Name: Agent Rewind: How to Safely Deploy Unpredictable AI
Uploaded: 2025-10-23T05:46:25.009Z
Duration: 1 h 8 s
Description: Agent Rewind: How to Safely Deploy Unpredictable AI

Transcript for "Agent Rewind: How to Safely Deploy Unpredictable AI": Welcome, everyone. Excited to be with you today. I'm Matt Pichelio with Rubrik. I'm joined by Sol Hill South also with Rubrik. And, we're here to talk about one of the most exciting topics of, of our times and, but also a concerning one as well, which is the frontier in AI known as agentic AI. As we know, agents can make decisions. They can take actions on their own. But, with that autonomy, becomes some risk. And so I'm, excited to be joined by the one and only Johnny Yu, who is a research manager at IDC. And, he's been tracking this space very closely, from an AI adoption perspective and really excited for his insights. Johnny, we'd love an an introduction from you, and then we can jump right into the conversation. Sure. Thank you, Maddie. So I'm Johnny Yu. I'm part of the, infrastructure software research group at IDC. And, software research group at IDC. And, my my main focus is actually on data protection, cyber recovery, and cyber resilience. But you're gonna see how how AI is, a jump to guy AI especially is going to tie into that very, very closely. Well, let's jump into it. So, Johnny, IDC is called, AgenTic AI the next frontier beyond generative AI. Where are enterprises today on their adoption journey of of of, AgenTic AI? Sure, Matt. So recent, IDC research has found that about 35% of organizations, who are at least dabbling in AI as in, like, you know, they're at the very least in the POC stage. Certainly, some of them, some of their AI initiatives might be in production or whatever. But we wanted to make sure that this this survey was specifically capturing the organizations that were experimenting at AI. And we found about 35% of organizations, are working with agentic AI and have at least one agentic AI initiative or workload that's, running in in, in production. And when compared to things like, Gen AI, predictive AI, and machine learning, those percent this this percentage is relatively small, like, 35% compared to, I don't know, like, 70% plus maybe on these other ones. And that really kinda goes to show that, you know, those GenAI, predictive AI, and, you you know, machine learning, these are things that have been around longer. There's a little bit more maturity there. But when it comes to agentic AI, we're still, organizations are still not quite there yet. And this this this this can be a good thing, and we'll get to that later. But I really wanted to kind of point out what the main difference is between those other categories of AI and agentic AI. So, when I'm talking about, you know, those other categories of AI, I'm thinking about things like, you know, your your copilots, your chat DBT, your your perplexity. These these all, what I would say, prompt based AI, where you kinda ask a question and you get an answer. There's kind of a certain level of of, a heavy level, I would say, of human touch where you kinda have to ask the right questions and get the right follow ups. And if you're still not kinda getting the answers that you want out of that AI, you just ask a different question. You know? But agent API is different. It's about autonomy. It's about it's less ask a question, get an answer. It's more assign a goal, and then the AI agent takes the necessary actions to achieve those goals. You know what I mean? Absolutely. So what this kind of introduces now, there's this new kind of challenge where a you have a nonhuman entity taking action and potentially creating nonhuman error. And so when when we think about nonhuman error for, like, those those those other kinds of AI, like, yeah, just oh, I didn't like this answer. Ask it up. You know, re reframe the question a little bit or tell it not to use these sources or whatever, and then you ask it again. When it comes to agent the agent take AI, you don't have that kind of luxury. You kinda your your AI agent took an action and it was a mistake and that's it. You know, you don't you don't go back and go, like, you know, hey, can you please fix that? You know? So this goes back to what I said earlier about why I'm I think it's actually a good thing that we're about at, like, 35% of organizations are at at, like, you know, have one, HEMTIC AI initiative running in in production. It's because it means we have time to build guardrails before adoption reaches actual critical mass. We're at a point where, you know, let's think about all the ways that we can we can stop an AI agent's, mistake from becoming a a a true and total disaster. Let's work on fixing that now before, you know, everyone's running this this stuff in production and then, you know, we're seeing failure left and right. Love it. Yeah. For sure. We're we are early. Right? As an industry, as a whole, we're we're early to the to the show. But, help me understand, like, what are some of the benefits, from early adoption when it comes to AgenTic AI? Like, why even take this step forward? Well, there's a lot to be said about using AI in general and then AgenTic AI in particular. And mostly, it's through optimization and, like, productivity gains and just automating as much as as you can. You know, there's a lot of repetitive tasks that go on in the world of the I of of the IT administrator. And then there are ways to optimize workflows and and so that's where a lot of the benefits come from from, for companies that are, that are experimenting