After Hugging Face: 3½ Ways Frontier AI Is Making Cloud Recovery Strategies More Critical

  • Rubrik
After Hugging Face: 3½ Ways Frontier AI Is Making Cloud Recovery Strategies More Critical

About This Webinar

Recently, an AI agent broke out of an isolated environment, escalated its own privileges, and found its way into production infrastructure. It took an attack path no human had thought to model. Security teams caught it and shut it down that time. The next one might move faster.

Hugging Face isn't a one-off. It's a preview. Frontier AI is already changing what runs in your cloud, and most recovery plans still assume incidents unfold at human speed, with a person in the loop to catch them. AI doesn't work that way.

No product tour. Just a real conversation and a few frameworks you can take straight into your next architecture review, tabletop, or board update.

Prevention and controls will fail sometimes. When they do, recovery is what's left. Better to pressure-test it on your own terms than during an incident.

3½ ways frontier AI is raising the stakes on cloud recovery:

  1. Change outpaces review. AI writes production code in most organizations now. That means more changes than anyone can check, and more bad ones slipping through.
  2. Failure moves at machine speed. A misconfiguration, corrupted dataset, or compromised credential spreads in minutes, not hours. Recovery speed becomes the whole game.
  3. The blast radius is bigger and blurrier. Agents and non-human identities sprawl across clouds and SaaS, so damage crosses services instead of sitting in one datastore. The hard part isn't whether the data exists — it's finding a clean state and re-assembling the business service.
  4. The ½: autonomous agents that modify infrastructure, data, and permissions on their own. Hugging Face proved the capability is real. Most companies aren't running it in production yet, but it's coming fast, and almost no recovery plan accounts for an agent that goes wrong.

Speakers

  1. Cal Al-Dhubaib Principal Technologist Rubrik
  2. Matt Castriotta Field CTO, Cloud Rubrik
Full Transcript Mackenzie Putici Hello, everyone, and welcome to today's special webinar event. After hugging face, three and a half ways Frontier AI is making cloud recovery strategies more critical. Today's webinar is sponsored by Rubrik and produced by Actual Tech Media. My name is Mackenzie Pudice with Actual Tech Media, a future b two b company, and I'm excited to be your host for today's session. Now I'm sure you've all heard by now that recently an AI agent broke out of an isolated environment, escalated its own privileges, and found its way into production infrastructure. This may be like one of those famous law cases we always refer back to. And, well, it took an attack path no human had thought to model. Security teams caught it and shut it down this time. However, the next one might move faster. And hugging face… Sorry. Hugging face isn't a one off. It is a preview. Frontier AI is already changing what's running your cloud and most recovery plans still assume incidents unfold at human speed with a person in the loop to catch them. And AI doesn't work that way. We know today's webinar, there's no product tour, just real conversation and a few frameworks you can take straight into your next architecture review, tabletop, or board update. And, well, let's be frank, prevention and controls will fail sometimes. When they do, recovery is what's left. Better to pressure test it on your own terms than during an incident. So as you can tell, this is gonna be a great conversation today as we dig into everything hugging face and how you can respond. But before we get our speakers up here on stage, I just wanna go through a little bit of housekeeping, and then we'll get that show on the road. So I've already seen a lot of you introducing yourselves in the chat tab, but if you're just joining us, just getting in the door here, if you look to the right hand side of your console, you'll see a chat interface, and there's a few tabs at the top. Chat, docs, q and a. So let's quickly go through those. And first things first, if you haven't already, feel free to introduce yourself. Give us a quick hi, hello, hola, bonjour, aloha. Shoot a greeting in there. Let us know where you're tuning in from. Love to see the energy flying around here. And, of course, think of this as your sounding board throughout the session. If you have interesting feedback, comments, ideas, anything you wanna leave with the speakers, let us know. We'd love to see you're responding to the information. Alright. We've got Paul coming in here from The US East Coast. We've got Ryan from San Fran. We've got Rob from San Diego. We've got Colorado Springs, Nebraska, Bethlehem, Porter Ranch, New Jersey, Atlanta, Jackson, Mississippi, Texas. Awesome. So great, Brad. Great to see you guys all popping in here, introducing yourselves, giving a quick hi hello. This is gonna be a great session. So as I mentioned, fire off there in the chat, anything and everything. We