That's now. Businesses are adding AI agents to their workforce. Agents that access systems, make decisions, and complete work autonomously. These agentic teammates are doing more than ever at machine speed. But when they go rogue, the consequences scale just as fast. Thousands of agents, thousands of workflows, one connected system. One mistake replicated everywhere. Resilience has to evolve alongside AI. Not just protecting the data, but understanding your agents, knowing what they access, confidently restoring them, and what they rely on. Agent resilience means you can adopt AI with confidence, point in time recovery, fully restored. Cohesity, the goat of AI resilience.
Good morning, good afternoon, and good evening. Welcome to Cohesity Catalyst, the AI and data security summit. I'm Emily Chang and I'll be your host today. We're diving into how to protect your systems, your AI agents, and your data, and how to build enduring cyber resilience from the ground up. You'll hear from the builders, cyber security experts, and business leaders shaping the industry and leave with battle-tested strategies that you can put to work. We've got an incredible lineup. Sanjay Poonen, Cohesity CEO and president. Chris Kempczinski, CEO of McDonald's. Jason Clinton, deputy CISO at Anthropic. Thomas Kurian, CEO at Google Cloud. Dr. Matt Wood, chief AI and technology officer at AWS. Carrie Briski, VP of GenAI at NVIDIA, and more great speakers, including a lineup of customers from CVS Health, RWS, and Singtel. And we're really excited to welcome our guest speaker, Brett Goldstein. You know him, of course, from Ted Lasso and Shrinking. Brett is an Emmy award-winning actor and writer. He'll be sitting down with Cohesity CMO, Carol Carpenter, for what's sure to be a fun, provocative conversation. Before we dive in, this is a fully interactive event. We invite you to use the chat, tell us where you're joining from, share your ideas, post your questions in the Q&A panel, and participate in the polls. Our experts are online to provide answers to your most pressing questions. So grab your coffee and let's kick things off. I'd like to welcome Sanjay Poonen, CEO and president of Cohesity.
Hello everyone. Welcome to Cohesity Catalyst. I'm Sanjay Poonen, CEO and president of Cohesity. I'm going to talk to you today about the cyber resilience imperative. We're living in an incredible time. People call it the mythos or the post-mythos moment where AI models have gotten to be moving at incredible speed, machine speed. They are significantly more dangerous and can do a lot more damage than humans can. Many of the type of attacks we've seen are now also agentic if you look at what's happened most recently with the Hugging Face incident. So what you have to prepare as you think about your posture is you go back to this NIST framework which shows the left hand side of it being detect and prevent and the right hand side of this being recover and remediate. I like the adaptation of the Benjamin Franklin line. An ounce of resilience is as important as a pound of detection which means you must invest in detection prevention technologies like CrowdStrike and Palo Alto. They're key partners to us. But just as important and many of our customers, the last line of defense is Cohesity, where we are their recovery and remediation solution for all of their data. Now, why do people specifically pick Cohesity and why we were built for this moment? Number one, we were built on a zero trust platform. That data that's ingested from various different sources is an immutable platform. That platform has the fastest cyber recovery of anybody. The scale at which we work. The largest companies in the enterprise in banking and public sector in the tens and hundreds of petabytes rely on us. The largest companies rely on us for that resilience. And we ensure that our platform has the highest level of cyber security. And these are not just banks and public sectors but critical infrastructure that many governments rely on. One of the frameworks we have seen that's very important for CIOs, CISOs, but quite frankly even for boards and CEOs to adopt is a notion of building a minimum viable company or minimum viable government if you're the public sector. What does that mean? You look at all of your applications and you categorize them with the applications that need to come back in an hour if they're most important maybe like your identity systems in a day, in a week, in a month. And then those apps that don't come in a month, you could probably not worry about because you're not using them. We've studied hundreds of customers and we've come up with a beautiful model first and only in our industry called the five steps of cyber resilience. It starts with first off ensuring all of your workloads are protected. Starting with identity. Secondly, building a cyber vault which is airgapped. Number three doing a number of threat scanning, anomaly detections and sophisticated security steps on your data. Number four, practicing application resilience in an agentic approach. And number five, discovering all your data, building security posture management and an AI security posture management. In fact, that step five could often become the first step. So it doesn't matter if you start with step one or you start with step five. In many cases, customers are discovering their data first, what's here in step five, and then driving a data protection process. We're very fortunate at Cohesity that we were the first company in our space to enter Project Glasswing with Anthropic. We've had a strategic partnership with them but also OpenAI with GPT 5.5 cyber in Project Daybreak with Google Gemini 3.5 who Google's an investor in Cohesity. We were the first to drive a partnership with NVIDIA in our space in open secure AI alliance. We're the only company in our space that's working with NVIDIA and Microsoft on the open weights and American AI leadership initiative. The reason this is important is you want a company that's in the inner circle of these frontier labs and the hyperscalers working with them to understand what the threat vector lies and using their models to ensure our code is fully fixed and scanned with the latest of these technologies. And we're the only company in our space doing that. So, we'd encourage you to work with us as we build a future for you where you can play defense on your data. That's resilience. But then you can use the power of AI to also play offense for your strategic future. Thank you very much.
