Enhancing Cybersecurity with Embedded AI – Gopi Sirineni, Axiado

This is Textron tv. Hey guys, thanks Ventre. We're here with Gopi Sni, who is c e O for Oxy.

They're a company that is driving AI and cybersecurity into the hardware because well after years of all this stuff running up in the user space, it may be time to rethink where this stuff really shouldn't run the most optimally. Gopi, welcome to the show. Thank you to Michael.

Uh, thank you for bringing, uh, giving us an opportunity to work on this one and, you know, talk to your audience. Definitely sure we can talk progress. So walk us through, I don't think a lot of folks know what you are doing per se.

We understand that there is AI and we understand that it can be accelerated in hardware, but how does that play out in cybersecurity? So we specifically do, as you talked about, everybody's been doing, is an application level on AI's been used. So we come into the hardware.

So we're not replacing any of the cybersecurity software solutions in the market today. We're actually augmenting. So lemme give you just an example.

If you look for cybersecurity, there are thousands of thousands of companies already trying to solve something and you know, cybersecurity starts from, from an application level to all the way to how you penetrate into how the bad actor comes into and all that. And most of these softwares, you know, addresses different variation. So what we trying to do is, or we trying to make it simpler so we address a ransomware, you know, protection against detection and protection against any of the ransomware attacks happen.

2020 5% of the attacks in the markets are ransomware based. And what everybody does is just let me explain what the ransomware and then trivialize this one to make it easy. And then I can go further what exactly our solution is.

Ransomware means a bad actor came into your system somehow, however way he came into, he comes into your system and in a simple way, he takes over your system. Now two ways to take over. One your hardest encryption, which is your content, and the second one is to take over the system as a super user.

So lock up your system, ask for ransomware. In both cases, unfortunately there's nothing you can do other than pay the ransom if you care about the platform access back to you or content which is there on that. And most of these solutions in the market are software driven.

There's something called port of entry means bad actor coming into your system to a a kernel or a user access and application driving a server, driving whatever that is running on, on a platform, on a physical access to that also. So he came into, let's, for the sake of it, call it the port of entry. All these software are trying to stop a port of entry and everybody says that right now.

They go by something called a immutable route of trust means every software, whether it's a Palo Alto, whether it's Snowflake, CrowdStrike, you look at it, all these guys are Zscaler. They'll talk about they use a zero trust model and also use a, a root of trust or hardware root of trust. And they use a word immutable.

Unfortunately, if the root of trust is compromised and this no longer is an immutable situation and none of these zero trust models work, that's exactly where we come into the picture. We give you a true immutable root of trust protection that it's not compromised. And also we give you all these zero trust models, pretty highly augmented and more reliability, et cetera.

Let me pause there and then we can go further in the details. How that is implemented, Is this evolving into a game of latency because we have all these security tools, but to your point, a lot of havoc can be wrecked immediately. So we're basically then in a game of trying to limit the scope of the breach.

So are you really getting at the point where if we drive a little further in the hardware, we can respond to these attacks in a way that uh, contains them more aggressively? So it's not on the latency side as I talked about is more specifically addressing the ransomware means a system control and management side is the one. So we're not on the data paid data, data path to stop anything.

We're on the control and management side means we're the first one to boot up on the system, make sure that check the system properly and it's secured or protected, et cetera, tested authentication, all that. And any applications are running, they get an attestation from us. So we're not on the data inspectors, but we are on the think like this way a thief coming into the bank to rob her and he can come to whatever the doors he's coming through, he'll turn off your security cameras and all that, all that people stop at each level.

The cameras are checking, the thief is coming in, et cetera, but end of the day he has to come to your, you know, money chest, which is here is a hard disk kind of platform. So we are at the bottom. We don't see the port of entry where they're coming.

But when they're coming and taking over exactly the action happening, that's where we stop. We may, we may be the only architecture detects while the ransomware attack is happening, not the after the fact and then protect you right there itself. So in this way we're not on a any latency path of anything.