love to know what you're thinking and feeling. Also, let us know if you have any technical issues in the chat. Hopefully, that's not the case. And a quick browser refresh usually does help if there's any issues with video, audio, slides not loading, not syncing up. But if you do have persistent issues, reach out there. Myself and Riley, who's coordinating our webinar here smoothly in the background, we'll be here to help. A couple other things you can do to enhance your viewing experience, bottom right of your viewport, if you hover over it, you'll see closed captions. You can enable those. If you're on a mobile device, there's a light mode you can toggle on to sort of simplify the console, helps on a smaller screen. There's also a new catch up feature here in the console, with a little arrow. So if you guys get interrupted, have to step away, if you're just joining, coming in late at any point, hit that, and it'll give you a quick ten minutes summary of what just happened or a summary of the last ten minutes that took place. Alright. With that, couple of other things. If you have questions today, please throw them in the q and a tab so we can keep an eye on those, and they don't get lost in the chat. And then in the docs there, we have got just a link to rubric.com. So if you wanna learn more, get deeper, go ahead. But as mentioned, this is not sort of a product focused demo sort of day. This is more real talk, more just, hey. This is where we're at with the the new state of things. So, but you check it out there. And last bit of fun, we have a $250 Amazon gift card up for grabs. We're gonna draw that at the end of today's session. You do need to be live and present on this webinar to qualify for that prize. So, feel free to stick around. And at the end, we will award that to one lucky winner. We've got a thumbs up. Yes. There's also a reaction section that I forgot to mention. So if you guys love anything, wanna give some applause, give some hype, feel free to do that throughout the session. Alright, everyone. With that, I think we have about done it. So we have two presenters today. Excuse me. We have Kyle Al Duhaib, and he is the principal technologist over at Rubrik, and then Matt Castriada, who's the field CTO of of cloud at Rubrik. And, basically, I'm gonna hand it over to you, Matt, set the stage, and we're gonna dig into it. Take it away. Video Welcome everybody to, three and a half ways Frontier AI is making cloud recovery strategies more critical. It is, becoming top of mind, these Frontier models like, Anthropics, Mythos, Opus, Fable, as well as, OpenAI's GPT five dot six family are really, sort of changing the game for cybersecurity. We like to say that, that these frontier models have moved the cheese for security. And, what we're finding, obviously, is that these models have very innovative ways of finding their way, out of sandboxes and into other companies' environments. And, and as as evidenced by the hugging face, incident just recently with OpenAI's models. These these frontier models take a path that no human has thought to model, and security teams are responding against a machine attack with with the human response. And that's not gonna work in today's day and age. So what we're gonna talk about is ways that, you can ensure that you have a recovery posture for your most critical data regardless of where that sits. We're gonna focus a lot on cloud today, but Cal's gonna give you some, some some tips and tricks, across your data state on ways that you can do that. You… We'll get into the three and a half here in a bit. Just by way of introduction, my name is Matt Castriotta. I'm Rubik's field CTO specific to our cloud go to market, and with me is Cal Dubaib, Al Dubaib. Cal, can you introduce yourself? Yeah. Absolutely. Matt, really great to be here today with you. I'm principal technologist here in Rubrik. Got a deep background, been in the AI and data science space now for fifteen years. And I focused much of my career on the intersection of AI risk management and getting these AI systems into production. So so really excited for this exchange of knowledge here, Matt. Awesome. Appreciate it. Why don't we just jump right into the content? What are these three and a half ways, Cal, and and how can companies prepare themselves? Awesome. Awesome. Well, there's been no shortage of examples, with breaches over the past few years. They're both increasing in frequency and increasing in their consequence. But the reason why it's becoming a topic yet again today, and just with recent news of what happened with Hugging Face is we've seen the very first example of a prominent, frontier model actually executing an autonomous task. And the one thing that I I wanna highlight here that I'm actually having more conversations about is in the case of Hugging Face, it really was just an agent going after data and data exfiltration. The next flavor of this is a malicious actor directing an AI system to then manipulate and fool your own AI agents into overreaching. And part of the