Thank you, Sanjay. You've just heard why cyber resilience is so important in today's AI age. The strategies for security and data protection that we've relied on today are going to change a lot. Now, let's bring on someone with firsthand perspective on how Frontier AI is impacting cyber security and what organizations need to know about the escalating threat landscape. Jason Clinton, deputy CISO at Anthropic, sits down with Sanjay to discuss the rise of machine speed attacks, what happens when AI agents are compromised, and Project Glasswing, a defensive cyber security initiative that Cohesity is honored to be part of and support.
Hello everyone. A warm welcome once again to Cohesity Catalyst. I'm here with Jason Clinton who is the deputy CISO of Anthropic. We're actually at the Anthropic building. Jason, well, welcome. Thank you for being with us.
Why is this time different now? What's different about the threat vector?
Yeah, I mean I don't need to convince anybody listening to this that this is a big change in 2026. We've had the quote mythos moment. We've seen advanced cyber models now from Anthropic and broader places in the ecosystem. And so, you know, anyone who had any doubts before that models were going to be contributing to both the offensive and the defensive, you know, from a dual use perspective side of the challenge that we're facing from cyber security now has a lot of evidence that it's true. We've seen thousands of vulnerabilities fixed in major products from major vendors that everyone listening to this is familiar with. So if you just think about what it means for the average defender today they have to plan for a world where there's this vulnerability tsunami coming at them and we are seeing advanced cyber models now which are open weights as well and accessible to anybody to be used on the offensive side to drive a fully autonomous attacks and because of this every security leader has to think differently about their security program. I think that's a phase change in industry that we should all be planning for.
The nature of these attacks are more deadly and they're moving at much faster at machine speed.
Yeah. So we have like a very little time to prepare for what's coming. And so Glasswing and Mythos and you've been a great partner in that as well. You know, as we've been thinking about enabling defenders to move quickly, we need to enable a world where every company is prepared for breach. And so I have like three very specific recommendations that I want to hear them. Obviously if you're designing for breach, you need to design for resiliency. So, so products like yours which can help people recover and decrease that mean time to recovery is an important part of any defensive posture. And then you know you just think about every one of the defenses that we put in place as security leaders has to include some kind of detection response component. But we also need to eat our vegetables. We just need to fall in love with the basics again. So there's a lot of things that we can do to prevent the bad guys when they get in from moving laterally. We can use detection response. You can use AI systems as a part of the defense and then vulnerability management and I know you all have done an incredible job trying to figure out where there are vulnerabilities in binaries and where we top. Yeah I think those are the three things like vulnerability management, have a robust detection response program where you're moving at machine speed on the defensive side because as that model intelligence goes up for the attackers the model intelligence will go up for the defender at the same rate so you're having a good alignment between those two curves and then the final thing of course is just thinking about resiliency and doing the basics and to prevent us from having those long outages when the bad guys do.
Well I want to pick up exactly on that last point which is assume breach and then have a resilience process which is that's all we do you assume it's a matter of time before some big US brand could get hit. You have to assume you're going to get breached and then have a fast recovery plan. Talk about both the left side of NIST which is detection and prevention which is where a lot of Anthropic's work is and how important the right side is recovering resilience is and you covered both those in those two pieces.
Yeah. I mean both are important. Starting with detection if you just think about what is happening today with every company trying to figure out how they can automate their SOC. There's so many examples of SOC startups. I think there's like 60 of them now. Those are examples of a very simple and straightforward observation that a model is extremely good at doing analysis of any kind of alert that's coming through. So you can do false positive rejection at a much higher and higher fidelity rate than a human being would have been able to do previously. So autonomous SOC is like a huge part of the solution space today. My perspective from the Anthropic side is how can we scale the defenses with the amount of the flood of attacks that are going to be coming from these new models. And so sometimes this is called automated remediation. In my view that's part of it but the bigger picture is what can we do to isolate and limit the amount of damage that the attackers can do inside that detection pipeline. And then on the automated recovery and the process of driving remediation, Anthropic is in the situation now where you may have seen in our blog post two weeks ago we just shared that we used Claude as a part of the incident response process where we're combing through literally petabytes of logs trying to find all the places where something might have gone wrong and using that to drive our remediation process. That can feed into recovery. That can feed into database backup restores, knowing where you have damage and limit and understanding the full blast radius, faster than any incident response or even like a crisis response team, assuming you have a breach and you bring in an outside firm, it might take them months to tell you what's wrong. Today, you can have a model tell you in a couple of days at the most.