So we are actually on the underneath of everything watching. And another example of let's say if you think about a building, the ground floor is the hardware for us. So it doesn't matter how the bad actor comes in multiple floors in here, operating systems and stuff, we will be there to monitor those two stop.

So it doesn't add any latency for us and completely watching as a external process. That's why the external process solution for us compared to what is an X 86, what you just talked about is an x i six out of the host process. When you are running on a software, that's where you add a latency.

But if you are completely outside, you know we call an out-of-band management, out-of-band controller management. Think like you have a security cameras outside the building to watching you versus inside the cameras. It's too late.

If somebody had already came in camera as detector but the bad actor already came in. But versus I'm watching from outside can tell you before you even knock the door or trying to breach the door. Mm-hmm.

That's the difference for us. Once the attack is detected, then what's the remediation process that you guys kick off? So there's a multiple setups you can use in ops.

We are a silicon and a card provider to a system. Let's say if it's a Google using it or a Dell using this, our solutions in there, they can set it up how the action should be so we can detect not only the, a physical access by somebody based on our profiles and AI engine, neural engine learns the behavior for you. If any only happens like example I talked about bad actor trying to tamper the door outside today.

You won't know until he comes in, breaks in. But we'll know based on the behavior that somebody's trying to attempt on that we can stop that. It doesn't need to be just a physical access, uh, content.

What he's trying to do or is able to come up in a solution trying to come up inside. Whether it's an operating system, whether it's in a driver, whether it is in a particular application you're running any of these we can even a container, virtual container inside the container also we will be able to tell. So we'll detect those things after detection.

It's an O E M or A C S P to what they want to do. One, you can log the file push into the cloud so you don't want to stop it, but you know that data collection happening. Two, if that matches to the bad profile we already trained our A engine to then it will be able to stop because we know that's a bad pattern.

Our bad things are happening there and that's two and the three, if we don't know means we detected the N M E and we matched to bad profile that didn't match to there's something still ally there. Then you can run into the cloud to match to the new ones with our, you know, generate generative AI engine match to outside to see if the latest one, if it's not the one present, one I already trained, but something else what has been talked about, we can match it to. And then if it fits that match we will do stop it.

If it don't, we log the file, push it to the cloud so that you don't take over, then we have something to come back. So the protection, the stoppage and the detection, you know is our job stoppage and mostly by the O E M job part, the user jobs. How far away are we from seeing this chip in hardware then?

Is it starting to be built in or once of in two Weeks. So we've been sampling our chip, uh, with our solutions for last four months in a two weeks. Uh, in our O C P conference.

This is the go O C P global summit in San Jose Convention center is around 4,000 people comes up for those we'll be showing around six or seven customer boxes powered by us. Is it your sense the bad guys are using AI already? I mean where are we?

Is this evolving into an AI arms race? Uh, it's obvious. You know, the first first person to use always is bad guys, right?

So it would be the one, so the brains or computer doesn't matter. You create a computer engine to somebody, it definitely bad actors is also smart to do it. So these generative AI I G D type applications, it gives you kind of as you talk about combating on each other, who is outsmarting other uh, here what we're trying to do is take the human away from it.

The machine learning and neural puts on the platform, not just on the cloud. The difference for us is each platform has a brains to protect by itself as a sideband carrier rather than in ice. So the bad actor can only run based on applications you can provide to the guy.

Means that's on X 36. We don't have any applications running other than ourself on security. And that's why we are, you know, kind of outsmarting and also can stop any futuristic, you know, zero day attacks based on the what are what we already taught to it.

Then you'll be able to predict a futuristic, most of the attacks we use as zero day, but technically they're not a zero day, they're somewhere completely nine out of the 10 of them is deteriorated out of what we already know. It's just a human does not know right away. It'll take some time to make it to be zero day to hey we know this is somewhere 2017 found out.

Whereas a machine can learn that AI engines where we can do it pretty quickly can tell and stop those prediction. If seven out of the 10 things matches to that, you can protect it. Okay, let's stop it.

If it's three, yeah. Okay. Put the log file.