reason why, this is becoming hard to manage is in the past, we we had traditional software that had clear failure cases. It either worked or didn't work. You would stress test it, you would unit test it, and you would kind of know the bounds, and it would be very clear if a failure occurred. With the introduction of narrow machine learning and AI models, the types that would predict a future state, apply labels to items, model data, for example, you had some small tail chance of these models failing. And there were ways that you could build observe ability to kind of detect when these models were dealing with atypical datasets. It got a little bit harder, but we still had a good handle on it. Today with AgenTek systems and the enterprise, they can operate over multiple steps with any combination of data sources and applications and in some cases, have the ability to take actions to change, configurations in your environment. And so the surface area is greatly expanding, and it's virtually impossible to test every possible edge case in advance. And if you are like 80% of our customers, chances are, you know, you have your your data on one of the, big cloud providers, whether it's AWS, Azure, or Google, and you're more likely adopting more than one of these. Now most large organizations actually have on average about 89 different SaaS applications in addition to, some of the largest organizations have upwards of 200. And what this results in is basically a, proliferation of data that's increasingly harder to get a pulse on, that may or may not include sensitive information in data sources that could be unstructured or harder to monitor. And now you have AI agents that are able to crawl this data exhaustively. So unlike in the past where a human user wouldn't use all the privileges available to them, these AI systems, they're relentless. They will uncover and find sensitive data, pull it into their context, and may actually make configuration changes in your environment that you are not aware of. Now on top of that, the reality is that most organizations have dozens, if not hundreds of these accounts across their environment, probably segregated by region, each one treated as a distinct management point. And so the, the scale of what you need to monitor and what you need to be able to react to not only is moving at machine speed, but is far beyond the scope of what humans are able to keep up with anymore. Now, what happens when one small component here, Cascades into a failure or an AI system is acting off of inaccurate information, and then makes a decision that then has an impact on one small component of your environment that results in a failure. Earlier this year, Amazon experienced one such incident with code that was AI generated that led to, about 6,300,000 orders getting dropped. So even the big names are not immune from this. And the reality is we're dealing with a… The… A great sprawl of these AI systems that are cascading across your tools and data sources and multiple different users. And many of my own conversations with CIOs and and CISOs, the reality is that they often end up uncovering that there are more agents and AI tools operating within their environment than they're even aware of. And having a good sense of what agents are operating and touching what systems of record and taking what actions is surprisingly hard. Fewer than 23% of organizations actually have full visibility into the AI systems operating within their environment today. These AI systems also misbehave in very strange ways. Just one small example here that I wanted to highlight is the idea of prompt injection. And this could be something as innocuous as an agent that you're tasking with search that comes across a website. And that website has now been compromised and hidden within it is some HTML injection that instructs the agent to exfiltrate information or misbehave in an unexpected way. Now, something that we often stress at Rubrik is that just because you have backup turned on doesn't mean that you're well prepared to recover from an incident like this. Most of the downtime spent, that leads to weeks between a critical breach and recovery is actually going through and sifting through every possible snapshot to determine a recovery point that you can restore to that isn't gonna also adversely impact your environment. So I know that was a rapid fire overview of some of the challenges here, but putting that all together, the three and a half reasons why AI is making cloud recovery a lot more challenging is, one, the blast radius is getting a little bit blurrier and a little bit larger. The pace of changes happening with your environment far outpace this… The capability of humans to be able to manually review. Three, these failures are now moving and cascading at machine speed, And the AI systems operating with your environment already have a tendency to overreach. And very soon, they will be the targets of attacks by threat actors leveraging Frontier AI against you. So with that being said, let's, let's