Prasen Shelar
Every security pro needs an AI twin
In this interview, Prasen Shelar, co-founder and CEO of Axari, explains how his company builds AI employees for cybersecurity: digital twins that sit alongside security professionals in Slack or Teams, learn how they work, and take on the tactical coordination work that he says consumes 60 to 70 percent of their time. He describes how a twin earns trust in stages, from learning to assisting, delegating, and commanding, and walks through vulnerability remediation as an example. He also covers his path from a late-night food delivery startup in India to Demisto, Palo Alto Networks, and Fylamynt, how future security leaders will learn when busywork is automated, why Axari targets resource-constrained mid-market teams, its bring-your-own-cloud and bring-your-own-model approach, and why he believes AI will make security worse before it makes it better.

Prasen Shelar is co-founder and CEO of Axari, which builds AI employees for cybersecurity: digital twins that work alongside security professionals in Slack and Teams and take on their tactical coordination work.
Prasen started his first company, MetroMidnight, a late-night food delivery service, in India before earning his master's at Carnegie Mellon. He joined Demisto as its first product manager, building more than 500 security integrations for its SOAR platform, and stayed through the Palo Alto Networks acquisition. He then led product at Fylamynt, which NetApp acquired.
“The coordination intelligence is still in the human head. That needs to be brought out to help them with the drudgery work.”
Prasen's core argument is that security teams spend 60 to 70 percent of their time as the human coordination layer between their tools, and that defenders need AI to manage AI. AI will make security worse for a while, he says, before tools that take on that busywork make it better.
"Defenders need an unfair advantage right now, because attackers already have that," Prasen explains. "Our goal would be to not limit yourself with the headcount."
The conversation
In this conversation: Josh Rubin (Host, CTO Studio) and Prasen Shelar (Co-Founder and CEO, Axari).
I am the co-founder and CEO at Axari. It's essentially an AI employee company for cybersecurity.
AI employees. So you spent most of your career, I believe, working as a CISO in cybersecurity, building teams and products. Now you're building a product to replace some of the companies you used to sell to?
I prefer calling it AI employees, because people can relate to it pretty easily. But we're creating a digital twin for security professionals that will attach itself to you, learn about you, get your context, and help augment the work that you do. Security professionals are always bogged down with work and they need more bandwidth, and this is our way of giving them a superpower: making sure they have another copy of themselves that's focused on the drudgery work and the tactical work, which humans don't actually want to do. The learning we've had so far, after talking to so many people, is that 60 to 70 percent of the time security professionals spend right now is mostly on tactical work. So it does help to have a digital twin, an AI employee, helping you out.
Well, it's funny. You've seen The Matrix, I assume. Have you seen the original Tron? This feels like Tron. This is the digital version of myself that I send out into the environment. It is me in digital form out there. It is there to pick up work, but I imagine it is also for provenance. This agent is attached to this person, and when this agent does something, it is doing it on behalf of this person, because in the security sphere you need that level of provenance.
Yes. Although it is quite subjective in terms of the process we follow. It has to build the confidence to be used on your behalf. So the first phase is learning about you. The twin sits next to you in Slack or Teams, wherever you run your collaboration, and it understands what context you're coming from, how much work you do, and how much of it you want automated. Learning, assisting, delegating, and commanding: that's the process we follow, which builds the confidence to do that.
So you're installing something within their various systems to track them?
Not track them, but help them.
Of course. But how is it learning? How are you training this model on all of this stuff? It has to be in my Slack to see what I'm doing in Slack. It has to be in my meetings, potentially, to hear what I'm talking about and engaging with in those meetings.
Think of it as an AI employee that is part of your Slack, with the concept of a Slack app or a Teams app. It is connected to all of your tooling: your email, your calendar, your security tools, and of course all of the messages on Slack and Teams that are going on on your behalf. It reads all of that, classifies it, understands your behavior, and then it starts to help from there.
Is it sitting in on things like a Google Hangouts video call? Because text is one thing, but so much of the conversation happens verbally.
Yes. There are a lot of note takers these days, Fireflies, Granola, all of that, you may have seen. We leverage the data that's coming from the existing stack rather than creating our own. All of the vertical vendors right now, in different domains in security as well as non-security, we leverage them to give us the data to learn the behavior of the user.
You have to train your twin, which I fully understand. And this is operating from a cybersecurity standpoint. But from the outside, that sounds like a cybersecurity nightmare: I have a thing that's tracking me and following everything, and it's in my email and in my calendar. How are you protecting intellectual property, privacy, all of those things within that framework?