How do you view Anthropic the future of agents and either the unintentional malicious damage they may be able to do? I mean obviously I want to talk about the Hugging Face incident. How should CISOs and security practitioners be thinking about the future of agents and their security?
If you're thinking about agents that are as you said you're on behalf of agent, your sidekick. That thing is using your credentials. It's using your credentials to transmit your OAuth tokens to whatever data stores your corporate network is using. And that is a very well understood and very straightforward thing to sandbox and understand what the blast radius is. And then you have the second thing that you alluded to which is we have these agents that are long running persistent. Some kind of business process runs in them. They're running as a system service account. They might be the account that runs your automated code scans for your CI/CD system or it might be the account that you use for customer service agent or it might be an account that runs some finance accounts payable intake process. Like those are examples of an agent that has long-term memory that has a state that has limited permissions to only do the thing that that part of the business needs for that business to work well. And when that thing goes rogue or goes off of the distribution from the goals that have been defined by the business or it gets compromised by an attacker or otherwise damaged through any kind of a breach, there's different recovery and different things to plan for. But from a security controls perspective, this is back to that principle of least agency again. Like only give it the permissions that are necessary for that thing to do what it needs to do to do the job and don't give it anymore.
I'd like to switch a little bit to Mythos Glasswing. We were admitted into Glasswing and we got to work on two things. One was our customers began had been SC who were inside Glasswing were scanning our binaries already and were giving us vulnerabilities they found. That wasn't an alarming number, but we found them. And then when we got inside Glasswing, we were able to do source code scanning. And now we could do the binary. And to me, it was amazing, Jason, because I'd never in being in the software industry for 25, 30 years seen something that could work on binaries with effectiveness, do these chained kinds of vulnerability findings. And we made a commitment of course very quickly which were well underway to kind of fixing the vulnerabilities. But it's extremely powerful. We've been sharing back with Anthropic our findings. I think we were one of the few partners who used the most tokens during that time to get our code which we were grateful for.
What is your vision of how the future of these code scanning and vulnerability tools will play?
This is like maybe the most important question for strategic. Speak away. Yeah. So, the reason I'm saying it's the most important. So, imagine that models will keep getting cheaper. And I just want to draw a connection between a couple of dots to make that claim. If you compare the intelligence of Haiku 4.5 today to GPT-4 from three years ago, Haiku 4.5 is way more intelligent than GPT-4 was, but it's very cheap to serve fractions of pennies on the dollar for what GPT-4 cost to operate 3 years ago. And you can find multiple places in the world where there are benchmarks whether it's ARGID, polyglot, there's a couple of others where you just see the cost for a particular benchmark to pass going down down down over many years now of trend line and so roughly what you see is a roughly 10x year-over-year improvement in the total amount of cost to serve a particular level of intelligence and so you know I would make the claim that we will see advanced cyber capabilities in Sonnet class models in six months and then approximately 18 months from now we'll see advanced cyber capabilities in Haiku class models and so if you're thinking about this from a strategic perspective as a security leader especially in somebody in DevSecOps and you're thinking what am I going to do for code scanning and what the future looks like you should be planning for vul zero what I mean by that is there will be a point 24 months from now roughly when it will be considered malpractice in software development to not have incredibly cheap advanced cyber models running on every line of code that goes through the IDE, the CI system probably do blackbox penetration testing in the continuous delivery environment and at affordable pricing. And affordable pricing yes that's right and so the cost will keep going down because of these scaling laws implications.
Well there you have it folks we're just honored to be one of their top partners involved with Glasswing and Mythos all in on cloud code plot cool work, anything else you're doing, Claude, whatever it is. So, we look forward to this. Jason, thank you for spending some of your time.
Look forward to making all of you, our joint customers, successful with the integration and the best capabilities of both Cohesity and Anthropic. Thank you very much.
Hello everyone. I'm here with my really good friend Thomas Kurian. Thomas, good to see you. I've known Thomas for many decades. Short sort of story, he and I grew up in the same town in Bangalore. Our families know each other. He went to the rival school. We actually competed with each other at SAP and Oracle. And then I when I left and joined VMware, I said I was so happy to be finally his partner and friend first at VMware and at Google. We did some good things together. How you advise CISOs to think about right now they're all they're thinking about is scanning code. How would you get them to get their head out of that and saying would you think about the X and the Y and the Z dimensions of what could go wrong? You need a different approach.