bring Matt back into the conversation here and talk a little bit more about what you can do about that. Yeah. I mean, I I think it's a great overview and, obviously, very, very relevant today. I have, am in front of customers constantly that are either at the beginning of their cloud journey, the middle, or the end of their cloud journey. And and, what we're finding in the cloud, specifically as these, as as vibe coding becomes more popular and code agents become more popular, I think the approximation is about 90% of organizations today are writing code through through some sort of agentic, agentic tool, is that's, that really puts a spotlight on config drift. And we see config drift constantly within within the customer's cloud estate, especially mature cloud customers. And that config drift can manifest itself in lots of different ways. It could be vulnerabilities that these agents are obviously able to chain together to do nasty things. In the case of Hugging Face actually hacking into another company's database or it could just be, you know, an an AWS key that happens to be publicly accessible and, an agent gets a hold of it and that's how they in auth… Authorize into an environment, authenticate into an environment. So so again, code agents, definitely in use, extensively. And and really, the the challenge is is that these agents are designed to be autonomous. They're designed to make changes autonomously, but only a third of our customers actually trust agents to make autonomous changes. So the human in the loop is still is still a a factor here. And, it's really what we feel like is dampening, is dampening the, you know, the ability for agents, Agentic, to really take off because, because of that mistrust in in what an agent can do. So it's not about a human in the loop. It's about a machine responding to a machine action. Right? And I think that that's really what we're going going for here, and and that's and that's what I'm stressing to to our customers that are… Have very distributed data states. Well, these these agents, they're relentless. I've been spending a lot of time over the past week just unpacking and digging into the OpenAI incident. And Crazy. It's been really fascinating. I mean, Entropic wasn't the only organization that was hacked in this process. There is, another service running on, Modal. And, apparently, the OpenAI model was trying to look for an environment where it could stage its attack, and it exhaustively searched for every publicly accessible endpoint. And this was actually… This environment of Modal was improperly configured, and it invited essentially the AI in. So these vulnerabilities, they exist. They're already in the environment. We just never used to come across them with the density that these AI models can't because, again, they're exhausted. They're relentless. Yeah. Yeah. And that… And again, and that that that… I think that's what… That really puts the spotlight on, on ensuring as you use infrastructure as code, because, again, a lot of our customers tend to… Are are tending to start to standardize on some sort of infrastructure as code platform, whether that's Terraform or Ansible or whatever they're using to instantiate infrastructure and configuration, that they have to do this in a very meticulous and thoughtful way and then keep track of configuration drift moving forward. Because that configuration drift is really what these agents can exploit. Right? And as they find how that configuration drifts and they find ways to essentially extend their reach, they're gonna exploit them, and they're gonna do it at machine speed. And that's the challenge that that we're faced with is it's… We're seeing exploits that a human wouldn't see necessarily through an attack chain, and then it's moving at a 100 x the speed of a human. Right? So then the the the response, detection and response essentially becomes mood at that point. At that point, now you have to you have to recover or limit some sort of, damage that, you know, that that autonomous agent might be doing. Yeah. Now now you talk to a ton of chief information officers and and CISOs. When it when it comes to actually dealing with some of the the the preparation for this, what are some of the practical challenges and constraints that they're running into with their cloud environments? Yeah. I think the first thing is is is as I mentioned that time to response, the mean time to response. The, you know, it was already measured… Breakout times, CrowdStrike has already… Had already before the agentic agent already measured that in seconds, meaning the amount of time it takes an attacker once they gain a foothold. And we say attacker loosely here. An agent does… Isn't always have malicious intent. Right? But but, you know, work with me here because we're we're we're going down the sirens. Sometimes, you get it. But… Yeah. Exactly. So so, you know, that that mean time to response had already needed to be measured in seconds because the breakout times were somewhere in the twenty four, twenty seven second range, meaning the amount of time it took an attacker to be able to extend their reach beyond the system that they initially gained a foothold into. Now that breakout time is measured in milliseconds. Right? Because, again, agents can chain exploits. They can they can authenticate across the environment, move laterally, access databases on a… Unauthorized or actually authorized, but on a… One agent, anthropic. One agent. 