Black Hat USA 2026 | Sumedh Thakar, Qualys

Clips
More from Black Hat USA 2026

Sumedh Thakar

President & CEO

Qualys

play_circle_outline Prioritizing Exploitable, High-Impact Vulnerabilities to Break Attack Chains and Reinforce Defense-in-Depth

play_circle_outline Agent-Based Scanless Detection and 24-Hour Zero-Day Remediation Against AI Autonomous Exploitation

play_circle_outline ML/AI-based patch reliability scoring to reduce outage risk

play_circle_outline Validation of exploits using safe payloads to confirm real exploitability

play_circle_outline Data governance: controlling training data and preventing sensitive data leakage

play_circle_outline Risk Operations Center: Translating Cloud, Identity, and Misconfiguration Risks into Dollar Metrics for Boards and Cyber Insurance

Info

Sumedh Thakar, Qualys

Sumedh Thakar

President & CEO Qualys

search

(INTRO)

Krista Case

>> Hey, welcome back to theCUBE. We are approaching the end of day two here at Black Hat 2026, but the conversations are still going strong around the fact that vulnerabilities can be discovered, weaponized, and exploited much faster than organizations can assess and remediate them. So we do see that security teams are struggling with the volume and speed of change. That's really outpacing these traditional approaches. And at the crux of that problem really is their ability to prioritize what to fix and then take the appropriate action. But there certainly is some hope here in terms of evolving this approach to making better cybersecurity risk decisions. Sitting down with me today is Sumedh Thakar, President and CEO of Qualys. Sumedh, thanks so much for joining theCUBE.

Sumedh Thakar

>> Thank you very much for having me.

Krista Case

>> Absolutely. So before the cameras were rolling, we were talking about some customer conversations. I know you said you've been getting some really great feedback.

Sumedh Thakar

>> Yeah.

Krista Case

>> And it's consistent with what we're hearing from a lot of CISOs, which is that they're trying to build confidence in terms of understanding that they're prioritizing the risks that matter the most to their business. Are you hearing that too? And if so, what do you think is the missing link there?

Sumedh Thakar

>> Yeah, that's a great point. I think the question is, why are we talking about prioritization? If we could fix everything, we would have just done it, right? And so I think in many ways this is not a new issue. For the last few years everybody's been struggling with not having enough resources to fix everything. The problem just has exploded even more with frontier models, right? So it's not like we had enough human resources anyway to fix everything. And so now the questions are coming even more, which is I cannot fix everything, so by nature I need to prioritize. And so then the question is how do you prioritize and what you ask is how do I build the confidence in the prioritization, right? And I think there's sort of like a couple layers of prioritization here. It's just the basic prioritization based on the threat intelligence. If things are actually being exploited, there are a lot of theoretical vulnerabilities out there that have no way to be exploited. The second prioritization comes from, is this actually exploitable in my environment? So while a vulnerability could be very exploitable out there, CISOs have put defense in depth. They have put other controls like a firewall, like an EDR, which actually can block these exploits. So the second prioritization comes from, can I actually run these exploits to confirm that they are exploitable so now you can reduce the findings even more? And then the last part is, yes, even if it is exploitable, is it something that leads to a big loss to my business or no? And so I think what we see is that you might have millions of vulnerabilities but only 1 % of those actually are exploitable and then only a small percentage of those actually lead to a business loss.

Krista Case

>> Right, it's a great point, Sumedh. We can't waste time chasing, like you say, these vulnerabilities that aren't going to impact our business at the end of the day. So I understand that Qualys has had an announcement on the vulnerability management and scanning front. Now, vulnerability management traditionally has relied on scanning.

Sumedh Thakar

>> Yes.

Krista Case

>> But can you talk through why that approach isn't working anymore? And maybe some of those gaps that Qualys saw that needed to be addressed.

Sumedh Thakar

>> Well, thank you for that question. I think it's very interesting because I have been at Qualys for 23 years. I was one of the original engineers that worked on the platform. And at that time, vulnerability management was two pieces, scanning and then fixing. And as the volume of vulnerabilities exploded the last many years, you now have three pillars of vulnerability management. Can I detect quickly? Can I prioritize quickly? and then can I fix quickly? And in response to the new threat model, every CISO is having to explain to the board what is your approach to AI -based autonomous exploitation, and their response cannot be, we are going to hire more people. So you have to have a response that basically grounds yourself in saying, we will have some form of autonomous response. And I think that's where the fixing is the part that gives you the most bang for the buck, then you have the prioritization, which helps you reduce what you need to fix, but your scanning has become even more important because if you're not able to find the issues in the first place, what are you trying to fix? And so those three pillars of quick detection, which we call agentless scanning and Agent Insta, second is validation of exploits with Agent Val, and then the third is the ability to actually remediate vulnerabilities, not just patch, but remediate with different approaches, which is Agent Sara. So all these three agents together are giving us hope that today we have the ability to go from detection to remediation in what I call a zero -day remediation. In the first 24 hours, we can get things fixed. It is possible today, technology exists, and we're excited to bring that to our customers.

Krista Case

>> Yeah, so zero -day remediation. Can you walk through, I guess, maybe some of the challenges that a customer might encounter as they're trying to make this a reality? it's definitely a lofty goal.

