover. Okay, so we're very focused on the right side of this. Detection and prevention is on the left side of this. Okay, so Palo Alto and CrowdStrike are good examples of companies that detect and prevent, right? Recovery is all on the right side of this. So we assume that all these attacks are going to get in, and you're the last line of defense of data. So what's physical security? These are like the badge systems that you open your building. Network security is parallel. Endpoint security is CrowdStrike. Application security is ones that cover code. But once you get through all of these layers, you get to the data. That's us. Okay. So agentic attacks just are coming at you harder. They're coming at you harder, and they're potentially more deadly, right? So it doesn't cost as much to create an agentic attack. You could — I mean, you could be a nation state, but you could also see gangs using it. So I think there's going to be a surge in these types of agentic attacks, like we saw in the OpenAI Hugging Face. Some of them obviously were unbridled ones that came out of the labs, in that case OpenAI, but you could, you know, use an open-source unbridled example, open weights, and create something from a gang sitting in some country in Africa or the Middle East. Who knows? That's possible. So we say, listen, part of our proposition is say, assume you get hit, okay? If you get hit, that five-step cyber resilience, we will teach you how to recover really fast. So our posture at these companies: assume you're going to get hit, assume breach, and then let's build you. So it's a little bit like, you know, here in California, we deal with earthquakes a lot. So everybody, whether you're living in — this building is five or six floors, but if you go to the city, the Salesforce Tower is, I don't know, 40, 50 floors, maybe much higher. Whether you're in a five, six-floor building, one-floor building, or 50-floor building, you practice earthquake preparedness just because you never know when it's going to hit you. You don't want to be in a place where, oh, that happened. Guess what? The elevators don't work. You got to use the stairs. All that practice you do. So we assume he's going to get hit, and then practice how quickly you can recover.

So an agentic world just makes that recovery practice even more important because it's going to come at you faster and it could potentially be more deadly. What if an agentic attack attacks a nuclear plant? First stage was maybe just a hospital, and all the — or maybe in a hospital where all the surgical equipment in the intensive care unit is disabled and you can't use any of that equipment for life-saving. All of these are very possible. All of that equipment in a hospital is now on the internet. They're running on some operating system. Often it's Microsoft. Okay. Could be attacked. Anything on the internet, whether it's a nuclear plant, whether it's a hospital, whether it's critical infrastructure — power plants, water plants, chemical plants — it could all be attacked. So we obviously want a safe world, not just for the United States but every country. Even a nation-state country — you think North Korea wants their own people to suffer from a headache? No. They may plan attacks in the rest of the world, but they want their — so everybody, I think, wants a safe world. Nobody wants to have their own people die. So I think like this is a good time for the world to basically say, hey, this is the way in which we think the future could be secured. I wrote a blog when Dario wrote out his blog about pacing the enterprise. You can all read it, and Michael will send you a copy. You know, he said pacing the enter — and my title of it was, pacing the enterprise is, you know, important, resilience is urgent. Okay. Do you need a pace? In other words, to me, pacing the enterprise is like putting a brake on a car. But you can never run a car just with a brake, otherwise you'll never leave the parking lot. So the idea that we're going to suddenly stop AI from innovating is a false claim. That's the accelerate. We got to keep innovating. But it's our job to ensure that we can recover really quickly. We have a whole set of things in the car. There are seat belts, there are airbags, okay? You know, we ensure that, God forbid, if you have a crash, you can recover very quickly. The likelihood damage happens is low. That's what we do. I don't think the world needs to be just in a place where you're living in fear that AI is going to destroy humanity. That's our view. So I wrote a blog on that, not intended to combat against Dario, but to basically educate the world that while you're pacing the frontiers, that might be important, you know, even more urgent is to ensure that every enterprise has a resilience strategy.

That's called air-gapped cyber vault, and yeah, many governments do that. If you look at the DoD, the Department of Defense in every country, they have a completely isolated environment. They do that. But not just the US government. German army does that. Israeli army does that. But you can't run the world without a connection to the internet, right? It's very hard. World get — Yeah, exactly. You said that right. So several secure networks may need to have that as a special part of it. Certainly in the way in which we do cyber vaulting, we do that. But yes, you can't isolate yourself from the world, right? You can't live in a bunker completely. So you can have ways by which you get the best of both. You put your data, a third copy in a box. That's our product called Fort Knox. It's a cyber vault. You connect to the internet, you get the data, the third copy. Your first and second copy is inside your network. The third copy is disconnected. So if something happens to your bank, you've got a third copy that's in a secure bunker that's completely disconnected. That's going to air — that's a best practice we have in Fort Knox that, you know, Vasu can cover in more detail. So we adopt that type of philosophy, air gap, for some of the things we do. But even our most progressive companies, whether it's the banks or defense, have to have an internet connection for many of the things they do, right? They're sending wires. I mean, that's all on protocols and networks.