18,000 actions. Yeah. Right. Yeah. And in how much time. Right? Yeah. I mean, it was probably measured in seconds. So that's the challenge that we're running into. And I think that's the biggest concern right now of of any… All of the c suite that I speak to. And so what we keep making clear to our customers is until the response side of this catches up, because it's gonna take time for the response side to this… For this to catch up. As a matter of fact, that model incident you mentioned Yeah. A frontier… A Chinese frontier model was used to stop the the OpenAI frontier model from doing additional damage. So so it's actually now a a a gain of one month. It was a control. Right. We used another model as a control against that model, which is the future of cybersecurity. That is what we're gonna need to do. But across the board, we're not there yet. So until we get there, it becomes all about recovery. How do I ensure that when my environment is impacted, whether that's if that's a an unauthorized read from a database or whether that's a change to an s three bucket that happens to be, critical to my organization, how quickly can I get the business back? What is my mean time to recovery? And I think that that's that's sort of where, where the, you know, where where it's moving more is more towards the end of the kill chain, the ability to be able to get back and how quickly we can do that. And and what we're asking our customers to do is the same way that they do quarterly Doctor exercises, they should be doing mass recovery exercises more. Yeah. Ensuring that they have a place to recover too. They have an isolated recovery environment. These are all things that are becoming top of mind as as as recovery becomes more more more prevalent and and relevant. I mean, that tracks with my experience working on the AI side of things. It it became very clear that you needed to have some observability in place to understand how the AI system was operating and when it was encountering edge cases or weird things, and, of course, guardrails, table stakes now. But whenever I would ask people, okay, and how often have you actually tested these? Kinda… You get you get crickets on the other side of it. And so like I can't I can't underline that enough. In a typical environment today, can you talk to us a little bit about what happens when prevention and controls ultimately flare up, for an organization that might not be as mature in their resilience and recovery journey, what does that typical timeline look like? Yeah. Unfortunately, what it does is it manifests itself as a course of discordant data points. And I think that that is… That's the problem that we have is is these data points in piecing that puzzle together right now or dependent on a SOC analyst or somebody that's looking at a SIEM or some sort of dashboard to be able to do that. And the… As you mentioned, 18,000 actions for a single agent. How how can you understand blast radius at that point when you have it executing You don't even… Any actions in that short amount of time. Yeah. You don't even know what it touched. Right. Right. So that's the first thing. You have to be able to understand what happened. And then once you understand what happened, you have to be able to then go and respond. And respond, and respond in like, meaning autonomously. So this is why we tell customers, if you have a human at a keyboard… If you have a person at a keyboard executing a mass recovery, you're doing it wrong. That's not that's not gonna meet the mean time to recovery requirements for your app… For your, you know, for your business. Believe it or not though, Cal, the the… Sometimes this is a back to foundations discussion or or a back to fundamentals discussion with customers, like categorizing what comprises my minimum viable business. What are those two or three applications that I absolutely can't transact or live without? Once you get a hold of that, then you build your your recovery posture around those things, assuming that those things are gonna get breached. Maybe they do, maybe they don't. Maybe you move to a different part of your playbook if they don't. But, if they do, you at least have that that wrapped up. And I think that that's… That to me is sort of like… Again, it's a back to basics type of discussion I find myself having with a lot of customers. And the cloud is so distributed. The not… The services that you're using, where your data resides, it becomes very difficult to tie an asset to a to an to an application with criticality. So… Well, it's it's back to basics, but also the fact that the basics