Sumedh Thakar

>> Yes.The good news is it's doable. We have 150 million patches we've deployed in the last 12 months. Out of those, 40 million today are already autonomously deployed without any human intervention. So what is the pushback for remediation? The pushback always is, well, what if the system goes down? So that's the pushback on patch management and remediation. And so what we are focused on is to say, well, how can we build confidence that you can patch without worrying that it's going to be an outage? So the first part of that is how do we help you make sure that if you want to reduce the possibility for an outage, fix the least amount. So that's where this hyper -prioritization is important so you can get down to only 1 % that you fixed. Second becomes, well, is there an alternative without a patch that I can apply a configuration fix that will prevent the exploit and I have more time to patch later? So what we have done is leveraging these frontier models ourselves. we have created the ability to create mitigations where you can actually maybe just change the permission of a file, maybe you disable a service, but with that, the potential of a compromise becomes a lot less and you don't have to patch right away. And the third and the last piece there is to build confidence in the IT team's ability to deploy the patch without worrying about an outage. We took the half a billion patches that we have deployed in the last few years where we see every rollback, we see every error that we get, and we build an ML AI based patch reliability score that allows us to do that as soon as a vendor releases a new patch, we can actually tell you this is a high reliability patch and it does not need a reboot. So you feel higher confidence that I can remediate in zero days because it's a high reliability patch and I do not need a reboot. And so with that, we can, look, at the end we are managing risk. So it's not that we have to fix everything, but if we can take enough off the table, the potential of a compromise goes down significantly. So auto -remediation can fix 20, 30 % of your stuff, but that 20, 30 % can really reduce the potential of an entire chain of attack happening successfully.

Krista Case

>> Absolutely, because we're seeing that, right? These AI -driven adversarial attacks, they can chain together these vulnerabilities. So it's a great point. And I'm glad that you called out sort of how you're validating the confidence, right? Because that is something that we're seeing security teams, especially if they're trying to integrate AI as part of their processes, which is necessary to do in order to keep pace, they're needing to understand that they can trust the AI. So are you building confidence with customers?

Sumedh Thakar

>> Exactly, right, I think that's where the combination of the data, the analysis, the inference, and the AI models, because we built the harness, so we are part of Glasswing, we are in Daybreak, so we are basically collecting and leveraging these models of AI to build mitigations and to build safe exploits for customers. Because as you saw with OpenAI/Hugging Face, that whole thing, they don't care, the autonomous AI does not care about this, the system shouldn't go down. It'll try 10 different things, one of those could bring the system down. So how do we leverage this AI capability to build out exploits that are safe with higher confidence for the customers so that they can test themselves without having to worry about an outage. The same thing for the mitigation. High confidence mitigations that are tested, that are created by reverse engineering exploits through AI, and then giving those to the customers so they have much higher confidence remediation that they can do, and all of it is based on not just using AI as is out of the box, but building confidence around the AI capabilities so that you can use those for remediation.

Krista Case

>> That makes a lot of sense. And Sumedh, can you talk, maybe just double click a little bit more in terms of how you're validating these vulnerabilities. I know you mentioned that a moment or two ago.

Sumedh Thakar

>> Yeah, I think a very simple way to look at that is if there's a bunch of different doors that are there and the traditional approach of prioritization will tell you maybe these 20 doors could lead to a compromise, right? I think what we're now focusing on is saying, hey, we're just going to try to open those 20 doors and we'll tell you by the way, only five of those open. And that we are doing so that now you have a much more focused approach and that's where that exploit validation comes in where it's not a complete red teaming exercise, but what we can do is when you see a vulnerability, we send a safe payload. A safe payload could be as simple as saying, hey, can you resolve this DNS name? If we see the DNS resolution come through, we don't need to update your system, or write anything. We can tell we could compromise it. So now that gives you a higher confidence in saying, hey, by the way, my controls didn't work. I got actually an exploitation that happened, so that becomes my priority. Versus if we send a payload, EDR blocks it, firewall blocks it, and it doesn't execute, then you have more time. So that, again, reduces the potential of an outage because you're not fixing everything, you're fixing the one that works. And so that's where we do a lot of the validation, by running very, very safe ways to actually test if the exploit works.

Krista Case

>> That makes sense, thank you, Sumedh.

Krista Case

>> Of course.

Krista Case

>> And I know another area you had an announcement. This week, my understanding is around AI governance. We've been hearing a big focus on run time and how do we put these, you kind of described, a harness, how do we put the right guardrails in place, because enterprises are just racing to try to adopt AI. And so for our security teams, they need to not get in the way of that, they need to enable that. So can you talk to, I guess, maybe what you're seeing customers struggle with from that perspective?

Sumedh Thakar

>> I think we are at the early innings of AI, and so I think the initial challenge is shadow AI, right? And how do you figure out what's in my environment? And I think it's almost like not having a shadow AI problem is a business problem, which means that you're not actually using AI in the business, which is not good for business. So I think today it is very important that every business actually is using AI. And then we kind of were at the beginning where everybody wanted to stop shadow AI because they're like, I want to know who's using AI so I can block them. Then we saw the pivot to saying, well I want to know who's not using AI so I can send them somewhere else.

Krista Case

>> Yes, right.

Sumedh Thakar

>> So now we're kind of rolling back to saying, look, I need to be able to have the visibility of who's using AI, what are they doing with it, can I put some guardrails around it, where are my models, where are my GPUs, and I think I've done this long enough in cyber that no matter what new technology comes, it always has these components, which is discovery, whether it was cloud, whether it is going to be quantum in a few years, the question will always be the same, where do I have it? Can I assess it, I cannot fix everything, can I prioritize it, and then at the end, if I don't get it fixed, what was the point of this dashboard tourism?

