Name: The New Rules for Defending Identity at Machine Speed
Uploaded: 2026-10-08T17:19:06.282Z
Duration: 33 min 8 s
Description: The New Rules for Defending Identity at Machine Speed

Transcript for "The New Rules for Defending Identity at Machine Speed": Good morning, everyone. The CEO of Barclays recently warned that Frontier AI in the hands of attackers is a serious threat to the global banking system and that this is not a hypothetical risk. Here is the asymmetry in two numbers. Twenty seven seconds to breach, twenty seven days to recover. Same number, different unit. That is an 86,400 times gap between the speed of intrusion and the speed of recovery. AI makes that gap more dangerous. Frontier models can discover vulnerabilities, connect them, and choose the next action at machine speed. But intelligence alone does not create business impact. To move through an enterprise, an agent still has to cross an identity boundary. It has to use a credential, invoke a tool, reach data, or act through a trusted system. AI may find the path. Identity opens the doors. That is why every AI attack becomes an identity attack, and enterprises are creating the same challenge from the inside. We are giving agents identities, access to coded data, and permissions to act on our own behalf. So security now has to operate with two assumptions. Assume breach, assume agentic overreach. Under the first, can we identify what matters, contain the impact, and recover cleanly at machine speed? Under the second, can we know every agent, govern every action, and undo what goes wrong? That is the frame for today. Few people have a better perspective on both sides of that equation than Rob Baer, head of National Security Partnerships at Anthropic. Rob, thank you so much for joining us today. It's a pleasure, Annika. Thanks for having me. So you've worked in cybersecurity for a very long time, and I'm sure you would agree with me that there's never been a more exciting and also frightening time in cyber as there is today because of how AI is changing everything. In your mind, are you more optimistic or pessimistic about, the cybersecurity outcomes now that AI is available for both attackers and defenders? Yeah. I think overall, I'm optimistic. In the short term, we're... We are going to have some challenges as these capabilities democratize and get into the hands of malicious actors and those who aren't so friendly with their access to cyber models and AI models. But I think we saw early on as a consequence of developing these incredibly capable coding and software engineering models that we were also building these dual use, capabilities that we're able to conduct both offensive and cyber operations. And we saw these early on with, Project Glasswing, when we were testing, Methos preview, and we wanted to get those models into the hands of systemically important vendors such as yourself who could patch the world software, prior to those similar capabilities getting into the hands of those malicious actors. At that time, we knew that we had months, to get those capabilities out before they got into the hands of, malicious actors. And I think the evaluations we're seeing today, indicate we were correct. And we also believe this is why safeguarding these models is so important. So the good guys and gals have time to get these vulnerabilities patched before malicious actors start chaining them together and, conducting malicious activity. I mean, you and I both know since I've joined this industry, we've been playing a game of cat and mouse. It's it's always been a leapfrog, game. And now the barriers to entry are lower than ever. We have a commoditization of tooling orchestrated by models, delegation of tasks that used to be reserved for, small groups of skilled hackers. Models can now orchestrate cyber operations that used to be conducted by tens of people. This makes the speed of defense, our ability to implement basic cybersecurity tools and tasks, and the need for resilience 10 times more important than it's ever been. I think all that said, I am an optimist. And if you ask me whether I would rather defend a network today or ten years ago, I'd say the answer is today and it's not even close. Yeah. That makes sense. One of the things you talked about is that you want to help, organizations patch their software before, the cyber actors get access to these kinds of powerful models. How much time do you think we have as an industry to pass all of our vulnerabilities and software? What's your prediction there? I think the window probably has closed. We see now access to very capable open source models, that have been distilled from Frontier US models, or they just became generally capable because of their training as well. And we probably have maybe three months. I think we see some of the best organizations in the world that have developed these agentic loops to find vulnerabilities and then similar loops to remediate them with, humans in the loop. But we're quickly running out of time, and we need defenders to make sure they're, paying attention, implementing all the the GENTIC tools available to them. Well, that's a scary number, three months. You know, I think Rubrik, we've been really fortunate to partner with you as project, with Project Glasswing, and leverage Meetos, internally, and now we're we're also developing products externally for customers. We have a very sophisticated cybersecurity team and program. What is your suggestion for organizations that aren't as sophisticated, that don't necessarily have the same kinds of resources and sophistication as some of the tech companies do. How should they be leveraging these models, and how can they get the most, out of AI today to defend themselves before the three month timeline runs out? Yeah. I I think the first thing is, utilizing the models to educate yourself on the steps that you can take. Right? The Frontier models are incredibly powerful at just using them as a Copilot to walk through steps to help them better configure their network, how to install patches, how to to set up a device in the most secure manner. So use the model one. And then