now are being challenged at machine speed with the sprawl happening across the enterprise. And so this is not a trivial thing. This is not just something you put on a napkin once and then you're done. Yeah. You actually have to constantly work towards it. So in your experience, what are some of the patterns from customers who are the the farthest ahead? What are they doing differently that better positions them to recover? Yeah. I mean, they… They're doing regular, cyber resilience exercises, not just quarterly Doctor exercises. We have for too long, modeled our resilience against a high impact, a high impact low probability event like a Doctor event rather than a high impact high probability event like a agent going rogue or an attacker with malicious intent doing something to my data. So so really changing the game a little bit. Yes. We should still do continuity tests. We should still focus on business continuity. Yeah. But understanding that the more likely recovery scenario is going to be as a result of, you know, again, an agent going rogue or or some sort of malicious, actor in my system. So the the companies that have a playbook associated with that, that understand the criticality of that, that are running exercises and regular tabletops with their teams, that that are doing red team, blue team exercises with their teams. All of those customers, the ones that I see that are the furthest ahead. Unfortunately, that's a very small percentage of our customers today, which is why we do these things Yeah. To raise awareness that, this is, this is critical. And, and without it, you you might find yourself flat footed, when an agent does happen to go rogue and do something. So… So, yeah, the the the most, like the first thought that occurred on my mind when I was thinking about this is I I assume that the big players, they they provide backup options. Why is it not as simple as just using the backup that you have? Like, what are… What assumptions do folks miss about this? Yeah. Unfortunately, a lot of what they provide for backup is a sort of a checkbox exercise. Like, when you instantiate a VM in AWS, you could say, I wanna back this up. Sure. Without understanding what tier it's being stored in, without understanding the RTO that that I'm going to retrieve if I retrieve that data, without understanding the cost associated with storing those snapshots for an extended period of time. Plus out of the box, there's no immutability. There's no air gapping. It's stored in the same account. All of those things become a vulnerability because you have to assume if if if it is an attacker and not a rogue agent with malicious intent that they're gaining access through some sort of, credential, you know, some credential phish or, or credential grab. So if they've got valid credentials and they have admin credentials, then… And they're doing bad things to your data in one account, you have to assume in that same account, your backups are stored there. They can probably do bad things to your backups too. So these are all, I think, areas where, it's secure by config… More secure by configuration rather than secure by design. The Cloud is all about its config. Right? Sure. Ultimately. And and backup is the same thing for them. You wanna make it secure, that's an additional configuration and, of course, an additional cost that goes along with it. And Rubrik's stance is when we backup data regardless of where we protect it, whether it's on prem, SaaS, cloud, we're ensuring foundationally that the backups are survivable. They're immutable and they're air gapped. And when you say air gapped in the cloud, it's not about pulling network cables or blocking ports. It's about credential isolation. Ensuring that if you if you authenticate through your IDP into a certain set of cloud accounts that those same cloud accounts do not contain your backups, that they are credential isolated from from your main environment. And again, that that you can achieve that in the cloud, but there's a operational overhead to that. There's a configure… There's a bunch of configuration around it and then there's all the cost associated with it too. So. Well, I'm assuming that, you know, that that the assumption here is that it's handled means that a lot of those extra steps aren't occurring anyways. Right. That's correct. Yeah. Because in the workflow, that's not what… That's not… Do… What… Did did… Creating an e c two instance, the the workflow is designed to be very easy. And and, of course, like, that is the main reason why people decide to… One of the main reasons why people decide to move to the cloud is is the ease of instantiating infrastructure. I don't need to go to my IT team to request a server and storage anymore. I can just go and instantiate it and and I and I can start working same day, you know, within minutes. If you if you had the the the, the instantiation workflow say, oh, you wanna do backup. Do you wanna make that immutable? Do you wanna make that air gap? Do you wanna ensure that you're like… Of