with and then, deploying, AI agents. And me coming from the world of data protection and cyber resilience, I think, you know, maybe a little biased here, but they I'm gonna use it as an example. So especially in the world of data protection and cyber resilience, I would say 95, some even say 99% of the work that goes into successful, a successful cyber response to like a ransomware attack or like even a disaster disaster recovery, some kind of IT outage, 95% of that work comes from preparation of being ready. So, you know, so making sure this workload has the proper resources so that it can meet its RPOs and RTOs and then, oh, this new storage repository has just been discovered. Let's make sure it's properly configured and it's secured. And then this whole running running fail over and recovery test and all that. This is all prep work, but this is all what you need to do to really, like, you know, up your chances of a successful, recovery from a from an outage, whether it's for a cyber reason or just a natural disaster or whatever. It's also repetitive and tedious, and no nobody wants to do it, but they kinda have to. So part of the benefits of using AI agents is that it offloads that type of work onto the agents instead of humans. And so human beings are are better able to take on projects that aren't so much, like, maintaining what's necessary to to to, you know, have that good cyber response and more towards what kind of project can I take out to push IT forward, you know? And they could be something as simple as, hey, that that that that hardware refresh that we've been putting off for a while, I finally have the man bandwidth to do it, you know. To, like, you know, something else that's like there's been an an entirely new initiative altogether. Like, maybe you're going to, switch a different portion of your infrastructure. And again, you just did not have the manpower to do it because so much of your work has been tied up in doing the prep work to cyber resilience. And that's actually a pretty big deal. So much of it is prep work. So much of it is busy work. At the same time, we need it because we want a a reliable recovery when something bad happens. That's kinda, like, one of the big benefits that, we've mostly been seeing in terms of, like, the early adopters. And there is something to be said too about, not only does do you free up, do do you free up, like, more more human bandwidth to take on other projects, but there's something to be said about being able to scale further into more missions, critical workloads. Now that you have systems in place to, like, kind of take care of what I would call, like, some of the low hanging fruits. Like, you know, I've been calling it, like, busy work or the repetitive and tedious stuff. But, honestly, the the end goal here is to try to start making AI get into the more mission critical stuff. That what that does is it does open up a lot of risks, especially when we start talking about AI mistakes and, you know, certainly, we'll get into that. But, before we really jump into that, I also wanna point out in some IDC research, which, we found that AI projects are among the top two top three areas that are least likely to face, budget cuts. So it just kinda goes to show that, organizations aren't pumping the brakes when it comes to AI. And so we really need to, like, just start thinking about these risks before we get to a point where organizations are are deploying a a h agentic AI, like, you know, wholesale at a large scale. That's a good transition, I think, into exactly that point. I think, you know, it seems like a lot of promise. There's a lot of budget being pushed behind this. I think the the benefits are there. We can see them. You just described and articulated those clearly. But I talked to a lot of CISOs and CIOs, and and they're they're concerned. They're up at night, right, thinking about about these issues. What happens if AI goes awry, if an agent goes rogue? Like, what do we do? Can you maybe articulate some of those challenges that you've you've come across through your conversations and research? Sure. So as we kind of talked about with AI agents, there's you're kind of letting an AI agent make some decisions and some of them could be crucial depending on, you know, what keys to the kingdom you've handed to this AI. And so there's there's a certain level of unpredictability to what these AI agents are going to do. And this is kinda true of all AI. I mean, anyone who's played around with chat chat GPT has probably run into some really weird answers. Some of them confidently incorrect. And so to say, like, agentic AI is got not gonna be prone to that is pretty naive. So there's kind of some expectation that while while most of the time it will probably make the right the make the right call, there's gonna be situations where it isn't. And you just wanna be make sure that you're prepared for that. So one of the things as a CISO you should be really, really concerned about and, you know, some of the ones I've spoken to and why they've not really let their, Agentic AI touch some of the more sensitive systems is because you need kind of like an audit trail or some way to kind of figure out, okay, why did this AI agent decide that deleting all these backups was the right call? Where exactly where did things go wrong? How do you pinpoint that point of failure? And how do you fix it, you know? It's kinda similar to there gonna be guardrails in play in place from stopping a human from doing those, you know. There's multi factor authentication, all kinds of things