Krista Case

>> Right, right.

Sumedh Thakar

>> We were just looking at dashboards, we're not doing anything. So that same thing is happening with AI. So what we did with TotalAI is to leverage the Qualys footprint customers have and create the ability for them to figure out their shadow AI, what's happening in the client machines, where you're running AI in the cloud environment, et cetera. So you can first get the visibility. Once you get the visibility, we can test the models, we can see if the model is doing something that it should not, and then we can help you build out policies to be able to say we have some governance capability where unauthorized AI engines are not being used, the AI engines are not being fed data that they should not be. So it's something that is going to come up more and more in the next few months as a priority for customers as soon as they get through the current challenges that they have. AI security and governance is going to be definitely something that is going to be important for a lot of customers.

Krista Case

>> I agree, Sumedh. And we're certainly seeing that as well, that customers are trying to turn the lights on, right, and kind of solve the shadow AI piece of it. When you think about the stack that's needed to govern AI, what do you think are going to be maybe some of the key control points?

Sumedh Thakar

>> Yeah, I think there are a couple of control points here that are very important. First, discovery and inventory, having an upgraded inventory. Then you have the hardware components of AI. I think we also have to think about that some of these are not necessarily a security issue. Some of these are just a business issue. How many tokens am I spending is more important sometimes than the security part of it. So the inventory is something that is going to be important. After that, do I know all my models? Where are they running? It becomes important. Testing the models to ensure that they're not giving you answers that you should not be giving becomes important. But then the last piece that is also very important is what data is being used to train these models becomes very important. So if I know a model is being trained, is it being trained on my employee information? That should not be. If it's a support chatbot and I ask it, what's the employee salary? is it going out and talking to an MCP server to give that information, that becomes a control point. And then at the end, we are going to have to produce reports. For regulators, we're going to have to produce reports for the board to show we know where our AI is, to show we have tested it, to show that we have actually put guardrails around it, and that we are continuing to monitor it, right? But at the end, AI is one risk. You still have cloud risk, you still have identity risk, you still have misconfiguration risk. So all of that needs to come together in a singular risk view, where what we really see today is the building of a ROC. Like you have the SOC, which is after breach. Proactive risk management is Risk Operations Center. So how does AI security plug into that? And then in three years, it's going to be quantum security. But at the end, companies have a limited budget, and where should they spend that money is going to be based on where is the risk. If AI is the risk, spend more on that. If quantum is the risk, spend more on that. And the ROC actually helps you get that visibility.

Krista Case

>> Walk us through the ROC, because I know I've had some conversations with your team about the Risk Operations Center and how it's complementary to the Security Operations Center and doesn't necessarily displace it. But Sumedh, I'd love to hear from the horse's mouth today.

Sumedh Thakar

>> Yeah, we are having really great conversations because the SOC, I know I described that as snipers, right? They are focusing on somebody in my environment now and how can I get rid of them? That's their main focus. Because they are not thinking three months, six months from now, et cetera. And so what has happened is that we have done a good job in creating SOCs with SIEM and XDRs and looking at log data, et cetera. But when it comes to proactive risk management, the question is, I have cloud risk in one dashboard, I have identity risk in a different dashboard, I have on-prem risk in a different dashboard, and so we end up with dashboard tourism. I just have dashboards. But if I ask you, you have a business unit that makes 500 million a year, what's the potential dollar value loss on that business unit? You cannot say because every risk with a different score is in different areas, right? So the idea of the Risk Operations Center is how do we elevate the conversation of risk? How do we take vulnerability as only one component of risk? Second component is misconfiguration. Third component is identity. Bring all of that, harmonize all of that, and then get into threat intelligence, business context, remediation, and compliance, all to be very focused on risk itself, and then acknowledge that risk management exercise has three components. One is mitigation proactively. Second is consciously accepting a certain risk. And the third is what you cannot accept and what you cannot mitigate, you need to transfer to a cyber insurance company. So the ROC will allow you to have a conversation with your cyber insurance company to say, here's my sustained risk level, here's the additional residual risk that is left that I cannot accept, and what is the cost of transferring that risk to a cyber insurance company. So the Risk Operations Center is more of a strategic plan for CISOs over the next couple of years, is implementing a ROC so that they can answer the question that the board wants, they don't want to know the number of counts et cetera, the board just wants to see, is 90 % of my revenue under my risk appetite?

Krista Case

>> And that's, it really kind of cuts to the chase and it puts it in those board level terms.

Sumedh Thakar

>> Exactly, yes.

Krista Case

>> So you kind of referenced that you're having some great customer conversations around the ROC. What are some of the points of visibility, if you will, that you think security practitioners don't have today that the ROC is giving them? Is it kind of connecting those points that you mentioned?

Sumedh Thakar

>> Yeah.Well, exactly, right? So if you have an application that's making you 500 million a year, there's a separate tool that looks at the cloud piece, there's a separate tool that looks at identity piece, and a separate tool that looks at configuration piece. And so if you are asked the question, you have individual visibility, but if you say how much loss could I have if there is a ransomware attack on this 500 million business unit, and where is that risk coming from, you cannot explain that because it's distributed across multiple tools. The good thing about the ROC is that it takes data from existing tools, pulls it into a single platform, normalizes the different risk scores, because some are on 10, some are on 100, normalizes all of that, and then puts them in the context of your overall risk appetite. And so the visibility is there from an individual perspective, but then the CISO's job is to really bridge that gap between technical findings to what it means to the business. And the ROC really enables giving them that visibility of having those technical findings translated into business language and explaining to the board how much dollar value loss it could have and then where are we in our journey to ensure we reduce the risk of that loss. There is nothing like zero risk. So, I always say to people, by not doing anything, you're already accepting a risk.