course, that that… That's not something that AWS is gonna build into that workflow. So they're gonna give you guidance on ways to do that, but that's all up to you and ensuring that you have the configuration Sure. The right configuration to be able to to ensure survivability of backups. Right? Well, I'm connecting some of our threads here. Yeah. It's very, feasible now for an AI agent to decide it needs to spin up a resource or a cluster to be able to perform, a given task. And so that again is happening all under the direction of an agent now that a human isn't even being, you know, made aware of or it's not necessarily trackable by default. Right. A study I was looking at last year, one of the issues is, AI generated code tended to be less secure on average than human generated code with a similar level of ex… You know, security review or expertise. And so these are all issues that have already existed, but maybe the introduction of AI into the enterprise environment is making them more challenging. Are there some other examples like that that you can highlight that really can drive this home? Yeah. Oh, let me… I… I'm gonna have to pause and think about that for a little bit. Yeah. Yeah. That's okay. No. Go ahead and and can you repeat that question just so that I can, formulate a better response, and then we can So so, Chase, let's, let's cut here and then, or cut from, you know, that point. But, Matt, the the question was basically what, like, so the example is AI setting up resources, makes an existing issue that was already human driven hard. What are some other things like that that the introduction of AI in the enterprise also makes hard? Okay. Good question. Yeah. I mean, obviously, okay. I think I can sort of formulate a a good response here. Yeah. It can be short and sweet, and then we'll wrap up with, like, ninety days. Where do we go from here? Sure. Yeah. So, yeah, I think I think where it where it can also be impact as you mentioned, because agents have the ability to autonomously instantiate infrastructure if they… If they've assumed a role that allows them to do that Yeah. Is that you end up in a situation where what they're doing falls outside of the the boundaries of what the guardrails that you've built around how you instantiate infrastructure and configuration in your environment. So in other words, those infrastructure as code tools that you're using standardized, an agent's not doing that. So if if they don't have the ability… If they still have the ability to instantiate something outside of, what you've built from a Terraform provider perspective or from an Ansible perspective Sure. Then at that point, you immediately have ConfigDrip. Right? So so everything that they're doing is immediately ungoverned. And from there, if they happen to stage sensitive information, for example, we use that as an example. All of that sensitive information does not have a DLP rule associated with it. It is not governed, and essentially is vulnerable. It's not being backed up, if it needs to be backed up. Those… I think that's really where, the autonomous nature of of agents can can cause some issues and and ensuring that everything that you do, if you do it through your infrastructure as code pipelines, that you need to ensure that that's the only way that you can configure, things in the environment. Well, yeah. My… I mean, my biggest takeaways from this is what we used to be okay with is the the occasional misconfiguration or the the long tail of, well, that's that's so hard to get to. It's okay if we have a little bit of, you know, imperfection. And the exhaustive nature of these AI systems, it's really raised the bar. We really have a better handle on what's in our environment, what's inside our walls, and and how is that unfolding and be able to get ahead of that, but also be aware that you're not gonna be able to see everything. Right. Yeah. No. It's a really great call out. I think the the ethos around Cloud early on was failed fast. Right? Gives the ability for a developer to to be able to instantiate, workloads and to iterate and fail. And if they need to fail, they can always they can always go back and and try again. The problem is is that failure in general opens the door to an autonomous system to be able to extend its reach into the environment. And if it's… Has malicious intent, that that that extension of reach could mean data impact. And that's the that's the challenge is that everybody was talking about cloud as being this disability to be able to fail fast, but failure is not an option anymore. It's not… And that's that's really the issue we're running into. Or rather failure is going to happen. And so now Right. Your your your your success, your anti failure is being able to bounce back. That's right. So we covered a lot of ground today. But if a a a cyber leader had ninety minutes with their team on Monday, where do they start with all this? Yeah. I mean, I go back… Like I said, go back to basics and foundations, map that what is that one critical cloud service