to make sure that you have something in place that no human can just suddenly go into your system and make a big decision like that. Once again, I'm using deleting your pack ups because I come from a way of data protection. But literally, insert other large major major infrastructure change here, and you kinda get what my point is what what I'm getting at here. And then there's also something we might not have to look out for at this stage, but, you know, I can guarantee you it'll happen further on in the future, is at some point in the future, there will be exploitation. There will be manipulating. There will be some some trickery that that, bad actors will use to to force AI agents to make some really bad mistakes on purpose. There are gonna be ways to exploit this. It just hasn't happened yet because, you know, we're not at a point of mass mass adoption. But I can guarantee you when we are, bad guys are gonna be figuring how phase two extort them. And, you know, I think it would be naive to not, like, you know, start planning around it or thinking, like, oh, that's gonna be so far in the future that it's it we don't have to worry about it now. It's like, well, actually, now's probably the perfect time to start thinking about it because, I mean, that's just kinda how it is with with cyber criminals is they're always gonna be jumping ahead and looking at the next thing that they can exploit. And then one last kind of, like, I don't wanna say it's an anecdote because it's happened often enough then. We have a IEC research to to back it up. Historically, there has been a trend of organizations adopting new technology too quickly and then having to kinda walk back their investments. And so, I have data from an IDC study about four years ago, about containers where about 50% of the respondents said that they had to repatriate 50% or more of their workloads that they had containerized, because they had realized after they had, you know, gone through all the trouble of, refactoring it to work in containers that, all of their data protection and security, you know, connectors, everything that everything that they had previously been using to make sure that the data has is properly backed up and can be recovered into these newly factored, workloads. And all the security features around it no longer work, and so they had to repatriate all that. And it's all just because it's like, well, they were told that, you know, containers are great. It's the hot new technology and, you know, you're gonna get so many so many benefits and then they were just moving too fast and not letting these other adjacent systems that are very deeply connected and, you know, these safety nets, basically. Like, they weren't letting that stuff catch up or understanding what the consequences would be when they when they make those switch over. And so, you know, I think it would be naive to think that that's not gonna happen within Agentic AI, but I think it is, you know, our job as people in the industry and as technologists to, like, you know, minimize that as best as we can. Absolutely. Yeah. And I think those organizations that went through that repatriation process would love some resilience tools at their disposal, right, to help get them back on track, which is a great segue into, you know, how should we start thinking about this when we think about agentic AI? How should we start thinking about resilience within within that context? Yeah. So when we when we think about resilience in a general term, it's more about making sure that no matter what happens that we can get back up and running pretty quickly. And for the most part, I I would say the majority of things that would, you know, jeopardize resilience are probably, like, human mistakes and then malicious, malicious attacks. And now we have this new factor to come in that that's coming in, which is the non human mistakes. So now we have something that's adding to that resil that resilience story is not only does resilience must encompass, okay, somebody, enter or, like, you know, entered a command wrong or, you know, as I keep saying, like, you know, deleted the wrong repository or configured something incorrectly. Like, you know, human mistakes that are kind of expected and within the realm of reason. We also have to obviously protect against the malicious actors, the cyber attackers, the ransomware, you know, all that all that fun stuff. And now one new factor that can really jeopardize resilience within an organization, which is now we have these AI agents that might make a mistake, that might cause downtime, might make it very important, application that your business depends on and, then make it, you know, suddenly not not able to work for, you know, an extended period of time. That's something that could affect your resilience. Kind of similar to how we have ways to address the common unintentional mistake, we have our ransomware, you know, playbooks on what to do. We have playbooks on what to do if there's a disaster, if there's a fire or, you know, a huge power outage or something like that. We need a similar kind of playbook for AI mistakes as well. So, you know, how do you pinpoint where the mistake happened? You know, how do you get the context or determine what were the circumstances that made that AI make a bad decision. Because, you know, usually, there's some kind of logic to it you you would hope that led to the AI making what turned out to be a bad decision. But, you know, if you can you wanna be able to at least have some system in in place to figure out those circumstances