Krista Case

>> Yes, absolutely, like we were talking about, businesses are recognizing that they have to accept a certain level of risk, and I think it's really powerful to be able to translate this into dollars and cents, because it's really difficult to do when we think about something like a ransomware attack in your example.

Sumedh Thakar

>> Yeah, yes and no, at the end of the day, the reason you have to put cyber in dollars and cents is because cyber is being used to protect dollars and cents. So that's like when you get car insurance, you have to say, I'm going to pay $1 ,000 to insure a car that is worth $50 ,000, right? If you don't know what the car is worth and you don't know what you're trying to protect, then you are taking infinite risk. And so I think it's always been warranted. It is becoming more and more important now that you cannot fix everything. You have to be able to explain the prioritization in terms of saying, why am I focusing on this or why am I not focusing on this? Because the business loss is not there. And I think we have now built up over the last 20 years enough data points through insurance and ransomware payouts that we have data available as we partner with a company that gives us that kind of information where we can say, if you are a business that is in this segment, in this geographical location, on average, how much ransomware payout has happened for industries in your business? And so you can project from that.

Krista Case

>> I see.

Sumedh Thakar

>> If I have a $500 million business unit, typical ransomware payout has been around 10 million. Okay. And so now you can put a number to it. So do you want to spend 10 million? No, but maybe you could say, look, if it's a 10 million payout, probably I want to spend a million to make sure that I don't have the 10 million payout. So I think that more and more of the data points are available and we're bringing it together, but I think I'm excited to continue to work in that direction.

Krista Case

>> Yes, yes, certainly. So Sumedh, you've kind of described yourself as a frontier AI optimist, right? And I think there's a lot of potential doom and gloom here, but reading between the lines, does that indicate you think as an industry over the next, call it, six to 18 months, that we're going to have some practical learnings and really kind of make some progress in terms of our ability to prioritize and respond, everything we've been talking about today?

Sumedh Thakar

>> I am very much optimistic when it comes to that. I think unfortunately some of the frontier AI rollouts have created sort of haves and have nots, and so there are geographies, countries, and industries where they don't have access to some of these models. The good news is that platforms like Qualys and some of the other cyber platforms are able to leverage these models to create more and democratize the outcomes of these frontier AI models so that customers can leverage this within their environment without themselves having to invest in using, leveraging these models. And so I think I'm optimistic because, yes, attackers are using AI for creating attacks, but defenders also have access to similar technology, just looking at it from a different perspective. And so one of the optimistic views that I have is, in theory, if every organization that writes code has the ability to run this kind of frontier AI model on their code base, they will find all their vulnerabilities themselves, and you will never have zero -day vulnerabilities. because a zero day happens when an attacker knows something you don't, but if you're running AI models you will always know. The flip side is that then the risk moves towards the zero day remediation because as soon as you release a patch, they will reverse engineer using AI, but the good news is that we at Qualys can also do the same. Once a patch is released, we can reverse engineer using AI, find the exploit, and then use that in a positive way to create compensating controls and to create a testing methodology that is safe. So I do feel like AI is definitely bringing positive outcomes for cybersecurity for a lot of organizations, and they don't necessarily themselves have to be investing in running these AI models. They should hold their partners and vendors responsible for ensuring that the appropriate amount of AI is being used to give them an outcome, which is at the end of the day what they care about.

Krista Case

>> And that's a really fair point, Sumedh. Maybe a concluding thought. So if I'm a practitioner and I'm evaluating vendors for trying to help me navigate my risk management. There's a lot of AI washing, especially if I'm in the CISO role, I'm being just inundated. What would you look for in a vendor to sift through that noise and really make sure that this vendor is going to actually deliver something that's actionable?

Sumedh Thakar

>> That's a great point. I think you have to, before you even go look for a vendor, you have to ask yourself, what is the outcome I'm looking for? And can I have a solution that will find malware is not an outcome, right? If I'm looking at business risk, I'm looking at, for me to avoid a $500 million loss, I'm going to create four layers of controls as my defense in depth. And these are the four ones. And then you have to be able to look at the vendors and say, look at the end of the day, AI or not AI? When you're looking at a vendor, there are three things everybody's looking at. Number one, is it accurate? Because if the solution is not accurate, it doesn't matter how cheap it is. Number two, is the solution fast enough? If it's accurate but takes six months, it's of no use. And then the last point is, is it cost effective? So is it actually helping me do the first two but without having to spend more money than I earn? And so I think when you're evaluating a vendor, those outcomes are what you should be looking at. And the vendor who's leveraging AI the right way, and, for example, there are vendors in our space that are charging the customer by the token. Now as a customer, you don't want to deal with that. You're looking for an outcome. So what we do as an example is we create a harness in the back end that picks the appropriate model. So we're not charging customers by the token. So they are getting an outcome that is a predictable outcome from a cost perspective and they're getting accurate, fast, and cost effective ways that they can actually take to resolution and that ultimately I think is what everybody should be looking at is what is my final outcome in risk management and does this vendor help me do things that are accurate, fast, and cost -effective?