Look at it more from the point of what you can do to secure it, in terms of permissioned executions. The twin has an understanding of you, but how much of the data it creates is only behavioral, rather than the raw data you get from a person? We have different stacks we've built which create observability around what exactly the twin is doing, what permissions the twin has to execute on behalf of the user, how much the user is in the loop, and how many of the actions are executed in a sandbox. So there are a lot of controls we put in place around the twin that make it more trustworthy, so the human is more confident doing that with a digital identity.
Okay, we're going to back up for a second, because I think we started with the second point as opposed to the first point. You operate Axari. You've been working in cybersecurity for years and years. Give me a little bit about your background and how you got into this.
I sort of call myself a dreamer, because I started my first company back in India, called MetroMidnight, where we delivered food from 9 p.m. to 4 a.m. That was a different idea than security, but it helped me grow as an entrepreneur. I came to the States in 2017 and did my master's at CMU. The first company I worked for was Demisto, which did incident response automation. We built the category SOAR, security orchestration, automation, and response, when analysts were remediating incidents on Excel sheets. It was very manual. That was my entry point as a product manager into security, and as part of it I built more than 500 security integrations for the product. That helped me understand the different verticals of security and the complexity of how a human deals with different security tools, and that led me to who I am now, in terms of the workflow automation engines we've built. Demisto got acquired by Palo Alto Networks. I worked there for about two and a half years and led the team, and worked with Prisma Cloud as well to build a remediation stack, so I learned about the different stacks of security at Palo Alto Networks overall. Then I joined another company, Fylamynt, that got acquired by NetApp, and now I'm here building Axari. All of that background has led me to lean towards a product mindset for security.
It sounds like you were in the perfect position, starting from the ground, to see ground zero: every step someone has to go through to eliminate points of failure and shore up security in a system. That allowed you to step in and say, okay, I can automate this bit, I can automate this bit over here. And that compounds over time. So now, at Axari, what problem are you specifically trying to solve with the tech you're rolling out?
I think this is an excellent question. Even though you have all of these vertical, domain-specific tools, there is a layer of work that only humans do right now, and they are accountable for it. If I need to secure myself against ransomware, if I need to patch a vulnerability, if I need to be completely compliance assured, this level of work is a runtime human layer right now. Humans take the overall understanding from the tools, stitch all of the information together, and carry it towards the final outcome: finally the vulnerability is patched, finally I am secure in terms of compliance, or all the audit readiness has been done. For a lot of these human actions they're accountable for, there has been no tool that helps. And humans are not able to keep up with the pace at which AI is coming, where the vertical tools and the attackers are going crazy with it. So we help humans at that worker layer. We bring an AI teammate that becomes a first-class citizen of their team, increases their productivity, and helps them achieve the outcomes they are accountable for.
I want a specific example. Give me some specific thing where these new AI tools can come in and say, a human was only able to do this before, and now this AI tool is taking it on and is accountable, free, and qualified to do this work.
Yes. Say a vulnerability is seen by Snyk or that kind of product, and a human is notified who is now responsible for patching it. The work that comes after this is to understand the vulnerability, prioritize it, look at the assets, then actually patch it with the developer, put it into the overall software development lifecycle, and get it remediated. There are a number of steps after the vulnerability has been flagged to get it remediated. That is completely manual, and there is no layer that helps a human track it. If this takes hours or days, that is where the digital twin can come in: stitch all of the context together, reach out to the right people, understand what the next action should be, help coordinate it, and autonomously do the drudgery work the human was accountable for, keeping humans in the loop and getting their approval wherever required. Getting that drudgery work off people's plates and helping them fast-track it is immensely powerful.
Absolutely. But now the hard question. That drudgery work is how you learned to do what you're building right now. This is the junior question. When you are automating all of the drudgery, when you're taking that base-level, Excel-spreadsheet-style work out of the equation, how does the next generation come up? Where does the next group of CISOs come from?
Okay. I think there are two levels of information a person has when they work in security. One is the worker knowledge: how much do I know about the vulnerability and how to patch it? The other is coordination intelligence for an organization: how does this work in my organization? What humans do I have, what systems do I have, where is the information locked up? That coordination intelligence is only in human heads right now, and it's not being commoditized by the AI labs. The worker knowledge, to patch the vulnerability or look at compliance, is being commoditized by the AI labs. But the knowledge that lives in people's heads around the processes, and it's a network-of-n-nodes problem, that coordination intelligence is still in the human head. That needs to be brought out somehow to help them with the drudgery work and the automation. Understanding the worker knowledge is something one has to do to evolve as a security professional. But even if the coordination intelligence is automated, it doesn't really bother the security professional, because it's just understanding the context, stitching things together, and getting over the coordination challenges people are facing right now, because security is extremely fragmented. So much happens in so many tools. Even if the context switching and the orchestration are automated, there's not much to learn from them. But the in-depth understanding of these concepts and the way they operate, humans need to do that regardless, because if they're missing the context there, the gap in understanding will percolate into their organizations.