that led to the AI making that mistake. And I feel like that's kinda half of the equation, you know. You figured out where it went wrong. Great. But, like, how do you quickly get back up and run running? Or, you know, how do you undo that mistake? How do you revert everything back before that mistake was made? Because, you know, we now realize it's a mistake and we need to undo all of it and go back to where we were so that, you know, we're back to that resilience, you know, of when something bad happens, how quickly can you get back to full functionality, you know. And so I think without the ability to kind of, like, undo the mistake after you've identified it, then there's no way to really deploy AI agents safely. Because if a mistake made by an AI agent can, you know, grind your business to a halt, then, like, who's gonna take that risk? You know, who's going to how do you confidently deploy AI agents in a scenario like that. Right? Absolutely. Yeah. I think, the slide might have given it away, but Rubrik, we've been working on a solution for for this exact problem called the agent rewind. Johnny, you've had a first look at it. I'd love to understand, like, what stood out to you, on this capability? Yeah. So I think that addresses a lot of the, a lot of the points that, we were kind of making earlier about, what happens when, when an agent an AI agent makes a mistake. And so, this, this kind of, integration that's going on here with agent rewind and like, with with with Predabase and and, Rubik's Security Cloud is it means that from the Predabase side, there is a way to to trace what the the AI agents were doing, you know, like, what prompt was it following, what tools did it have access to, what action did it take on those tools to, like, kinda like, what it what it brings up here is accountability for what the a AI agent is doing. You now know, okay. This is what it was working on, and I have that visibility. That's great. And then the Rubik, half of it, the Rubik Security Cloud is bringing in the ability to do that that disaster recovery portion or that that that, you know, reverting things quickly back to an earlier state. And so I think, like, the combination of that was what really stood out to me was it's a matter of you have Credit Base kind of supplying all that context that you need so that you see you see everything and you see where the mistakes were, and that can be used to guide the recovery portion of it, which is the Rubik Security Cloud coming in. Now it is a safety net for AI agents actions, And you can't really confidently deploy AI agents if you don't have a way to to fix their mistakes. You're just stuck being afflicted by by by downtime, and that's obviously what you wanna avoid here. Absolutely. No. That's good. Thanks for that take, and we appreciate it. And maybe some natural next question for for enterprises that are listening in today. You know, where where they gonna get the most value? Where where is agent rewind the most relevant for them today? What would be your what would be your answer to that? Well, I think there's certainly that category of organizations that are very far along and, you know, have one or more AI agents running in production, then they're definitely gonna get a lot of value out of this for sure. But even if you as an organization aren't quite at that stage yet, I think that this is something that's a good foundational, capability to have. Because, presumably, if you're going to be using AI agents either, you know, you're in the p POC stage now and, you know, maybe you're you're running it in some not so critical applications, but, presumably, you're going to expand. You're gonna go beyond this point. And so as I keep saying, it's better to, like, have those guardrails in place and, like, not make the mistakes later after the AI AI agent has made a mistake and you realize you don't have a way to really fix that. So as long as you have some level of aspiration to have your AI agents start, you know, touching some of the more critical applications, then I think that's where you're gonna find the most value, in in, in agent rewind. Obviously, the I would say the usefulness of AI rewind in general, would, like, grow in proportion to how costly an agent an AI agent's mistake is gonna be. So, like, again, if you're mostly using AI agents in some very noncritical stuff where, an agent making a mistake is like, okay. Whatever. We'll just, you know, we'll just fix it the next cycle or something. Like, you know, it's it's not that big a deal. Then sure. You don't really need agent rewind. Presumably, you will eventually develop, Agentic AI to the point where it will be touching more sensitive stuff. So then your value the the agent for agent rewind is just gonna grow, in proportion to that. Makes sense. So we went over a lot so far. From your perspective, what's, like, the biggest takeaway for for folks listening in today? Well, I would say that, IDC data certainly points out agentic AI is is, is not stopping. It's definitely growing. AI adoption in general is growing. And we definitely know that there's a lot of benefits to be reaped in deploying age, AI agents. But, honestly, the undo button or at least some kind of way to ensure that agentic AI does not lead to, you know, downtime because of some major mistake. Like, I think that's pretty foundational, pretty nonnegotiable. Like, if you if you don't have that, then you're running a lot of risk in in letting AI agents, touch anything too sensitive. So I think that's one, like, one of the major major themes there. And