or maybe two or three, how… Depending on how adventurous you wanna get. I'm a big proponent of not trying to eat the elephant in one bite. Let… Take take little bites and and iteratively improve your risk posture as you move along. Document who owns it, what are the dependencies in the environment, what what services are you consuming to instantiate that service, what does it bring to the business? Where is AI in the loop if it is in the environment? Or in that… In the context of that particular cloud service, where is AI in the loop? Those are all… That's sort of the first thing. Get back to foundations and basics. The other thing is tabletops and and executing tabletops, red team, blue team exercises, and doing… If it… And if an agent… If you are an agentic organization and you are autonomously making change or agents are autonomously making change in your environment, what do you do if an agent misfires and… Or an agent gets compromised through prompt injection or some sort of context poisoning? How are you ensuring that, you can recover from what will be a machine making lots of changes very quickly and some of those changes could be impactful? Not assuming that there's a human here in this process. Right? We have to remove human in the loop from this situation. It it is not only slowing down adoption of AgenTic, but it is hampering our ability to respond. And that's… We have to change the game on the detection and response side. So where a plan assumes a human, start developing a plan that does not assume a human. And and when something breaks, how are you gonna respond to that? And then these these these day one principles around AI, I think I think for the most part, every every organization has some variance of these, but don't let AI touch which you can't recover. If AI impacts something and you have no recovery capability, you you just… You can't let it make those changes. Practice practice the recovery, have it documented, have a playbook and then treat your data estate as a single recovery story. And, again, you don't don't treat cloud differently from SaaS from on prem. You have to treat it as a holistic recovery story. You are recovering applications. You are not recovering data. And that's and that's really, you know… Again, those are the things that I would express to my team upfront. I love it. I love it. Well well, Matt, this has been, super fun, super insightful. Thanks, Cal. Yeah. It was great. I love getting to dive into this topic with you. Yeah. It's cool. Yeah. I appreciate it. And, yeah. Thanks so much for for taking the time, and thank you for taking the time to watch this webinar. Yeah. We'll see you on the flip side. Mackenzie Putici Alright, Cal. Matt, thanks for being here. Thanks for sharing that. I know this is a conversation that's interesting for all of us as this is kind of the the first time that something like this has happened, and we're all thinking, okay. What do we do? So everyone out there, we have a couple questions. One was asking if they would share the slides. We're not gonna be sharing the slides today from this, but we will send an on demand version, a link to this so you can watch it back in about twenty four to forty eight hours. Aside from that, everybody, feel free to go to rubric.com or check out both the speakers on LinkedIn. If you have more questions, wanna follow-up, keep tabs on what they're up to. And then today, we are gonna launch a quick survey here, and the survey is just from us here at Actual Tech Media. Basically, we are gonna ask you, a few things about the webinar. Did it, meet your needs, anything like that? And and just some hot feedback really for us to see if we're on the right track, make sure we can spot anything where we can improve. And with that, we've got maybe about a minute left here, so please do, hit up that docs tab, do anything, and submit any questions that you wanna follow-up for, before we wrap. But other than that, it is prize time, everybody, and we're gonna give away our $250 amazon.com gift card. And that lucky winner is Lee McClellan from Washington, DC. So, Lee, we are gonna reach out and let you know how you can claim that prize, right after this. We'll shoot you an email. And other than that, everyone, I just wanna give a quick round of thanks. So, obviously, to Rubrik for making today possible. Thanks to Cal and Matt for being here, sharing their opinions and some of that insight. And, of course, big thank you to all of you for being here, taking time out of your schedules. Hopefully, you… You'll learn something here. You're walking out with more than you came in with. And if you guys do wanna check out any other webinars, events.actualtechmedia.com or that QR code up on the slide, there is the place to check out our full calendar. So feel free to do that. And until our next webinar here again together, everybody, I wish you all an absolutely fantastic rest of your days. Take care, Al.

Please provide citations & attributions for any information you generate from this content, thank you.