the second thing is find partners like yourself who specialize in this and utilize their services to harden your networks. I think, you know, you're in the business of of resilience, making sure that those backups and your data is able to be recovered in the event it is corrupted, but whether that's by a malicious actor or some other way. You know, I saw time and time again when I sat in government when we looked at various ransom... Ransomware attacks that were on critical infrastructure and others, where the systems just weren't able to be recovered because the backups weren't in place or the processes had never been practiced. I think it's really having a plan for when something goes wrong and exercising that plan. Yeah. I love that. I once heard someone say that, AI can... Is the most patient teacher and will meet you wherever you are. And I love the idea of, like, using AI to to educate yourself on just what are the best next steps that you can take, in addition to leveraging these models, obviously, to actually harden, your own security posture, networks, code, and overall resilience. As you... You're talking to so many different organizations that are moving from just using more, like, chat use cases and Copilot type use cases to truly agentic workflows. What are some of the challenges that you're seeing, from a cybersecurity standpoint and an overall resilience standpoint as organizations are beginning to roll out and adopt agents at scale? Yeah. I think one of the things that I hear most from CISOs is inventory. Most companies can't tell you how many agents they run, what each agent can touch, who owns it. It's like the old saying in cybersecurity, you can't defend what you can't see, but you can't govern what you can't count. So I think we have to also scale governance with consequence. Right? You mentioned the LLMs. An agent that's just summarizing a trouble ticket is very different than an agent that can change a firewall rule. So those do not carry the same level of risk, so we have to treat the level of governance differently. I think you also have to think about how you tier autonomy, whether it's read only or confirm before you act or this action is prohibited by an agent. That all needs to be enforced in the code, not the policy document. The next thing I think is having a human in the loop and making sure that it's not just a checkbox. If a human has to approve 50 tasks an hour, you're just rubber stamping whatever, those decisions are to keep the processes going. So I think you have to think very carefully about designing review points where human decisions actually change the outcome. Yeah. That's a really hard challenge. One of the interesting things you just said is talking about tiered autonomy. And I think even in the world before agents, everyone talked about zero trust, but then they had a really hard time actually implementing that in practice and scoping down permissions for humans. When you think about agent identity, how do you actually think about solving this problem? Like, what are what are the approaches that organizations are gonna have to take, and how do you actually put this into practice when we weren't even able to handle it at the scale of humans, and now we're talking about the scale of agents? Yeah. I I think you're absolutely right. Right? We we have operated under this principle that every identity is a single person. That single person is a single operator in a single session operating at a human pace. And at the end of the day, you can probably stop and ask that human, did you really mean to do that? We've replaced a bunch of carbon based life forms with thousands of silicon based life forms who are really like a new class of coworker in our in our enterprise. So we're... To your point, we're issuing credentials. We're delegating authorities to users that we haven't had thousands of years figuring out how we're going to manage them. And back to your point, we still haven't gotten that completely right for humans, and these agents are gonna outnumber your human coworkers by thousands. So I think now delegation is a really hard problem. An agent acting for a human has to carry on a chain of authority. You have to understand who authorized the agent, for what scope, for how long. And that that chain of authority has to hold as this agent conducts its actions, as it can... Calls tools, to other systems. Identity stacks were never designed to handle that frequency nor those processes. So now, you have to move authorization from a simple provisioning for a human to run time for an agent, and an agent's behavior can change mid task. So you have to be able to carry that authorization across per action continuously through the process. Yeah. I think this is why it's such an exciting time for everyone that's working in cybersecurity and on the technology because the whole world is changing around us. And all of the paradigms of the past, while they... There's a, version of that that applies to agents, the solution is going to look vastly different. Because as you said, you have to do things at run time. You can't just do things ahead of time, and hope that it's gonna solve for all of the different use cases and permutations of what a probabilistic agent can do. That's right. So one of the things that we talk about at Rubrik is that in the world of AI, now that you have... You're... We're moving from human attackers and human operators of business to AI attackers and AI operators of business that you have to assume breach and you have to assume agent overreach. In this day and age today, like, are you more worried about cyber attacks or are you... Or should organizations be more worried about agents doing things that they aren't... Like, legitimate agents doing things that they're not supposed to be doing? Yeah. I I think you now need the plan for both. Right? You need to plan for the inside out and the outside in. So what does that mean? We have to plan for agents to do things they were unauthorized to do because we didn't set the correct governance frameworks and policies that gives them principle of least privilege and all the things