So that's neural learning for a futuristic, we outsmart what is, what has been already known based on that expansion. Am I protector for next 20 years? Maybe not because we gotta continue to evolve with the bad actors evolving.

So it's a, it's continued to be on that way. So, Well hindsight is always 2020 but looking back at it, did we kind of overlook the whole hardware approach to the cybersecurity problem over the years and now we're kind of going back in from a hardware architecture pers perspective to address it. It so you, you hit it nail on that.

That exactly right one. So obviously everybody, if you look at the world, if you talk about there's eight out of the 10 people in the technology world, they'll talk about a biggest AI engine. They can do thousands of millions of courses.

They can compete together. How fast they can do all this stuff, including everything is running on the cloud and everything on that. How big of the data pipe and all that.

But all of them are protector or not. We always do later. And whatever you build, you build for the excitement then you trying to protect.

And fortunately that's normal to human behavior. This is exactly that. Everybody ignored this word that everybody built because as those are the needs.

2017, the a w s came up, everything went on the cloud. Great because I can throw a lot of computing engines and all that. I can do a lot of computing easier.

That's a great use case. But unfortunately if that's not protected, you are providing all the data into those clouds and the expecting those engines all are secure to run and give the compute results back to you. So there is something, an expose on that.

So it's always that way. It's not a bad, bad thing but that's just normal innovation cycle happens. It's a time that we need to push back to the ground up, build security.

That's where the, we may be the first ones. Xito may be the building, the TCUs the first one to do it, but there will be, you know, 10 comes up in that. It's like if you, you look at the search for thousands of software based cybersecurity companies including the ransomware protection, all that too.

Also that's because that was a critical issue and now we see that's not enough to stop it because we are computing the same machine example I gave it to that we do it, the bad actors are also has those access to the tools of those. So we need to counter that, that you can do from a completely day one turn on the power, a first, second system comes out. It need to be secure to do that.

That's where we come into the picture. Mm-hmm If we take whether it's a late or not, it's a related word but you know, it is, it's you know, better late than never. That's what he says.

So Do you think this will reduce the total cost of cybersecurity? 'cause I feel like we have a lot of stuff that we have wrapped around things over the years, whether it's at the perimeter or the endpoint or in the cloud, but um, is there a way to rethink that whole cost structure? Definitely.

So the whole cost of the ownership is what we, what the point I try to make it. Our product name is called t c Truster Computer and Control Unit. These are the chips.

Uh, it's like similar to deep use on the data path, which is linox Nvidia and then rocom type of guys will make, we'll protect the data path on there. We need to do a controlling management. So you are there.

So our chip TCUs are there to power and turn on itself. That means I'm securing think like you build a building not after the factory, if you know the San Francisco building is leaning towards right and all that, if you're from this area, they're trying to retrofit that thing later, it's too late because you're a hundred floor building already. So versus you do from a ground up itself, how much it's gonna take, what it takes and how to pull a power up, how to boot up, how do you run the applications, everything.

They authenticate all these models. That makes it more much valuable to do that doing this way. So What's your sense of how long it might take us to replace the existing infrastructure we have with more resilient infrastructure that has a silicon kind of based approach to cybersecurity Every day The progress is happening and we are luckily, you know, participating with this uh, group called O C P, open Compute platform.

This is pretty, uh, everybody in the data center guys are, you know, part of that uh, automation, Google, Microsoft, the world, everybody driving. So they adopted the models, uh, as a modular approach right now. Before you used to go on a full motherboard that would've taken three years for us to come into, but now pluggable card, you can replace a pluggable card with a old solution.

It's our solution. So it's a simpler to change around 20% of the market is already adapted to that modular, you know, last year, this year expected to be 80% and the following year is a hundred percent. So we believe that, you know, towards the middle of next year onwards, adaptation of the silicon-based solution with us and similar to people like us will be coming into the picture.

All right. Well folks, you heard it here just like virtual machines kind of got more and more of the into the instruction sets and everything starts to move into hardware cybersecurity, it looks like there'll be no exception and we'll all be better off for it. Hey, thanks for being on the thank you.

Thank appreciate in.