So you know in general there are sort of three four important things that we talk to customers about. The first one is the risk from AI models is not just to source code of applications. The risk can affect binaries. It can affect firmware. It can affect a physical device like a network switch by compromising the firmware of that. And the reason is that models have gotten really good at understanding not just code but also configuration like for example they can understand your network and go if I access this switch then I can shut down this part of the network etc. So one thing is that models now have access to a lot more information and their ability to reason on them and so you need to think through what you are protecting is not just scanning your source code. The second thing that we say is if you look at the rate at which models are advancing, you're going to end up scanning continuously and that can get both expensive but also just finding an infinite number of bugs you need a solution to help you repair it. So that's the second piece is scanning without fixing is not a solution. The third one is when you look at the evolution of this whole infrastructure, the volume of applications, the volume of data, the volume of all the systems that companies run. If you then say I've got new models finding new things all the time, you then end up having to deal with an extraordinarily large surface area to patch, maintain, and update all the time. And so you can't just say I'm going to start at the start and go till the end because it's such a large estate of stuff. And the fourth complicating factor there's no one model that finds a superset of all the issues. In fact what many many companies are now reporting is a cheaper even a small open-source model with a good harness can find many of the vulnerabilities that a top-end model can do. So when you look at it, you know, you have these four issues. It's not just about source. It's about firmware, binaries, network, everything. Second, it's extremely frequent because you're going to have to get comfortable that models are going to come out all the time and just the volume of these issues is going to grow on you. The third one is just finding without repairing becomes challenging and when you repair you need to think about what are you going to focus on repairing and lastly no single model is the solution you have to have a conglomeration of models like for example saying this time I'm going to scan with this model next cycle I scan with another model and so those are the four core beliefs we have based on all the data we've seen and based on what we're doing ourselves to protect Google Cloud.
Thank you, Jason and Sanjay. Now, while most of the industry discussion emphasizes the threat, they made a great case there for how AI will make a real difference for defenders. For more industry perspective, I'm joined now by Rob Saddowski, VP of product and solution marketing at Cohesity to review the key findings from Cohesity's hot off-the-presses survey of cyber security and IT leaders on the current state of their resilience plans. Cohesity just released new research on cyber resilience and there's a lot here worth unpacking. So Rob, let's start with the scope of the study. Who did you talk to and how did you approach it?
We surveyed 3,200 IT and security leaders across 11 countries because we wanted a large enough sample to see patterns by industry and region, not just have anecdotes. So really the goal was after this study, we could speak with confidence about where organizations stand with resilience today.
What's the main finding that came out of all that data?
The big topline thing was that the resilience plans that most organizations have in place, they really weren't built for the environment they're actually operating in. Now, the plans were often written a few years ago under different assumptions about incident response and recovery speed. And the gap between plan and reality is putting businesses at significant risk of disruption.
Where specifically are these plans falling short? Walk me through it.
Well, the first thing is that the plans assume that incidents are going to be fully contained before recovery starts. So, they think contain first and then restore. But in practice containment and recovery they increasingly they have to happen in parallel. Second big thing is that many organizations they just don't have a complete picture of all their operational dependencies. So they know what needs restoring but not everything downstream that it actually depends on. And finally recovery is kind of stalling in places that nobody actually anticipated.
What does that gap actually translate to in practice in terms of real business impact?
Where it shows up is directly in their recovery time. So the actual execution of recovery is running almost twice as long as planned. So that's the difference between a contained incident and a prolonged outage that can affect their customers, their revenue, and their reputation.
Two times longer than planned. I mean that's a striking number. Let's talk about another concept that comes up in the study. MVC, minimum viable company. Explain what that means and how that can make a difference.
Sure. So the MVC or minimum viable company, it identifies the smallest set of systems, applications, and data that a company needs to keep functioning during a crisis. So what it has the effect of doing is narrowing the scope instead of treating everything as equally urgent. So you can prioritize recovering what truly keeps the business running first. In MVC, what it does, it really makes recovery efforts focused and realistic rather than having to do everything all at once under pressure.
How widely is that actually being put into practice across organizations?
Not enough. Unfortunately, only about a third of organizations, 37% have actually defined an MVC. And of those, only 22% have actually tested recovery under realistic conditions. So what I'd say is that the concept is gaining traction, but proven tested readiness is really still rare.
Let's talk about the shift to AI because of course it is reshaping every industry, every business function. How is that impacting resilience planning specifically?