Krista Case

>> Well, Sumedh, thank you so much for sitting down to walk us through all this today on theCUBE. We really appreciate it.

Sumedh Thakar

>> Thank you for having me. It was a real pleasure. Thank you very much.

Krista Case

>> Absolutely. And thanks so much for joining us. We'll be back in just a couple minutes with some concluding thoughts from day two here at Black Hat 2026. Thank you.

Sumedh Thakar, Qualys

search

(INTRO)

Krista Case

>> Hey, welcome back to theCUBE. We are approaching the end of day two here at Black Hat 2026, but the conversations are still going strong around the fact that vulnerabilities can be discovered, weaponized, and exploited much faster than organizations can assess and remediate them. So we do see that security teams are struggling with the volume and speed of change. That's really outpacing these traditional approaches. And at the crux of that problem really is their ability to prioritize what to fix and then take the appropriate action. But there certainly is some hope here in terms of evolving this approach to making better cybersecurity risk decisions. Sitting down with me today is Sumedh Thakar, President and CEO of Qualys. Sumedh, thanks so much for joining theCUBE.

Sumedh Thakar

>> Thank you very much for having me.

Krista Case

>> Absolutely. So before the cameras were rolling, we were talking about some customer conversations. I know you said you've been getting some really great feedback.

Sumedh Thakar

>> Yeah.

Krista Case

>> And it's consistent with what we're hearing from a lot of CISOs, which is that they're trying to build confidence in terms of understanding that they're prioritizing the risks that matter the most to their business. Are you hearing that too? And if so, what do you think is the missing link there?

Sumedh Thakar

>> Yeah, that's a great point. I think the question is, why are we talking about prioritization? If we could fix everything, we would have just done it, right? And so I think in many ways this is not a new issue. For the last few years everybody's been struggling with not having enough resources to fix everything. The problem just has exploded even more with frontier models, right? So it's not like we had enough human resources anyway to fix everything. And so now the questions are coming even more, which is I cannot fix everything, so by nature I need to prioritize. And so then the question is how do you prioritize and what you ask is how do I build the confidence in the prioritization, right? And I think there's sort of like a couple layers of prioritization here. It's just the basic prioritization based on the threat intelligence. If things are actually being exploited, there are a lot of theoretical vulnerabilities out there that have no way to be exploited. The second prioritization comes from, is this actually exploitable in my environment? So while a vulnerability could be very exploitable out there, CISOs have put defense in depth. They have put other controls like a firewall, like an EDR, which actually can block these exploits. So the second prioritization comes from, can I actually run these exploits to confirm that they are exploitable so now you can reduce the findings even more? And then the last part is, yes, even if it is exploitable, is it something that leads to a big loss to my business or no? And so I think what we see is that you might have millions of vulnerabilities but only 1 % of those actually are exploitable and then only a small percentage of those actually lead to a business loss.

Krista Case

>> Right, it's a great point, Sumedh. We can't waste time chasing, like you say, these vulnerabilities that aren't going to impact our business at the end of the day. So I understand that Qualys has had an announcement on the vulnerability management and scanning front. Now, vulnerability management traditionally has relied on scanning.

Sumedh Thakar

>> Yes.

Krista Case

>> But can you talk through why that approach isn't working anymore? And maybe some of those gaps that Qualys saw that needed to be addressed.

Sumedh Thakar

>> Well, thank you for that question. I think it's very interesting because I have been at Qualys for 23 years. I was one of the original engineers that worked on the platform. And at that time, vulnerability management was two pieces, scanning and then fixing. And as the volume of vulnerabilities exploded the last many years, you now have three pillars of vulnerability management. Can I detect quickly? Can I prioritize quickly? and then can I fix quickly? And in response to the new threat model, every CISO is having to explain to the board what is your approach to AI -based autonomous exploitation, and their response cannot be, we are going to hire more people. So you have to have a response that basically grounds yourself in saying, we will have some form of autonomous response. And I think that's where the fixing is the part that gives you the most bang for the buck, then you have the prioritization, which helps you reduce what you need to fix, but your scanning has become even more important because if you're not able to find the issues in the first place, what are you trying to fix? And so those three pillars of quick detection, which we call agentless scanning and Agent Insta, second is validation of exploits with Agent Val, and then the third is the ability to actually remediate vulnerabilities, not just patch, but remediate with different approaches, which is Agent Sara. So all these three agents together are giving us hope that today we have the ability to go from detection to remediation in what I call a zero -day remediation. In the first 24 hours, we can get things fixed. It is possible today, technology exists, and we're excited to bring that to our customers.

Krista Case

>> Yeah, so zero -day remediation. Can you walk through, I guess, maybe some of the challenges that a customer might encounter as they're trying to make this a reality? it's definitely a lofty goal.