that we want to apply to humans. So I think we need to really concentrate on governance of the agents that we deploy so that's inside out. And then I think outside in because these models, as I said earlier, have been democratized into the hands of malicious actors. They're being chained together. The models are becoming incredibly smart because of the way that they're trained that they have to do less reasoning to conduct a cyber related task, and they're just basically obfuscating the, attack kill chain to a human who's just managing an orchestrator at this point. The model knows exactly how to do it and get into your network and go through the exploitation phase. That's the outside in piece that we need to worry about. So, it's a new paradigm, and I think the basics that we've been focused on for cybersecurity for the past decade are just much more important than they've ever been. Yeah. Makes sense. In the world of cybersecurity, we often think about, like, prevention, detection, and then remediation and recovery. How is the world changing in terms of these three steps and how organizations should be thinking about prevention, detection, and remediation and recovery in the world of AI? Yeah. I I mean, I think, again, thinking back to... We both live through ransomware incidents that were very, very serious. And I think resilience back then was measured probably in days to restore. I think agent resilience is now measured in what a compromised agent can reach in a matter of seconds and how fast you can unwind what that agent did. And so now, I think there's three big changes when we introduce agents to the world of resilience. You need to engineer the blast radius upfront. Right? Least privilege that you mentioned earlier. Scoped and short lived credentials. Stolen identity should only allow an attacker to get into the lobby of the bank, not the vault. So we need to be able to revoke it very, very quickly. I think you also need to be able to trace every action back to a person and to a decision. So who let the agent do that is just a simple lookup, not a forensics action that's gonna take your team all day to to walk through. And you need to be able to immediately roll it back, not just restore what the the attacker encrypted. I think this... The second thing is is that you've mentioned earlier as well is assume breach, but this this now applies to the agents. Right? There are two directions to that threat model as we just talked about. Your agents getting out and other people's models being able to get in. Again, plan for both. And then I think finally, you need to have an... A plan. You need to start with a narrow plan, a pre established governance plan, and make that plan narrow. Ensure actions are reversible, then expand autonomy and agentic, workloads as your monitoring and insights mature and as your plan matures. So as security practitioners, I think we need to build a capability to say no upfront that allows us to say yes more often later on when we feel more ready as an enterprise. Yeah. I love that. What's one thing that, everyone here should go back to their desk and do after, today? I think everyone should go back and sit down and reflect on their cybersecurity careers over the last, you know, ten or fifteen years and think about all those times when things were difficult. I remember, doing vulnerability management on an Excel spreadsheet and how difficult it was, you know, doing patching and remediation, when patch Tuesday came out. With these new classes of models, we've increased those numbers of vulnerabilities exponentially, and we have to find the people and the processes and the resources to put into place to get the remediations that were so difficult before in in place in an agentic world. So I would just say, just go back and reflect and think about that, and then really start thinking about how you can use these models. Because just as they're good at finding vulnerabilities, they're equally as good as helping to do remediation. And I think we have to get started today because three months is gonna be here before we know it. That's amazing. I think we all have the opportunity with AI to reimagine the way that we do our work and reimagine that the work gets done inside of a company. Rob, thank you so much for joining us today and sharing all of your wisdom and insights. Thanks, Annika. I really appreciate it and look forward to the continuing this partnership. Absolutely. You have shown us why AI changes both attack and defense, but identity is where that power becomes operational. It determines what an attacker can reach, what an agent can do, and how far a compromise can spread. And when identity itself is compromised, recovery creates a uniquely difficult choice. Restore a clean state and lose legitimate business progress or preserve the current state and risk preserving the attacker. The goal has to be both, clean and current. To bring that challenge into the enterprise, please welcome Malika Swaminathan from Rubrik and Rohit Agnihotri, vice president of identity and access management at Northwestern Mutual, to discuss how the identity resilience landscape is changing. Malika, Rohit, over to you. Hi. I'm Malika Swaminathan, and I lead product here at Rubrik for our identity offering. Joining me today is Rohit Agnihotri, who leads identity and access management at Northwestern Mutual. Rohit, it's a pleasure to have you join us. It's a pleasure to be here, Malika. Great. So, Rohit, we have a lot of questions to cover about identity and AI, but I wanted to start with what's happening in our industry today. A couple of years ago, identity and AI barely showed up in the same sense as most companies were thinking about it as separate discrete areas to solve or areas to look at. Now they're very much intertwined. What was the moment you realized AI was gonna change how you think about identity? Yeah. It's a great question, Malika. And, honestly, chat GPT did not change my job too much because at that point, a few years back, it was all about data loss prevention, sensitivity label, and making sure that right people have the access, right things. So