What it's doing, it's adding a new layer of complexity and most organizations are actually pretty behind on this. So what we found in the survey is less than half of organizations actually had a good inventory of where AI is being used across the business and only about 40% have incorporated AI infrastructure into their resilience plans. So even where usage is tracked, it's often left out of the recovery conversation. So there's a visibility gap first and then a planning gap on top of that.
Exactly. And it goes further still. More than half of organizations aren't confident that they could recover from actions taken by a rogue agent. So I think what folks need to take away is that as AI becomes embedded in their daily operations, this gap is going to be something that needs a lot more attention. Now, given everything you described, the plan, the reality gap, the MVC gap, the AI gap, are organizations actually responding or right now, is this still mostly just awareness?
Well, there's good news here. There is genuine awareness and it's translating into intent. 83% of organizations say they'll need significant, not incremental changes to their resilience plans. And for me, that's a meaningful signal. You know leaders recognize this isn't a matter of incremental change or small tweaks. It's going to be something significant but the real question comes in execution. How quickly they can actually close the gaps that they have identified.
So what would you recommend organizations focus on first practically speaking?
Practical I think is the really important word there. So the most important thing they can do is define that minimum viable company and test its recovery under realistic conditions. As they do that, they need to pay particular attention to dependency mapping, especially the identity system because that's where we see that recovery often quietly stalls. I think the second thing is to extend those resilience plans to formally account for AI systems and agents. Don't treat those as an afterthought. And finally, benchmark your progress against a structured model, something like Cohesity's five steps of cyber resilience, because this makes improvement something that organizations can measure over time, not just assume is happening.
What's a closing thought that you would leave with organizations as they work through all of this?
Yeah, I'd take it right back to where we started and really the title of our study. If your resilience plans were built even a couple of years ago, they likely weren't designed for today's environment. So get it updated. And the best positioned organizations are going to treat resilience as an ongoing tested practice, not something that lives in a static document, gets filed away, and revisited only when something goes wrong. And if you'd like to learn more, here's a QR code where you can download the full research report.
Really valuable perspective, Rob. Thank you so much for walking us through all of this.
Thanks for the great questions.
Now, we're going to switch gears. While we've been focusing on how AI is impacting security and resilience, how is it changing how companies do business, Sanjay had the opportunity to recently sit down with McDonald's CEO Chris Kempczinski. Let's hear more about how a Fortune 500 company like McDonald's has figured out how to scale AI responsibly. And after that, we're going to hear from CVS Health about their AI and resilience journey.
McDonald's is such an iconic brand, over 70 years old. But what's been amazing about the company is how you've reinvented yourself not just as an incredible restaurant experience but also adopting technology as you think about AI and its transformational potential to serve your 70 million odd daily customers coming in. What are some of the ways in which you guys are adopting and think about AI strategically?
The end of the day to stay relevant as a brand you've got to have that customer first mindset and you've got to be sort of anticipating and understanding where they're going and evolve your business to do that. And the opportunity with AI is for years and years we would have that sign out front billions and billions served but the reality is we actually didn't know who we were serving. They were all anonymous and now we actually do understand and know who's coming to the restaurants. So, how do we use AI at scale to actually start to customize the experience, give a consumer exactly what they're looking for? If you put yourself in kind of the shoes of a restaurant manager, they've got all these disparate systems. So, they've got an inventory system that they're managing. They have a crew scheduling system that they're managing. They're trying to pay attention to the drive-thru, all this different data that's coming. And today, it sits in different systems. And one of the things that we're spending a ton of time on is actually getting we call it kind of one pane of glass but getting all of those systems integrated. We think there's going to be a huge opportunity with AI. When you look at our restaurants, the difference between a top performing restaurants, top quartile and a bottom quartile is quite significant. AI is going to give us the opportunity using these tools.
Let's shift a little bit as you think about the audience here are people who are listening to this CTO, CISOs, CIOs, technology people. I'd like to get to general points of leadership, but what advice do you have for them? Many of them are very worried about in the cyber world agentic attacks. Yeah, they're looking now at AI as a productivity. How would you advise other technology leaders? The way you advise your own people and how they should be thinking about their posture in this era.
I think for in big legacy companies, the CIO or the CTO wasn't viewed as sort of being a P&L driving entity. It was more just kind of helping the trains run on time. And now with AI, the CTO, the CIO, you are like right in the middle of how a company thinks about the business going forward. I think that's the biggest opportunity right now that I'm certainly asking for my CTO is how do we help start to reimagine the company because as we were talking about so much of AI is not just about the technology, but it's re-engineering how we do things. It's making decisions on how you think about competitive advantages as I was talking about. So I think there's a huge opportunity around education. I think there's a huge opportunity also for CIOs, CTOs to actually be total business partners.