Sumedh Thakar

>> Yes.The good news is it's doable. We have 150 million patches we've deployed in the last 12 months. Out of those, 40 million today are already autonomously deployed without any human intervention. So what is the pushback for remediation? The pushback always is, well, what if the system goes down? So that's the pushback on patch management and remediation. And so what we are focused on is to say, well, how can we build confidence that you can patch without worrying that it's going to be an outage? So the first part of that is how do we help you make sure that if you want to reduce the possibility for an outage, fix the least amount. So that's where this hyper -prioritization is important so you can get down to only 1 % that you fixed. Second becomes, well, is there an alternative without a patch that I can apply a configuration fix that will prevent the exploit and I have more time to patch later? So what we have done is leveraging these frontier models ourselves. we have created the ability to create mitigations where you can actually maybe just change the permission of a file, maybe you disable a service, but with that, the potential of a compromise becomes a lot less and you don't have to patch right away. And the third and the last piece there is to build confidence in the IT team's ability to deploy the patch without worrying about an outage. We took the half a billion patches that we have deployed in the last few years where we see every rollback, we see every error that we get, and we build an ML AI based patch reliability score that allows us to do that as soon as a vendor releases a new patch, we can actually tell you this is a high reliability patch and it does not need a reboot. So you feel higher confidence that I can remediate in zero days because it's a high reliability patch and I do not need a reboot. And so with that, we can, look, at the end we are managing risk. So it's not that we have to fix everything, but if we can take enough off the table, the potential of a compromise goes down significantly. So auto -remediation can fix 20, 30 % of your stuff, but that 20, 30 % can really reduce the potential of an entire chain of attack happening successfully.

Krista Case

>> Absolutely, because we're seeing that, right? These AI -driven adversarial attacks, they can chain together these vulnerabilities. So it's a great point. And I'm glad that you called out sort of how you're validating the confidence, right? Because that is something that we're seeing security teams, especially if they're trying to integrate AI as part of their processes, which is necessary to do in order to keep pace, they're needing to understand that they can trust the AI. So are you building confidence with customers?

Sumedh Thakar

>> Exactly, right, I think that's where the combination of the data, the analysis, the inference, and the AI models, because we built the harness, so we are part of Glasswing, we are in Daybreak, so we are basically collecting and leveraging these models of AI to build mitigations and to build safe exploits for customers. Because as you saw with OpenAI/Hugging Face, that whole thing, they don't care, the autonomous AI does not care about this, the system shouldn't go down. It'll try 10 different things, one of those could bring the system down. So how do we leverage this AI capability to build out exploits that are safe with higher confidence for the customers so that they can test themselves without having to worry about an outage. The same thing for the mitigation. High confidence mitigations that are tested, that are created by reverse engineering exploits through AI, and then giving those to the customers so they have much higher confidence remediation that they can do, and all of it is based on not just using AI as is out of the box, but building confidence around the AI capabilities so that you can use those for remediation.

Krista Case

>> That makes a lot of sense. And Sumedh, can you talk, maybe just double click a little bit more in terms of how you're validating these vulnerabilities. I know you mentioned that a moment or two ago.

Sumedh Thakar

>> Yeah, I think a very simple way to look at that is if there's a bunch of different doors that are there and the traditional approach of prioritization will tell you maybe these 20 doors could lead to a compromise, right? I think what we're now focusing on is saying, hey, we're just going to try to open those 20 doors and we'll tell you by the way, only five of those open. And that we are doing so that now you have a much more focused approach and that's where that exploit validation comes in where it's not a complete red teaming exercise, but what we can do is when you see a vulnerability, we send a safe payload. A safe payload could be as simple as saying, hey, can you resolve this DNS name? If we see the DNS resolution come through, we don't need to update your system, or write anything. We can tell we could compromise it. So now that gives you a higher confidence in saying, hey, by the way, my controls didn't work. I got actually an exploitation that happened, so that becomes my priority. Versus if we send a payload, EDR blocks it, firewall blocks it, and it doesn't execute, then you have more time. So that, again, reduces the potential of an outage because you're not fixing everything, you're fixing the one that works. And so that's where we do a lot of the validation, by running very, very safe ways to actually test if the exploit works.

Krista Case

>> That makes sense, thank you, Sumedh.

Krista Case

>> Of course.

Krista Case

>> And I know another area you had an announcement. This week, my understanding is around AI governance. We've been hearing a big focus on run time and how do we put these, you kind of described, a harness, how do we put the right guardrails in place, because enterprises are just racing to try to adopt AI. And so for our security teams, they need to not get in the way of that, they need to enable that. So can you talk to, I guess, maybe what you're seeing customers struggle with from that perspective?

Sumedh Thakar

>> I think we are at the early innings of AI, and so I think the initial challenge is shadow AI, right? And how do you figure out what's in my environment? And I think it's almost like not having a shadow AI problem is a business problem, which means that you're not actually using AI in the business, which is not good for business. So I think today it is very important that every business actually is using AI. And then we kind of were at the beginning where everybody wanted to stop shadow AI because they're like, I want to know who's using AI so I can block them. Then we saw the pivot to saying, well I want to know who's not using AI so I can send them somewhere else.

Krista Case

>> Yes, right.

Sumedh Thakar

>> So now we're kind of rolling back to saying, look, I need to be able to have the visibility of who's using AI, what are they doing with it, can I put some guardrails around it, where are my models, where are my GPUs, and I think I've done this long enough in cyber that no matter what new technology comes, it always has these components, which is discovery, whether it was cloud, whether it is going to be quantum in a few years, the question will always be the same, where do I have it? Can I assess it, I cannot fix everything, can I prioritize it, and then at the end, if I don't get it fixed, what was the point of this dashboard tourism?

Krista Case

>> Right, right.