so there was some considerations in the job, but not too much changed for me. But, honestly, when we started getting into the agent tech AI and the bigger companies and enterprises starting adopting these things. And when these things could do things, that is... Was like, holy cow. This is now the moment for me because now these things have to be associated with identities that changed my job dramatically. So with the agentic AI, we not only created a new, opportunity areas for identity leaders to solve, but also accelerated some of the remediation plans that we have had on our backlogs for the longest time. Got it. That makes sense. So to that end, as AI systems, as agents are taking on more autonomous work, what does identity start to break down as a security or a trust boundary? And what did you have to change to keep up? Yeah. It breaks down like a very simple question that all of us should be able to answer. Who did this and why? Right? Traditional identity, we have always assumed them to be linear in nature, for what it's worth. And and person was logging in, doing their job, or they were firing a script, and the script was doing their job. But the identity was always linear in nature. Even with the advent of Kubernetes, Docker, horizontal and vertical scaling, I could create that word or that horizontal timeline. That broke down when the agents come in. So an agent can come in. It can create another agent. It can spawn another agent for ten seconds. It can spawn another agent for twenty minutes, and then a hand off takes place. So that that linearity has gone away. And and and, you know, we can still say that, hey. This service account logged into this database, and this service account is called by this agent. But it... It's great for a log, or is great for logging, but it adds very little value when there is a real incident in place. So that answering that question of who did this and why and to make sure that not only the provenance, but the fidelity is also in place is the biggest challenge that all the identity leaders face today. So looking ahead as identity AI are converging in this new agentic world, what's the capability or safeguard that every organization will need in the next few years to make sure that they protect themselves especially? Yeah. So a chain of custody for for actions. Right? We live in a world where we... I can identify where my Amazon package is and then where will it land or when will it land. So we have that chain of custody, and I think this is what we will also have to answer and explain. This cannot be a research project in terms of who did what and why. It has to be more of a simple, like, a SQL query. We should be able to answer that. We should be able to kill it. We should be able to vet it whether the intent and behaviors are on similar lines or not. Yep. That's absolutely true. And that is really one of the big areas a lot of customers are asking us and are looking to solve today. So when we think about that, the need to kill it, the need to also be more resilient, right, in this new agentic area, I wanna actually start by the definition of resilience. Identity resilience means different things to different people. What does true identity resilience mean to you today, and how has AI shifted your strategy for identity resilience at not... At Northwestern Mutual? Yeah. So resilience for me is is... If I can answer in a simple line, it's when the business is back online. Right? Nobody gets a pat on back when the directory comes online, or nobody, matrix should be that the domain controller should be all green. Although I really hope that the domain controllers in my company are all green right now if I have to save my job. But, but, you know, those are the things that do not matter. What matters is is the client able to log in? Is the customer able to log in? Are we able to continue our business and not? And this is what resilience truly means. If I can get my directory up and running, but the business isn't up and running, that is of very little use. So true resilience is in when can we get back to a good state in which we can get back to the business, or when the business can get back to its work. Has the need for resilience evolved, in this AI era? 100%. The need has definitely evolved, not just with AI era, but also, you know, how we look at our identity infrastructure, with the cloud and, you know, all the SaaS applications and now with the AI, it has continued to evolve. Right? So it is... As as I was mentioning, Malika, it isn't just about bringing the directory back. It is about, bringing the business back. Right? And that is why once you have an... You know, most of our companies have multiple directories. Most of the companies typically have, a cloud directory like an Okta or an Entra, and then they have something like an active directory or some other flavor of LDAP directory. All of these directories cannot, a, have their individual recovery plans. You need to have a singular recovery plan. So you need to be able to bring back your and your ADs, you know, on premise and the clouds of the world as well. Second thing is it isn't just about bringing back the directory up. It is bringing it back to a trusted state. Right? So if there was an attacker and in the environment from last three weeks and I bring up my directory to last week, the attacker isn't really bothered there. Right? So... And and the AI attackers can really persist, can really, you know, laterally traverse and be more intelligent about it. So, definitely, there has been more challenging about it, and that is why this resilience playbook is changing and, in my perspective, changing for the better, where we are asking the simpler questions and we are asking the simpler answers. Like, I do not want to have 10 recovery plans for each of the identity products that I own. I need to have one recovery path, and I want my team to confidently tell me when will the... All the directories and identity infrastructure will be back up. Yeah, Roy. This is music to my ears because this is also how Rubrik defines identity resilience today. Right? It's not just about