One of the things amazing about McDonald's is the leadership principles from 70 years of incredible brand experience. But as you came up the ranks as a leader, that you would tell the younger, you know, Sanjay and Chris from 20, 30 years ago, these are the kinds of leadership qualities that are important to leaders today.
Your mistakes, your things that you missed, those can be some of the greatest learning opportunities. And so that I think is one thing is, you know, don't be worried about kind of the, you know, am I going to make a mistake? What ends up happening is if you've got that mindset, you're always playing defense in your career and you're not taking risks and you've got to take risks. The other thing that I would say I've spent, you know, earlier in my career, I would spend a lot of time, you know, if I got offered
An opportunity or whatever, you know, thinking about like, okay, what am I going to get out of this? Is this going to help sort of advance my career? I way overthought things. Way overthought things. You know, if you get presented with an opportunity, your default answer should be yes. The default setting has to be yes, I want to go take that opportunity. Because no matter what that opportunity is, if you show that you can do that and do that well, it's just going to open up more opportunities.
When you think about the leadership principles that McDonald's has today that over the many decades, McDonald's has made great leaders out of. What is one or two of those that Ray started the company with?
Yeah. One is put the customer and our people first. And that's number one for a reason. We're a team. And that we operate as a team. We're a franchise business. 95% of our restaurants in the US are run by franchises. We don't own our own distribution logistics. We don't own any farms. We don't own any manufacturing facilities. All of that is done through partners. And so we're constantly thinking about yes, the company needs to win, but I need my franchises to win. I need my suppliers to win. And so I think in this business, those two in particular really resonate and have been key to our success over the last 70 plus years.
Great. All of you probably watching this have had chance grow through the golden arches and get something McDonald's. But when you do go there next, I hope you'll remember aspects of this that are good takeaways. I take away certainly this customer first mindset offering as a team that applies to all of us, not just McDonald's, but also Kist. Thank you very much, Chris. Thank you. It's been great.
For a long time, resilience meant uptime. Here at CVS, distribution just isn't systems going down. It could be patients at risk, a clinician that loses access to certain records, members waiting on a claim. For us, it's a patient safety conversation. The core principles that really shape our resilience and our recovery strategy is recoverability by design. We don't wait for an actual incident to think about the recovery. We operate on the assumption that adversaries will get in. The question becomes how do we prevent access to what matters even if they do get in and how we can recover clean. Resilience is just not an infrastructure problem. It takes application teams, security and compliance and executive leadership all working from the same playbook. We think about what attackers actually want to do. Encrypt production data, destroy backups. So we design against that immutable copies, anomaly detection and backup activity. Tight integration with our security operations center. Security works to keep adversaries out while resiliency makes sure that we can recover if there is actual impact. You really need both. Resilience without security is really wishful thinking. I push back on the idea that compliance slows you down. HIPAA, CMS, state pharmacy boards, those exist to protect patients. You treat them as guard rails that shape good architecture instead of roadblocks.
The guidance I would actually share with other executives, first test your recovery. A lot of organizations have RTOs and RPOs on paper that have never been validated. If you haven't proven you can recover under realistic conditions, you can't recover. It's that simple. The second, make it a leadership conversation. Your board really needs to understand recovery in business terms. The third and last one I would say is invest in immutability and isolation. Ransomware actors are going after backups. Now if your recovery copies aren't isolated, immutable and validated, your whole strategy has a single point of failure. AI creates a new surface area where we have to protect. Traditional backup and recovery wasn't built for that. Interface workloads, training pipelines, agentic workflows, making autonomous decisions. If this AI capability goes down and gets compromised, can we recover not just the service but the trustworthiness of the output? You need to know your training data wasn't poisoned and your model weights have integrity. We're treating these workloads the same as any critical platform. AI doesn't get a pass just because it's new. I think resiliency is a journey. The threats evolve, the technology evolves, your business evolves. The organizations that get this right treat us as a continuous discipline. That's the culture we're building here at CVS. And the partnership like the one we have with Cohesity is a big part of making it real.
All right. Now, we've got a lot of context, both industry and business. So let's delve a bit deeper into the technology. AI agents are being adopted at an incredible pace and as that happens they have to be included in resilience plans. I'm going to turn it over now to Swami and Teresa to tell and show you more.
If you are like most organizations, you are already rolling out AI agents. These agents have access to sensitive data and need to be scoped carefully as they are integrated into your production workflows. But here is a question. A few months from now, will you know exactly what they have access to? What memories and context they have built up along the way? And if one of these agents starts operating on bad context or poison memory, can you roll it back? Can you perform a precise recovery? That's what we are here to talk about today. How to bring cyber resilience to the agentic enterprise.