Sumedh Thakar

>> We were just looking at dashboards, we're not doing anything. So that same thing is happening with AI. So what we did with TotalAI is to leverage the Qualys footprint customers have and create the ability for them to figure out their shadow AI, what's happening in the client machines, where you're running AI in the cloud environment, et cetera. So you can first get the visibility. Once you get the visibility, we can test the models, we can see if the model is doing something that it should not, and then we can help you build out policies to be able to say we have some governance capability where unauthorized AI engines are not being used, the AI engines are not being fed data that they should not be. So it's something that is going to come up more and more in the next few months as a priority for customers as soon as they get through the current challenges that they have. AI security and governance is going to be definitely something that is going to be important for a lot of customers.

Krista Case

>> I agree, Sumedh. And we're certainly seeing that as well, that customers are trying to turn the lights on, right, and kind of solve the shadow AI piece of it. When you think about the stack that's needed to govern AI, what do you think are going to be maybe some of the key control points?

Sumedh Thakar

>> Yeah, I think there are a couple of control points here that are very important. First, discovery and inventory, having an upgraded inventory. Then you have the hardware components of AI. I think we also have to think about that some of these are not necessarily a security issue. Some of these are just a business issue. How many tokens am I spending is more important sometimes than the security part of it. So the inventory is something that is going to be important. After that, do I know all my models? Where are they running? It becomes important. Testing the models to ensure that they're not giving you answers that you should not be giving becomes important. But then the last piece that is also very important is what data is being used to train these models becomes very important. So if I know a model is being trained, is it being trained on my employee information? That should not be. If it's a support chatbot and I ask it, what's the employee salary? is it going out and talking to an MCP server to give that information, that becomes a control point. And then at the end, we are going to have to produce reports. For regulators, we're going to have to produce reports for the board to show we know where our AI is, to show we have tested it, to show that we have actually put guardrails around it, and that we are continuing to monitor it, right? But at the end, AI is one risk. You still have cloud risk, you still have identity risk, you still have misconfiguration risk. So all of that needs to come together in a singular risk view, where what we really see today is the building of a ROC. Like you have the SOC, which is after breach. Proactive risk management is Risk Operations Center. So how does AI security plug into that? And then in three years, it's going to be quantum security. But at the end, companies have a limited budget, and where should they spend that money is going to be based on where is the risk. If AI is the risk, spend more on that. If quantum is the risk, spend more on that. And the ROC actually helps you get that visibility.

Krista Case

>> Walk us through the ROC, because I know I've had some conversations with your team about the Risk Operations Center and how it's complementary to the Security Operations Center and doesn't necessarily displace it. But Sumedh, I'd love to hear from the horse's mouth today.

Sumedh Thakar

>> Yeah, we are having really great conversations because the SOC, I know I described that as snipers, right? They are focusing on somebody in my environment now and how can I get rid of them? That's their main focus. Because they are not thinking three months, six months from now, et cetera. And so what has happened is that we have done a good job in creating SOCs with SIEM and XDRs and looking at log data, et cetera. But when it comes to proactive risk management, the question is, I have cloud risk in one dashboard, I have identity risk in a different dashboard, I have on-prem risk in a different dashboard, and so we end up with dashboard tourism. I just have dashboards. But if I ask you, you have a business unit that makes 500 million a year, what's the potential dollar value loss on that business unit? You cannot say because every risk with a different score is in different areas, right? So the idea of the Risk Operations Center is how do we elevate the conversation of risk? How do we take vulnerability as only one component of risk? Second component is misconfiguration. Third component is identity. Bring all of that, harmonize all of that, and then get into threat intelligence, business context, remediation, and compliance, all to be very focused on risk itself, and then acknowledge that risk management exercise has three components. One is mitigation proactively. Second is consciously accepting a certain risk. And the third is what you cannot accept and what you cannot mitigate, you need to transfer to a cyber insurance company. So the ROC will allow you to have a conversation with your cyber insurance company to say, here's my sustained risk level, here's the additional residual risk that is left that I cannot accept, and what is the cost of transferring that risk to a cyber insurance company. So the Risk Operations Center is more of a strategic plan for CISOs over the next couple of years, is implementing a ROC so that they can answer the question that the board wants, they don't want to know the number of counts et cetera, the board just wants to see, is 90 % of my revenue under my risk appetite?

Krista Case

>> And that's, it really kind of cuts to the chase and it puts it in those board level terms.

Sumedh Thakar

>> Exactly, yes.

Krista Case

>> So you kind of referenced that you're having some great customer conversations around the ROC. What are some of the points of visibility, if you will, that you think security practitioners don't have today that the ROC is giving them? Is it kind of connecting those points that you mentioned?

Sumedh Thakar

>> Yeah.Well, exactly, right? So if you have an application that's making you 500 million a year, there's a separate tool that looks at the cloud piece, there's a separate tool that looks at identity piece, and a separate tool that looks at configuration piece. And so if you are asked the question, you have individual visibility, but if you say how much loss could I have if there is a ransomware attack on this 500 million business unit, and where is that risk coming from, you cannot explain that because it's distributed across multiple tools. The good thing about the ROC is that it takes data from existing tools, pulls it into a single platform, normalizes the different risk scores, because some are on 10, some are on 100, normalizes all of that, and then puts them in the context of your overall risk appetite. And so the visibility is there from an individual perspective, but then the CISO's job is to really bridge that gap between technical findings to what it means to the business. And the ROC really enables giving them that visibility of having those technical findings translated into business language and explaining to the board how much dollar value loss it could have and then where are we in our journey to ensure we reduce the risk of that loss. There is nothing like zero risk. So, I always say to people, by not doing anything, you're already accepting a risk.

