And we're back in San Francisco. We're here at RSAC 2026. You're watching theCUBE's live coverage. My name is Dave Vellante. I'm here with Christophe Bertrand. Jon Oltsik is also here. He'll be joining us later on. He's out doing his reporter notebook. He's going to come back and tell us what he's seen on the show floor, but we're really excited to have Bipul Sinha here. He's the CEO of Rubrik, good friend of theCUBE. Great to see you again.
My business friend, Jeremy Burton, I don't know if you know Jeremy.
He said something to me years ago. He said, "Don't fight fashion." And I look at the way you've positioned Rubrik as a cyber resilience platform and now you're sort of moving into AI operations. Very fashionable. You've been right a lot. What led you to these decisions?
I always say to my team that we should always be in love with the market, not with the technology. Because what happens is the market knows what market needs and market is always changing. When we started Rubrik, everybody was doing mobile, social, local, Uber, but we identified backup and recovery as a market that needed disruption and innovation. And we built our product and the company started on that basis, but the biggest disaster was cyber disaster, not human error or natural disaster. So we built our platform purpose built for cyber recovery, cyber resilience. And if you look at the world today, the world is going through AI transformation. The old paradigm of cybersecurity is dead, because it was built for human and human operator and human attacks. So world needs an AI governance engine, world need an AI security and compliance platform. And that's what we are delivering so that our customers can deploy agents with confidence.
So the pivot to cyber resilience wasn't actually a technology pivot. It was maybe a positioning change. Is that right? You were always built for that?
We were always built for cyber resilience. In fact, in 2016, 2017, we created a product for ransomware detection. And when we went to sell it to customers, what we heard was we don't have a cyber problem. We are looking for regular backup and recovery for natural disaster or human error. So we kept advocating about cyber recovery. We kept building our cybersecurity products. We built data classification engine. We built the scope of attack issue. Then we built data threat engine, but the market was not ready. And lo and behold, in 2018, 2019, when the ransomware started to happen at a scale, Rubrik was the only game in town in terms of a unique product purpose built for that moment to deliver fast cyber recovery.
Okay. So you were a little bit early maybe. You don't want to be late. You don't want to be way too early because there'll be no business there, but you kind of timed it, right? I mean, it's good to have foresight and a little bit of luck. Is that fair?
100%. You have to be two miles ahead of your customers. If you're 200 miles ahead of your customer, they'll write you off. But if you're just two miles ahead and showing them that you understand their problem, but you also anticipate what is around the corner for them, that's when you establish a long-term relationship.
So I have a question for you, which is a perfect segue. Is this now the time of cyber resiliency or cyber resilience for AI, or is it AI for cyber resiliency, or is it both?
The way I think about cybersecurity is traditional cybersecurity was built to prevent and detect attack, but you can't prevent the unpreventable. And now with AI assisted attack and AI acceleration, we are in a singularity moment where you cannot stop attacks as much as you try to stop it. So every business has to think about fundamental resilience. The second issue is with geopolitics, nation state actors, hot war and attack on data center or AI infrastructure, now we are truly living in a world where resilience is a national priority. Resilience is a societal priority. Resilience is a business priority, both in terms of digital infrastructure. So AI is a new technology wave and just like all the digital infrastructure requires resilience, AI also needs resilience.
Right. So our research shows at theCUBE Research shows that the nascent AI infrastructure is actually not well protected. Most organizations do not back up most of their infrastructure. So that's one aspect. The basic old backup, are you backing up this new application, workload, whatever you want to call it? And it's very typical. We've seen that many times through the years. New forms of technology come out. They don't get backed up because people don't wake up in the morning thinking about that. It's not about backing up though. It's about recovery. So that's number one. Number two, what we're also seeing is the AI infrastructure itself is being attacked. Literally, people are trying to go poison models or steal data or steal models, steal IP, or even negate all of the efforts by blowing up the actual investment. What are you seeing? Where are we in terms of the type of threats you've seen, what you've been able to help with? I'm very curious about that.
There are three aspects of AI protection. One is the AI infrastructure and data protection. That is no different than the traditional data security, data protection and recovery. So that's like the basic hygiene. You need to make sure your data has integrity and data has availability, otherwise you have garbage in, garbage out problem. Second piece is, how do you ensure that models itself are protected from poisoning and other attack? And then the third aspect is now as you are doing agentic work, agents have two problems. They hallucinate and they are also prone to cyber compromise. And if agents are compromised, then you have a semblance of an insider attack where somebody can actually run your business processes. So what we are doing is we are definitely doing making sure that the agent data and infrastructure is cyber resilient. We are ensuring that things that go in and out into your agents and model, they are kosher. And then the third piece is when agents take an action, is this action in compliance with your policies dynamically? We just announced Sage yesterday. Sage is the semantic AI governance engine where we are actually using AI specifically tuned and trained on your policies to be the guardian AI to judge agentic action. This is one of its kind industries first, guard railing AI products.
So Sage is what, a deterministic complement to the probabilistic agents, or is that right the way to think of it?
So Sage is a small language model specifically tuned on your policies so that it can judge where the agent action, whether agents are hallucinating or intentionally taking bad action or cyber compromise, dynamically stop bad action. So for example, Dave, suppose you have an agent that is supposed to send emails that you don't want the agents to send an email to everybody in theCUBE all the time. So what you want to do is although agent has the permission to send email, every time you send email, you look at the payload and say, "Do I allow it or not?" And that's a new dynamic control that only AI can do because it is about understanding agentic intent that old cybersecurity teams that is doing a rule-based policies can never catch.
So you built a semantic layer, essentially, a purpose-built semantic layer for your particular use case, customers?
That is actually the judge AI, guardian AI, to watch your agentic action.
And how do you think about these LLMs, the four big frontier models that are so powerful? Do you use them to train your small language model? What's your relationship with those?
The core issue is that if you're using the large frontier model to judge agentic action, then you have double, triple the cost of every agentic action. You need to have a low latency guardian model that is very, very accurate on your policies. And this is where a small language model comes in. And this is what the product that we acquired with Predibase, those guys were doing model fine-tuning and GPU virtualization to multiplex different model and reduce the cost of inference. So we took that technology and then purpose built it as Sage to be the guardian model to judge agentic action.
And your cost per token is so much lower.