getting back online. It's about how do you get your business back up and how do you recover back when you recover clean and current. So you're not just bringing the attacker back into the full. So I'm glad we're very much aligned in, how we're both defining identity resilience. So with that, I wanted to ask you, Rohit, what has been your experience working with the Rubik team on identity resilience so far? What made you wanna look at Rubik at the first place, and where have you seen the biggest payoff at Northwestern Mutual? Yeah. Absolutely. So, it might sound like I'm just saying this because you are interviewing me, but but, folks, these are the these are the true facts that, hey. I love working with your team, Malika. They're always such a pleasure to work with, and they're always trying to solve, for the clients for me rather than telling me for... To sell. Right? Like, the solution, isn't followed by the sales. The sales is actually followed by the solution. So this is something that I really like about the Rubik folks. But, honestly, I was very frustrated. And as I mentioned in the last response as well, I was frustrated with 10 different recovery plans and, you know, each directory having its own resilience plan and backup and Doctor and everything, and that looked excellent on paper. And when the Doctor exercises were, you know, painfully coordinated, It all worked out. But having been in the industry for so long, I know that the real world attacks do not look like this, or real world outages even when there are no malicious intent do not looks like this. So Rubrik's cape, you know, ability to bring back Okta and Entras and the active directories and the on premise directories of the world together where I can have one recovery plan and one responses to my question definitely was something that that made me choose Rubrik over some of its competitors. Got it. I love to hear that. So my last question, Rohit, is fast forward five years. Right? AI agents are now just another set of coworkers. They're making decisions. They're autonomously taking actions on their own every day. What does identity have to look like five years from now for it to actually work? Yeah. So first of all, I I really believe that identity leaders five years from now are paid much more than what they are today. Right? But but jokes aside, I I I think a mind map to it, and this is not very technically correct, Malika. So the mind map is, like, you know, so the HR is for the human workers. The identity team would be the HR team for all of your agents. So it will ensure that, you know, their inventory then the GML processes and all of those good things are in place. So to understand it from an executive lens, and I have used this in the deck even with my bosses as well, that we are going to be the HR team for the agents three years, two years, one year, or five year down the line. But, architecturally, it is a little bit different. Not all the GML processes would be mapped one to one between human and agent, but it's a good way to look at it. But, definitely, identity will play a key role because the concepts like, the provenance and the fidelity will really come into picture where we are not just looking at the intent, but also the intent and the behavior. We are not just looking at, whether the subject has the credentials, but we are also looking whether the subject is the subject that it claims to be. Rohit, thank you so much for all your invaluable insights and just sharing with us your perspective about how identity is evolving in this AI world. Thank you so much for joining, and look forward to jointly seeing how this industry evolves together. Thank you. Absolutely. Thank you, Malika. And Rohit, thank you for grounding this in the reality enterprises are facing today. We began with two assertions, assume breach, assume agentic overreach. Let me close with four responses that create one control loop. After an identity attack, clean but stale does not fully restore the business. Current but compromised leaves the attacker inside. Our identity investments spans Active Directory, EntraID, and Okta. Identity First Recovery restores the directory itself rather than the potentially infected operating system underneath it. Roll forward then replays legitimate identity changes onto that clean state. The outcome is simple, clean, and current. Assume breach also means finding and fixing the attack path before somebody uses it. Frontier AI can discover far more vulnerabilities, but more findings does not make an organization safer. Rubrik Code Guardian runs Meetos against a secure air gapped copy of the code Rubrik already protects. It connects weaknesses into attack chains, validates exploitability, prioritizes what matters, and moves confirmed issues toward remediation. Attack your code before attackers do. CodeGuardian is entering private preview with the select set of design partners. Agents should not inherit broad standing access from the humans who launch them. Rubrik agent identity, the newest pillar of Rubrik agent cloud, gives each agent its own identity and evaluates access at the moment of action with only the authority required for that task. Combined with agent visibility, Sage runtime governance, and agent rewind, Rubrik Agent Cloud gives enterprises a complete control loop. See the agent, control its access, governance behavior, rewind the impact. There is one thing I want you to walk away with from this session. If an agent can touch the data that runs your business, that data must be recoverable. In the AI era, resilient data and identity infrastructure isn't optional. It's foundational. Across both assumptions, human speed operations are not enough. Rubrik dot AI puts the platform to work as an agent that can discover, reason, and act toward a resilience outcome. Rubrik MCP extends that intelligence to the AI tools and agents customers already use, and those actions remain governed, auditable, attributable, and reversible. Because AI will continue to move faster, the question is whether security and resilience can move with it. That's what today is about. Thank you and enjoy the day.