And arguably the technology has changed. It's akin to: I am a carpenter that builds houses. I'm now a senior. I came up using a hammer and nails, but now all I use is a nail gun, and the juniors coming up are all using nail guns, and I'm mad that they can't use a hammer and nails. Well, they're never going to. That's actually not the point. Their goal was to build a house. The tool has changed, so I can't get angry that they don't have to deal with the same lower-level tools I had to.
A hundred percent. You don't necessarily need to process the information the way we did a few years back. The level of work and information has increased to the point that it has come to the application layer. At the application layer, I as a security professional should understand how to achieve an outcome. What I care about is that my company should not be breached, or I should have zero percent business disruption. How do I take care of that? That is what I need to care about, not the underlying orchestration. You don't need to understand all these concepts in depth to achieve the outcomes right now.
There used to be a bifurcation, obviously: I'm a back-end engineer, I'm a front-end engineer, I'm a security-focused engineer. A lot of that is collapsing with these tools. But I imagine in your mind there is still something that differentiates a person who has more of the security mindset from someone who has more of the application-building mindset. What makes for a good security developer?
This is actually a really complex question to answer, I would say, but I'll give you my take. The fundamental difference between a security engineer and an application engineer, or a general developer, is how you understand the controls around the application you're building. What are the different layers at which there could be loopholes, and what controls should I put in place to secure my stack? That level of understanding comes from understanding the controls around the particular application layer you're building. That fundamental understanding has to be there. Now, there are different levels in security. For instance, there are analysts who operate at the grassroots level: triaging and looking at the alerts, managing and enriching those alerts, forming incidents from them, and then remediating those. Then there are managers looking at some of these processes and orchestrating that work. Then there are deputies or CISOs creating the security program around it. All of this understanding is essentially about reducing risk: what you've been building and how it can be exploited. That understanding of risk, and of the complexity around the controls, is what you need if you're going to be a security engineer versus a general software engineer.
How big is your company right now?
We have around 20 people so far. We raised our pre-seed and now we're raising our seed.
Okay. Eventually you'll go down the series. What are you hiring for, or what are you raising for, right now? What unlocks growth for you?
We've identified problem-market fit. We definitely understand that people want to challenge the way they've been working with their existing security setup, and they want AI teammates. Now, to scale, we're looking at the different levels at which we can operate in enterprises or the mid-market. That repeatability, and the assumptions behind our existing AI teammate use cases, is what we want to validate with our seed round. So we'll be hiring a lot of go-to-market people and forward deployed engineers to support that. We have a really good stack in terms of product. We're just trying to scale from here with our seed.
So essentially where you're looking to invest is in the people who are going to go out and integrate. You have the pipeline and deal flow from customers saying, I'm ready, come integrate with me.
Yeah. We've built a network of CISOs that helps us get the leads and the overall pipeline, and it's been really powerful in that sense. Now we need people to actually reach out, validate, deploy, and grow that overall go-to-market pipeline. So we need sales folks, we need some forward deployed engineers, and we need people who will raise awareness of the product we've created.
What is the story you're taking out into the market with this? It feels like there are two stories you can take out with security. Either the fear story, it's coming for you and you need to protect yourself, or the value story, we're going to give your people their time back and you're going to save or make more because you're using my product. What do you see as your biggest value prop there?
RIP busy work. I think that's what we want to do. Defenders need an unfair advantage right now, because attackers already have that. If they have a product that makes them superhuman and increases their efficiency and productivity, that will help the organization achieve its outcomes better. The fear is definitely there, and it's being induced by a lot of vertical players. But our goal would be to not limit yourself with the headcount, to extend your headcount to be better and achieve those outcomes at faster scale and efficiency. That's the story we're taking out.
Who's the perfect customer for you?
We're targeting folks who are resource constrained right now in the mid-market. That's a really solid target which a lot of people don't focus on. There's also a lot of complexity being added because of the tooling and the fragmentation affecting lean enterprises right now, and we want to help these people remove that complexity from their stack by adding AI employees.
It's always harder to sell to resource-constrained people than to the people sitting on a pile of cash. Why are you targeting that sector specifically?