Krista Case

>> Yes, absolutely, like we were talking about, businesses are recognizing that they have to accept a certain level of risk, and I think it's really powerful to be able to translate this into dollars and cents, because it's really difficult to do when we think about something like a ransomware attack in your example.

Sumedh Thakar

>> Yeah, yes and no, at the end of the day, the reason you have to put cyber in dollars and cents is because cyber is being used to protect dollars and cents. So that's like when you get car insurance, you have to say, I'm going to pay $1 ,000 to insure a car that is worth $50 ,000, right? If you don't know what the car is worth and you don't know what you're trying to protect, then you are taking infinite risk. And so I think it's always been warranted. It is becoming more and more important now that you cannot fix everything. You have to be able to explain the prioritization in terms of saying, why am I focusing on this or why am I not focusing on this? Because the business loss is not there. And I think we have now built up over the last 20 years enough data points through insurance and ransomware payouts that we have data available as we partner with a company that gives us that kind of information where we can say, if you are a business that is in this segment, in this geographical location, on average, how much ransomware payout has happened for industries in your business? And so you can project from that.

Krista Case

>> I see.

Sumedh Thakar

>> If I have a $500 million business unit, typical ransomware payout has been around 10 million. Okay. And so now you can put a number to it. So do you want to spend 10 million? No, but maybe you could say, look, if it's a 10 million payout, probably I want to spend a million to make sure that I don't have the 10 million payout. So I think that more and more of the data points are available and we're bringing it together, but I think I'm excited to continue to work in that direction.

Krista Case

>> Yes, yes, certainly. So Sumedh, you've kind of described yourself as a frontier AI optimist, right? And I think there's a lot of potential doom and gloom here, but reading between the lines, does that indicate you think as an industry over the next, call it, six to 18 months, that we're going to have some practical learnings and really kind of make some progress in terms of our ability to prioritize and respond, everything we've been talking about today?

Sumedh Thakar

>> I am very much optimistic when it comes to that. I think unfortunately some of the frontier AI rollouts have created sort of haves and have nots, and so there are geographies, countries, and industries where they don't have access to some of these models. The good news is that platforms like Qualys and some of the other cyber platforms are able to leverage these models to create more and democratize the outcomes of these frontier AI models so that customers can leverage this within their environment without themselves having to invest in using, leveraging these models. And so I think I'm optimistic because, yes, attackers are using AI for creating attacks, but defenders also have access to similar technology, just looking at it from a different perspective. And so one of the optimistic views that I have is, in theory, if every organization that writes code has the ability to run this kind of frontier AI model on their code base, they will find all their vulnerabilities themselves, and you will never have zero -day vulnerabilities. because a zero day happens when an attacker knows something you don't, but if you're running AI models you will always know. The flip side is that then the risk moves towards the zero day remediation because as soon as you release a patch, they will reverse engineer using AI, but the good news is that we at Qualys can also do the same. Once a patch is released, we can reverse engineer using AI, find the exploit, and then use that in a positive way to create compensating controls and to create a testing methodology that is safe. So I do feel like AI is definitely bringing positive outcomes for cybersecurity for a lot of organizations, and they don't necessarily themselves have to be investing in running these AI models. They should hold their partners and vendors responsible for ensuring that the appropriate amount of AI is being used to give them an outcome, which is at the end of the day what they care about.

Krista Case

>> And that's a really fair point, Sumedh. Maybe a concluding thought. So if I'm a practitioner and I'm evaluating vendors for trying to help me navigate my risk management. There's a lot of AI washing, especially if I'm in the CISO role, I'm being just inundated. What would you look for in a vendor to sift through that noise and really make sure that this vendor is going to actually deliver something that's actionable?

Sumedh Thakar

>> That's a great point. I think you have to, before you even go look for a vendor, you have to ask yourself, what is the outcome I'm looking for? And can I have a solution that will find malware is not an outcome, right? If I'm looking at business risk, I'm looking at, for me to avoid a $500 million loss, I'm going to create four layers of controls as my defense in depth. And these are the four ones. And then you have to be able to look at the vendors and say, look at the end of the day, AI or not AI? When you're looking at a vendor, there are three things everybody's looking at. Number one, is it accurate? Because if the solution is not accurate, it doesn't matter how cheap it is. Number two, is the solution fast enough? If it's accurate but takes six months, it's of no use. And then the last point is, is it cost effective? So is it actually helping me do the first two but without having to spend more money than I earn? And so I think when you're evaluating a vendor, those outcomes are what you should be looking at. And the vendor who's leveraging AI the right way, and, for example, there are vendors in our space that are charging the customer by the token. Now as a customer, you don't want to deal with that. You're looking for an outcome. So what we do as an example is we create a harness in the back end that picks the appropriate model. So we're not charging customers by the token. So they are getting an outcome that is a predictable outcome from a cost perspective and they're getting accurate, fast, and cost effective ways that they can actually take to resolution and that ultimately I think is what everybody should be looking at is what is my final outcome in risk management and does this vendor help me do things that are accurate, fast, and cost -effective?

Krista Case

>> Well, Sumedh, thank you so much for sitting down to walk us through all this today on theCUBE. We really appreciate it.

Sumedh Thakar

>> Thank you for having me. It was a real pleasure. Thank you very much.

Krista Case

>> Absolutely. And thanks so much for joining us. We'll be back in just a couple minutes with some concluding thoughts from day two here at Black Hat 2026. Thank you.