AI agents are moving from pilot to production faster than any technology we have ever seen. And with that growth comes new challenges. Most organizations can answer basic questions about their core infrastructure. What's running? Where the data lives and what's protected. But agents are constantly evolving. They pick up new permissions, interact with your data, and accumulate memory and context over time. Keeping track of all of that is not easy. The industry is responding to this challenge with agent control towers, governance tools, and observability platforms. Those capabilities are important, but as agents take on more responsibility, visibility and control aren't enough. The stakes are high. Agents just don't recommend. They execute. They interact with data, trigger workflows, and make decisions with varying but increasing levels of autonomy. A single malicious prompt or coding error can trigger actions that are disruptive and expensive to recover from. Cohesity is investing alongside that observability and governance ecosystem enabling you to recover if the agent state is corrupted or compromised.
First you need to protect your agentic infrastructure. The platforms and components used to build, run, and manage agents. Second, you need to protect what agents access, applications, data, and the file systems they interact with. Our solution is agent resilience, which is now part of Cohesity Data Cloud. Agent resilience delivers a precise recovery path for both agents and the resources they impact. And it's designed to work consistently across agent platforms that enterprises are adopting today. We are starting with AWS Bedrock with support for Gemini Enterprise, Microsoft Agent 365 and Anthropic planned as well. Here is how it all comes together. Agent topology gives you a living map of every agent. The system it connects to, the data it can reach, and the dependencies between them. This allows you to understand potential impact before something goes wrong, not after. Next, we protect what no one else protects today. The agent state, its memory, its configuration, and the workflows that shape how it decides and acts. Most organizations aren't protecting any of it today. And that's the problem as agents become mission-critical. If something does go wrong, we allow you to recover precisely point in time with human in the loop back to a known good state. You don't have to rebuild the agent from scratch. Teresa, let's see how this works.
Thank you, Swami. Let's look at this from the viewpoint of a healthcare entity. As an IT admin, it's my responsibility to verify that I can recover applications according to my business level SLAs. One of my most critical agents built on AWS Bedrock reads important patient records and summarizes discharge notes. Leadership wants verification that I can recover if something goes wrong. Let's go ahead and take a look.
Navigating to agent resilience, I get immediate visibility into my agent topology. This gives me visual insight into all of my AI agents in a matter of moments. Reviewing my healthcare app specifically, I can quickly see all of my components, my agents, and I can see it's not being protected. If this agent starts modifying data or deleting records, I'm in trouble. I need to protect it. Now, within Cohesity Data Cloud, I open cloud protection and I choose sources. Well, there's already a list of registered sources here. AWS Bedrock is not one of them. So I'll go ahead and register the source. I will then choose AWS and start the registration. Since our agent runs on AWS Bedrock, this will be my selection. It's important to note that I'm not deploying anything into the agents runtime. All of this is agentless and API based. My next step is to create an IM role. This creates the right visibility and is done by downloading a cloud for template, uploading it to my AWS account, and confirming it's done. I then simply click register. Now behind the scenes, agent resilience starts discovering everything the agent depends on. Its memory, its runtime, and mapping it all out. It's that simple. My AWS Bedrock account is now managed.
Let's next start protecting my agent. Here we can see the configuration and the memories have all been autodiscovered. Agent memory holds conversation history and context. If it's poisoned, corrupted, or lost, the agent can behave differently. Even if the underlying platform is running fine. To protect, I select my AI agents. I then verify all of the policy settings. Everything looks right, so I'll go ahead and choose protect. And that's all. The agents memory and configuration are now protected.
A couple days later, I get a call from my incident response team. This is never a good thing. They're tracking suspicious activity from the healthcare app and they're recommending recovery. I'm going to go ahead and select the healthcare app to get started. Based upon the time of the incident, I know that this is the last known good snapshot before the issue. So, I browse all of the objects and select recovery. Now, I want to keep the agents long-term memories. This is what will give the agent back its past experience. I'm going to recover everything. Next, I choose where to recover to. And within minutes, the app data is back to a verified known good state. No rebuilding the agent, no days of downtime, just a clean roll back. My work is done here. So, Swami, back over to you.
Thank you, Teresa. What you just saw is agent resilience in action, protecting and recovering the infrastructure behind an agent and everything it touches. Now, imagine doing that for thousands of agents across your environment because that's where this is headed and at that scale, resilience can't be a manual process. So stay tuned for our session where we explore what's next in autonomous cyber resilience.
Next up we have Dr. Matt Wood, Chief AI and technology officer at AWS, who's joined by Cohesity CMO Carol Carpenter. They break down what it takes to deploy AI agents at scale and how platform providers help keep it all safe. Right after that, we've got some insight from Singtel on sovereign AI and trusted data.