Everyone you see in security, in the mid-market or anywhere, has turned out to be a developer. Everyone is building something internally. They're creating their own chief-of-staff and personal assistant agents and stitching things together. So they've built an invisible budget line anyhow to do the work we'll give them, which they can use to create these seamless workflows. That makes it easier for us to target someone who's already building it and who can actually buy, because they don't have the time to do it themselves. A bigger company has the resources and time to build some of these things, so there it becomes a build-versus-buy argument. But for a mid-market company, time to context and velocity are a really big challenge.
What is your moat, to prevent an Anthropic or a Codex from just coming in and saying, we're in all your stuff anyway, we can do this for you?
There are two things that give us an advantage right now. One is the domain expertise we have. We've had around 300 CISO conversations in the last six to eight months, and the corpus of information we've collected around what specific outcomes they want solved in the coordination intelligence problem, which we've turned into our own workflows. That's the fundamental moat we have right now, or what we'd call an advantage. I'll give you an analogy. When you're hiring and there are three humans in front of you, you choose the one who's the better fit for that particular role. So even if you see a generic player, you'd probably go for a domain-specific player to hire at that point. And when you hire this human, and they come into your organization, learn, understand, become more efficient, and now do the work at half the cost or with more efficiency than other people, it's really difficult for you to rip and replace them at that point. So the accumulation of organizational intelligence we get when we come in with our AI employee is extremely powerful, and that goes along with the advantage of our domain expertise and the platform we've created to support this orchestration.
Where does the agent fundamentally live, in the context or MCP or wherever you're storing that? Are you leveraging the frontier models to actually do the processing of everything, or are you air-gapping stuff and saying, here's a computer, you can on-prem this and you don't have to worry about everybody getting all of your stuff?
Yeah, and this is one of the major things we have: bring your own cloud. You can deploy this on-prem. And bring your own model: you can use your own model that's been allowlisted in your organization. And we optimize it even further. Say there's a task you do with a frontier model and it costs you $1. With security-specific optimization, using multiple models in that process, we can do it at a much lower cost and with more efficiency. So model optimization for the task, a multi-model approach, and deferred execution help us be better than what a frontier model will do for you right now. We're also model agnostic, framework agnostic, and tool agnostic. We're a platform that does all of this orchestration and management for you, with an interface that's an AI employee. That's better for you than working with an incumbent, or with someone like Claude or OpenAI, right now.
Are you focused on businesses in regulated or unregulated industries? Because once you introduce something like HIPAA or FINRA, it obviously gets more complicated.
We're working towards the regulations ourselves right now. We will go towards regulated industries soon, because that's where you really need an orchestration layer to turn down the complexity. But right now we're working on our own overall maturity in terms of how we achieve those regulations.
Last question, which is more of a culture and thought question. Is your product making the world better or worse, ultimately?
We're definitely making it better, because this is what I personally believe: AI is going to make things worse for us in security for some time. I'll give you an example. Right now you have different tools and humans managing things. Every layer is being AI-ified right now, and it's adding more and more complexity. It's making things faster, but it's adding more complexity to manage at the human layer, and that becomes a fragmentation challenge. So it's making it worse right now. But when you have a tool like ours that helps with this, it starts to make it easier, faster, and better. You need AI to manage AI, with a human in the loop, and that will bring down the complexity and help with speed and efficiency. That's how we come into the market and help with AI employees.
A vision popped into my head as you were saying that. With AI writ large, not just security, AI right now is incredibly disruptive. Disruption long term may prove to be great, but in the short term nobody likes to be disrupted. It is akin to cultural puberty. We're all going through it, and it feels weird. But maturity is what you get when you come out the other side. And what you're talking about is trying to build the harness layer, the orchestration layer: how do you apply understanding to these new muscles, this new capacity for thinking we've developed? So yes, disruption is hard, but the outcome, if controlled properly, is that you come out the other side an adult.
Exactly. And that's our vision: to help humans. We're not a typical security tool that just focuses on solving problems at the vertical layers. We're a product that exists because we care about humans, and we're building for humans to be better. That's the overall motto, I would call it, and that's why a lot of people are resonating with us: we're making the life of a human easier with our product. That's why we have such a solid network of CISOs. We have 30-plus CISOs as investors, 40-plus advisors on board, and 70-plus CISOs using our product. The level of network we've built is because we care, and people have really resonated with the concept that the human layer needs more efficiency, and with this AI employee that sits next to them and makes them superhuman.
Prasen of Axari, for CISOs, by CISOs. Thank you so much.
Thank you. This was exciting.
Be part of the
conversation.
Whether you’re a CEO or founder who wants to be featured or a company interested in sponsoring, there’s a place for you here.