I think for organizations that have been, kind of on the opposite of of, like, you know, same point, but coming at it from the opposite perspective where they would love to use, to have their AI agents do more, but have been held back because they're afraid to hand over the keys to the kingdom on the more critical applications. I think the other takeaway from this is, well, Agent Rewind does a good job of removing that barrier because suddenly it's like, oh, well, there's here's that safety net, of you've been wanting to take your AI agents to the next level. You wanna give them more autonomy over more critical stuff, but you're afraid of what they might do. Here's a way to fix those mistakes in case, you know, in case they do make those mistakes because, like, you know, no AI is gonna be perfect. There's going to be, some situation down the line where you're going to be, you know, you're gonna need something like an agent rewind or something that can that can, you know, fix those mistakes. That's great, Johnny. Well, we appreciate you, setting the stage for us. I've enjoyed this conversation. We've talked about a number of things, but most I I think that stuck up for me is, you know, we've heard about the resilience, the need for it when it comes to AgenTic AI, to start developing with competencies, to start to scale it out, to get those business benefits, from AgenTic AI. And so what I'd like to do now is turn it over to Sohil, who'll, take us a little deeper into the capabilities, how it actually works from an agentic rewind perspective, and even jump into a demo. So, Johnny, thanks so much. So he'll pass it over to you, my friend. Great. Thanks, guys. Thanks, Matty. Thanks, Johnny. Really good insights, really good context. So I'm Sohil Sheth, a go to market tech lead here at Rubrik. So I'm excited to to spend a little bit of time taking you through how Rubrik is specifically thinking about these challenges, these opportunities that Johnny and and Matt laid out, and then we'll go through, a quick interactive tour of agent rewind. So if you look at Rubrik's journey in this space, you got a a pretty good idea from from Johnny. But, you know, Rubrik's journey began focusing on the fundamentals of enterprise resilience. Right? So backup and recovery, Providing businesses critical ability to recover from a wide range of operational issues. So whether it was simple data loss, catastrophic events, ensuring business continuity. As the threat landscape evolved, Rubrik did as well. Right? So we expanded our focus to address the growing challenge of cyberattacks. So that includes developing sophisticated tools to help organizations not only defend against threats like ransomware, but also to rapidly recover from them. So, really, it was about a new standard of cyber resilience. So if you look at, you know, fast forwarding to today, standing at this next frontier of of enterprise resilience, we're addressing this new and and significant risk. Right? Mistakes made by AI agents. So to meet this challenge, we're introducing agent rewind. So if you look across the agent landscape, companies of all types are are building modern frameworks that are powerful and really easy to use. So we've seen the incredible way that AI and agents have been opened really to a huge nontechnical audience. So if you look at companies like Salesforce, right, they're staking really their entire platform and vision on this potential. And we do see some enterprises starting to realize that potential. And, you know, day to day, I'm working with lots of these companies who are exploring, platforms like Salesforce's Agent Force. And when you look at these platforms, especially when you're factoring in things like MCP, right, the the ability and and they make it easier than ever for agents to to really take these high impact actions across the business. So really what it means is that the blast radius, scope of impact, they've really gone up exponentially, but the audit, the tracing, the rollback capabilities haven't really kept. And then if you think of the scale at which all of this is happening, that's also exponentially gone up. So if you consider more of the the traditional nonhuman identities and and human error, It was more predictable. Right? You can only do so much damage if something went wrong. But now we're seeing ten, twenty, a 100 x the number of AI agents compared to the, nonhuman and human counterparts. So these agents are essentially taking all of the best and worst of the other identities here. So the agents are autonomous. They're taking rapid action at scale. But really, they can be unpredictable and unsupervised. So it requires a a completely new way of thinking about this challenge. So you end up with things going wrong in unpredictable ways that just doing more testing is never going to fully remediate. AI agents are incredibly powerful, but they can't be trusted, and and really that's by design. I like the saying that, you know, the the fact that they can't be trusted is actually their crowning feature and bug. And we've already seen real world incidents which show these really aren't just hallucinations. Right? They're executed and, in a way, expected actions. So if we look at this notion of AI resilience with Rubrik's unique position across enterprise, on prem, cloud, and SaaS workloads, we're able to solve a lot of the challenges that enterprises are facing when you combine autonomous AI agents with critical applications. And we do this in a single pane of glass approach, which really gives you that central control plane of