Okay. Well, maybe one or two more questions, and I'll turn it over to —

I mean, it's a very astute observation, right? I think we are in our space one of the critical companies that in security is important to our company's safety. So we take that very seriously. We have a high standard for that — that's why we were the first in line to go and make sure our code was scanned. Mythos, and we take that obligation — that's more important when we evaluate our product. I tell these guys, Vasu and a couple of his peers, through his other peers, there's three frameworks by which we apply our way: quality, time, features — QTF. It's in that order. All the security work we do is quality, ensuring that product quality is high. That's more important than time. If we're going to ship a product late, we will do that to ensure it's got these vulnerabilities fixed. Time is more important than features. Okay, in other words, if I'm going to say I'm releasing a product in a particular time, I don't want to be late. So we use that framework, and we take that responsibility very seriously. What they will do, the labor market, time will tell, right? When a person exits from this company, meaning they leave the company, we can either replace that person with another person, or we can replace that person with tokens. Which do you want? Often we need both. But we know the cost of a person, we know the cost of tokens, a fully loaded token cost for a person, and we can balance the cost of new employees that we need to add. With many of our most sophisticated engineers, I watch some of our most smartest CTO folks how their coding was four years ago when I got here and what they're doing today. In the past, one of our top engineers, he's in Portland, he used to code and they have a couple of junior engineers who do some of the work, sub-parts of it. Now, those are all agents doing some of those long-running paths. So I think there will be some tempered aspect of how we grow employees. The one thing I can say for sure: if you are in any part of our engineering team and you're not using AI, you're going to be left behind. You're absolutely going to be left behind because you're going to be replaced by somebody else who does use it. So that's a wakeup call to every engineer if you're kind of backwards about this. And there's certain departments like quality — QA, quality assurance is going to be completely transformed to AI, right? In the past you did all these automation tests to test, and AI can do that. So if you're not adopting AI, I mean, I tell them openly at company meetings, your job will probably get replaced — not in two or three years, in 3 to 6 months. So learn this quickly. So we're finding as we challenge people to do it, some people adopt, some people don't. And, you know, but I don't fear in society. I do think that there will be, just like, you know, every part of the industrial revolution — from steam engines to electric, from horse and buggy to cars, to planes and whatever have you — there were displacement of jobs, but the world as a whole became more productive. If I look at some of the thought leaders in speech who also talk about it, like Elon Musk and Jensen and others, they all believe the same thing. The example that Jensen uses all the time is the one I'm familiar with in Philips. He uses radiologists. He says, you know, in hospitals, people thought that these AIs were going to remove the jobs — radiologists actually made them more radiologists needed because more people are coming into the hospital for scans. And you need more of them. So it's not like radiologist jobs — they just got more productive. That's a good example that everyone could relate to. So how that will play out in the world, I don't know. Listen, a Waymo — if I don't know how you got here, but if you got here in a Waymo — the Waymo is displacing an Uber driver, no doubt. But I don't think the entire world is all going to be Uber drivers. So those Uber drivers of the next generation may not be their parents who were all Uber drivers. The next generation probably has to find a new role. It's the same way in — there's a state in the US called West Virginia, okay? Most of the people there 100 years ago were miners. But the next generation went to college, and they're not miners. So we have to help society realize, listen, if your father was an Uber driver, maybe now with Waymo — I don't know, maybe you study mechanical engineering and you're helping design a robot. So we have to evolve society, and I'm very optimistic that we will. And every technological shift is only produced — I mean, think in your world, right? What did people do in journalism before a word processor? They wrote in a typewriter. Did the spell-checking and the grammar-checking — and a word processor displaced the job of a journalist? Your role evolved. But if you were, as a journalist, still writing on a piece of paper and not using technology, you're probably getting left behind. 50 years ago, when the word processor came out, everyone who learned how to type learned a word processor, and they started using spell-check and grammar-check. And now — for you, you know, uh — stop right on the — It's good. He's got both. He's going to use an AI that takes a picture and puts it onto a — into a —

Julio, he's using a plot there. So he's doing —