So, there's so much opportunity and you yourself have shown some examples of innovation. You see lots of customers innovating with AI. How does Amazon think about that? How do you think about that balance between, hey, let's move fast. Let's get these agents into production. Let's change the workflow and really make transformational things happen with how do we make sure the data is secure, agents are secure, there isn't rogue behavior. Tell me more about that.
Moving fast doesn't mean moving irresponsibly. And so, our approach is that we look and our customers that are moving quickest with AI are also following this approach. You look way beyond the model when you're building an AI system and the model is an important component but increasingly the machinery that you put around it to guide the agents, control the agents, guard rails before and after you prompt the types of tools that that agent can interact with. There you really want a robust set of controls governance around the model in order to be able to allow that model to act in sensitive environments in with critical data with increasing workloads that are happening at the core of organizations. And so that enablement isn't at odds with speed. In fact, putting those guardrails around your model actually allows you to go faster.
Yeah. And I've heard you talk about security being a high priority for Amazon and as well as building it in which is what you're talking about making sure it's part of the foundation.
Yeah. In addition to wrapping the model and the guardrails and governance around. That's right. So we also, one of the interesting things about agents are that they tend to collect up a lot more information as they're doing their work and they benefit from a diversity of information which usually means you want to give them access to more and more information so they can be more and more successful. What you don't want is you don't want to share that increasingly broad sensitive internal information with anybody else. In fact, usually you don't even want internal usage to be visible for very sensitive workloads. And so our approach here is maybe a little different from others. By default we have zero operator access for when those prompts are actually run and when the agents are actually computing. And so we have end-to-end encryption and then when we get to the inference environment, we actually have a constrained verifiable environment with zero operator access. We can't look inside that environment by default. The lab, the AI model labs can't look inside that environment. We have no visibility. We don't want visibility. Your own organization can't see inside that. So you can keep sensitive information safe there. And that's where we actually perform the inference inside an environment which has zero operator access. And the code that runs inside that environment is verified and at runtime. And so we have hardware security inside that environment which verifies that only our signed code is running. And so it's a secure environment with high degrees of privacy by default which means that no one else is inspecting your prompts by default.
I love that. I mean what a robust what you just described is incredibly robust and that's what can give enterprises and public sector customers confidence in what they're building and deploying. So something we're announcing here at Catalyst and we've been talking about is what we call it's Cohesity agent resilience. So in those hopefully rare circumstances when agents do go rogue or you need to destroy an agent or and the ability to roll back and recover the agent infrastructure as well as the data which as you pointed out can grow it should grow and grow with time. So we're really excited about that and to partner with you all. We're announcing it on Bedrock first. So hopefully customers will learn more about that during the day.
Let's talk a little bit about the shared responsibility model which has been around for a while with cloud obviously. You know it puts data protection, recovery and configuration squarely on the customer side of the line. And we just talked about AI and agents and how they bring more complexity into the picture. So how can partners like Cohesity and how can customers think about their shared responsibility? How should they think about it? You have infrastructure, Amazon infrastructure, Cohesity is providing recovery. What else do they need to consider and think about?
Well, I think importantly as customers continue to move into production, they move actually pretty quickly and those agents that are moving into production quickly move to the core of the business and become one more incremental system that you have to build resiliency around. And so this is not something that you necessarily get for free and it's not something that you should ignore, but these are business critical operations inside your organization. And you need to be able to inspect and manage and backup and restore and control and govern those core business capabilities just like you would a database or your ERP system or anything like that. And so the maturity of the tooling to support that is starting to appear and why we're excited to partner with Cohesity on that. But it's a requirement for business critical agents just like it's a requirement for any business critical capability.
If you could wave a magic wand and you could say to an enterprise, let's imagine it's a large healthcare company, you are going to have an amazing transformation due to AI and you are going to be wholly and fully resilient. What are like those three things your magic wand would do for that company?
It's a great question. I think number one, develop a culture of just try it. There is so much that you can do with artificial intelligence today that is useful and economically valuable and cost-effective and you should just try it. And by trying it you can almost see you can feel around an organization the light bulb coming on above people's heads. The second would be to ensure that your security, privacy, resiliency and governance is up to speed. Governance is like has a bad reputation is kind of a dirty word in technology. It does have government. I get it. But it is the thing that allows you to move faster, not slower. And having it in place is another predicate to just making progress more quickly. So getting that foundation in is really important. And then giving those early innovators the permission to go try within the bounds of governance just some crazy things. You will learn so much.
It's fantastic. Well, it is such a pleasure to speak with you, Matt, Matt Wood. Very insightful and we're really happy Cohesity and Amazon are partners. Thank you.