answers. And what do you need answers to? Right? There's a whole host of critical business questions that have to be answered that you need to understand the implications before getting started. So if we look at agent rewind, what exactly is it? So at the highest level, it helps enterprises adopt Adjentic AI with confidence, and it does that by delivering two things. Number one, visibility. So you can trace agent actions across apps and infrastructure. And then secondly is this notion of reversibility. So undoing data changes, configuration updates caused by those actions. So it's not just observability. It's resilience. And it's important that enterprises have the confidence to go ahead and deploy agents like Copilot Studio, Agent Force, AWS, Bedrock, realize the full value of all of these advances, but with that confidence that they can roll them back when they go off script. So I wanna do a quick walk through of, exactly what this looks like. So here we see, our home dashboard within agent rewind. And if you look at the agent environment Voyager, for example, right off the bat, you're getting an inventory of active agents. Even just understanding what users across the business are using is that critical first step. And then if you think about the top applications in in the top right here, this is really showing you, when you think about actions, tool calls, functions, whatever you want to call them, that's the true value of these agents. Right? Autonomous reasoning in order to take what they deem is the right action. So we get insight into what applications are actually being impacted by the agents, and then we can break down, the different types of risk. So when you talk about assessing risk, we have a a highest risk widget as you see here across all of the agents, and there's different types of risks that are important to assess. Right? So if you think about access risk, so what is the level and scope of access or permissions granted to the agents? Does the agent have access to sensitive data sources? And then you can think about, you know, capability risk. Is the agent enabled to perform destructive actions? Can it delete? Can it modify? Can it override? Is it enabled to perform destructive actions on sensitive data? And then just general activity risk. Is the agent active? Is it performing daily or hourly high impact actions on critical apps or critical data? Is it performing these actions on non backed up data? And then finally, if you just think about rewind history risk, which is a a new concept rubric is introduced where we can look at whether the agent has repeatedly executed these dangerous actions that actually require frequent rewinds, by rubric or potentially even by by other systems. So if you look at some of these risk levels, Rubik's already for years been, you know, thinking about secure data backups and being able to do things like analyzing sensitive data or posture management. So really thinking about how the existing knowledge within our secure immutable backups can drive the evaluation of risk levels for agents that will be accessing all of these same data stores. So let's click into one of our agents here, our, Salesforce, sync agent here. And so the the agent map that you see here is really where you can start visually graphing an agent. And here we see that our agent and the underlying applications it has accessible with, quick indicators of underlying risk of the actions that are targeting these applications. So we're gonna click in and and see some pre flagged high impact actions. Again, Rubrik has, you know, some unique context here about Salesforce from just backing up every piece of data and metadata within these Salesforce instances and running it through our sensitive data detection engine, and other modules. So we click into this first action. Let's look, at this overview. So you can see now as we drill in, the overview has taken all of the traces, all of the raw logs, and broken it down in into natural language. So we get a summary of the action, for more intelligence around really explaining why this action was taken. And then you can think about, and look at the rewind plan here, and that leverages the context of the logs and the summary to to formulate this plan to, either manually or in an automated way, just undo, the high impact action here. And so in this example, we see there was a configuration change done on a important field in Salesforce. So we can click, rewind if that was unintended. And for a lot of our customers already leveraging something like Rubrik's Salesforce protection, you'll have access to every snapshot. You can even do things like, you know, take on demand snapshots, right, before and after the critical event. And this is where the real value starts coming in from combining some of the unique observability and intelligence around agent actions with, the existing resilience capabilities of of Rubrik. So I'll end it here. Just want to, add in here that if you're interested in seeing, more of a full end to end demo and and even learning more, we will be, at Dreamforce, starting in October 14, and we'll be diving in, deeper into some of the specific things we've seen, across the Salesforce ecosystem and, really beyond. So if you're planning about going, definitely reach out. We'll we'll be there. And, with that, Maddie, I'll, toss it over to you. I think we we have some questions. Yeah. So, Hill, that was great. Thanks so much for going through through that in detail. It's exciting stuff. I think helps really establish Rubrik as a, security and a an AI operations provider. So, excited for more things to come, from the team. I know we're we're cooking up even more. Two things for you guys that, come up. First, I think probably is for you, Johnny, and maybe, Silva, you can add on to that. But it was really around how agent rewind fits into a a larger resilient story and strategy. I don't know, Johnny, would you like to maybe take that question? Sure thing, Matt. So I think, agent rewind fits in as another tool in the IT team's arsenal when it comes to, like, maintaining resilience. So especially from, like, a CIO CSO stand point where maintaining resilience and maintaining uptime is, you know, of utmost importance, then this is just another tool for that specifically for agentic AI. And this kinda goes back to what we were talking about earlier, like, what were some of the, what what was the some of the things that were keeping CIOs and CSOs up at night is, like, you know, the potentially the unpredictability of of AI agents of, like, you know, making mistakes about not being able to chase those mistakes and all that. And, like, this is this is where this fits in is this is the tool for that accountability. It is giving that context there and then kind of, like, the whole like, what this all generally points back to at the end of the day is resilience and ensuring that any sort of mistakes, specifically in this case, obviously, we're talking about AI agents. But, like, what an organization is is more worried about on a more general umbrella term is any sort of mistake that could cause an outage, you want to be able to recover from. And you have your tools for if it was a human mistake, you have your tools for when it was a bad actor. And now here, agent rewind, here's your tool for maintaining resilience when it comes to, to agent decay. And I would I would just add, I touched on it a little bit in in the demo, but it really is a tool that's meant to fit alongside. Right? This isn't kind of a a separate silo tool. So the more context it has, right, to existing data, to the existing backups, as you saw in the demo, it's meant to hook into really all of the resilience that Rubrik's already doing or sensitive data detection, analyzing risk impact. So I think it's meant to it's really built to fit alongside existing controls. Great responses, guys. Appreciate it. Johnny, one more for you. A little bit about IDC. Where is the research pointing from an IDC perspective over the next twelve to eighteen months? Well, we'll be following, AI adoption in general. So and publishing more on on that. But, specifically, I think I will be looking closely at, how many organizations will end up scaling back their agentic AI deployments because of an costly mistake made by AI agents. I know it sounds a little, I have a feeling that that is something that will happen just just simply because it's a new technology that's gonna happen. And so give it some time, and I'm sure we're going to be able to get some numbers on that because I think it's entirely naive to think that there'll be people who are that there aren't gonna be organizations that will, you know, follow all the best practices and put the guardrails up in place. Like, historically, that hasn't been the case. I don't see it happening here, for a majority of cases, but I do think it is definitely worth it for us to try to mitigate that as much as we can. So that's certainly one of the things that IDC will be looking at. And I think I've hinted at this enough. Perhaps not in the next six months, maybe in a year, maybe a little further down the line. But once we start seeing how bad actors are gonna start exploiting agentic AI, we're gonna start looking at those numbers as well and seeing, like, you know, how are they doing it? You know, what kind of exploits are are common? What are the kinds of like, you know, what is the attack methodology or something like that, you know. So I think we're gonna start seeing that more, and I think, certainly, IDC research will start moving in that direction too. On top of, you know, tracking general adoption of AI because, like, those things are kinda linked. As this stuff gets used more and more, that means there's more money making opportunity for the bad guys, which means we start looking into that. And, you know, like, this this this this cycle of new technology, new people adopting it, reaching a critical mass, and then bad boy bad guys exploiting it. We've seen it before. It's going to happen again. So count on it. Well, excited to to to learn more from you and the team at IDC. And, with that folks, let's let's wrap it up. I think, Johnny, your insights were invaluable. Thank you so much. Thank you to the IDC team and and the partnership there. Sohail, thanks for showing us agent rewind, how you can, with agent rewind, you can deliver resilience, in practice. And then and then most importantly, thanks to the folks who joined, the discussion today. We appreciate it. I guess if there's one thing we can kind of take away from all this is that resilience is really the foundation of AI adoption at scale. Without it, you run the risks that Johnny so eloquently articulated earlier today. And agent rewind is really the one of the answers to that, that will provide confidence as you, let AI agents innovate without really putting your business at risk. So that's what we're all about, and and, excited to take the next step with those folks who are on the call. If there's any interest, please feel free to reach out. And again, thanks for making the time. Have a great day. Great. Thank you, everyone.