325. Addressing AI Frontier Lab model risks and opportunities
Inside the Strategy Room · 9 Oct 2026
Source: https://omny.fm/shows/inside-the-strategy-room/325-addressing-ai-frontier-lab-model-risks-and-opportunities
Transcript: https://podcasts.opennash.com/shows/inside-the-strategy-room/325-addressing-ai-frontier-lab-model-risks-and-opportunities-fe92b6/ (via OpenNash Podcasts, machine-generated)

From McKinsey and Company, I'm Sean Brown and welcome to inside the strategy room. AI frontier labs are presenting organizations with a wide range of increased risks and opportunities. Cyberattacks have become faster, cheaper, and more autonomous. Agentic AI is simultaneously shaping new workflows, while third-party dependencies are complicating supply chain risks. And let's not forget quantum is still coming. So today we'll discuss the implications of these frontier lab models for how organizations should evolve their cybersecurity strategy, what it means for their operating model and decision making, and how humans can still stay on top of the AI agents in this new world.

Joining us for the discussion are Marie Claude Nadeau, senior partner in our Bay Area office and the global co-leader of our risk and resilience practice, and Charlie Lewis, a partner in our Connecticut office and global leader of our cybersecurity practice. Welcome to you both. I'm really excited for this conversation.

Thanks for having me, Sean.

Thank you, Sean.

Okay, MC. We're seeing a lot in the press about frontier AI agents going rogue, doing things that their users weren't intending, and of other cases where they're doing what their users did intend. But those users are bad actors, and they're asking agents to attack companies by weaponizing structural flaws, and they're able to do it really cheaply and quickly. So it's easy to see why this is an important topic, not just for chief information security officers or CIOs, but also for the CEO, CFO, and entire C-suite.

Can you set the scene for us? So how should leaders prepare their organizations for these rapidly emerging risks, while also trying to make sure that they make the most of the opportunities that these developments can still present?

Well, as as you said, this is bigger than just cyber leadership, right? This is more than just more cyber attacks and cyber leaders having more work on their plates. There are several structural shifts that are happening at once and that are making it even more costly to be slow in responding to cyber attacks. Like you said, Frontier AI is making it really, uh, fast and cheap and autonomous, frankly, to actually, um, attack organizations. And the time between an organization noticing that a flaw exists and someone actually weaponizing it is collapsing from years or a year to a day or less.

And every control budget cycle change management process today has to, uh, change because they were designed for an older, slower clock, and this clock speed is increasing very, very fast. At the same time, AI is also changing how the businesses themselves operate. There's agents available, new workflows, new governance requirements inside these organizations that are changing how cyber leaders can react and and business leaders can react to these attacks. And, thirdly, critical dependencies are emerging well beyond your third parties.

Um, organizations used to be able to review their third parties at kind of a one layer level and identify dependencies or vulnerabilities there. Now, these vulnerabilities extend into all of the cloud data identity model infrastructure that they might be tapping into, even unknowingly. And then, you know, maybe something that's helping, uh, but also complexifying the response for cyber leaders is that the provider and technology landscape, um, that's available for them to even help respond to these attacks is moving very quickly.

These capabilities are evolving every day, and so they have to keep up to speed to understand what, um, what the best technology out there is. And then finally, the bar has just risen overall. There's a very high bar on your existing cyber fundamentals, the costs of weaknesses, even small weaknesses that exist in your organization that may have existed for a long time. That cost is increasing very fast, and the speed at which they have to be addressed is is increasing as well.

Thanks, MC. So how are companies dealing with this as as they develop and deploy new technologies? Are they just raising the bar on building quality into the initial release, or are they using AI to help make sure that before they release, some of these previously inconsequential vulnerabilities are no longer actually getting built into the system? And Charlie, maybe you could take this.

You know, Sean, I'll, I'll take that one quickly. You know, quality has, you know, I think companies have always thought about quality, right? They've always done their quality checks, they've always gone through and made sure, uh, that the application was working or the product was working. The shift now, though, is that you can't think about quality without putting security in there too, because you ultimately could have something break. I think about medical technology devices and the need to make sure that they are deployed securely and built securely to go into an operating room.

And the questions that might be coming from hospital, uh, you know, hospital buyers towards those medical technology sales and saying, what is that integration? And how does that look? Or when you think about, you know, products going into these broad data s uh centers and what they need, anything from some of the core cords that plug in to the potential large scale turbines that would run these. And as you think about these broadly, you start seeing then the simple applications like the one we're riding on right now.

And so what you're gonna see is one, an increased demand, but two, AI actually here to get to a little bit of of of what MC was saying actually sort of lowers that cost. It it speeds up the security checks for a secure development life cycle, um, in a way that does not slow the time to market. And allows the business to still achieve its goals. And then deploy its deploy that application securely in an automated fashion, with automated remediation and automated alerts, and all of this moves a bit faster.

In that way, we've seen organizations start using agentic threat modeling and then broader integration and automate automation of the workflows through their various SAST and DAST tools.

So we mentioned earlier that AI is also making it easier for attackers to go after companies. And so this is probably one of the big things that's keeping the C-suite up at night, right? So where is that having the biggest impact? If there's one area that the senior executive team really needs to pay attention to?

of course, and and Sean, look, I think I think that AI isn't necessarily creating a new attack life cycle. But what you are seeing is that it is compressing this attack life cycle, it is compressing various steps, right? And it's creating components and e escalating where where organizations may already struggle to defend certain elements within there. I think, you know, as you go through here, there's a great quote from a a blog that a regular security just put out, right?

That says for most of the history of cybersecurity, sophisticated attacks were scarce because the people who could run them were scarce, right? And AI changes that constraint across this entire life cycle and accelerates it too. I think that one of the biggest changes here is the compression of the time and effort to actually attack. And this gets back to that quote, right? Work that once requires specialized expertise and significant manual effort can be increasingly automated or augmented by AI.

And and I think about, you know, you would train specific folks who could break a certain kind of router, and they were very narrowly focused there. And now I don't have to do that training. I do think there are a few components that I want to highlight. Number one is recon. We typically forget what recon can do, but this often required, you know, substantial manual effort, right? And now it becomes increasingly automated. As you can think about mapping targets, what is the most valuable target that you can run into?

How do you think about accessing that target? And then what is the value you could get from said target after you deploy? And this is actually what I think is the most consequental consequential shift that is likely to put defenders on their back foot for a period of time until they can start moving forward and shift from sort of this offense dominated world to a defense one that's dominated by defense. But in reality, it's exploitation, right?

This is what matters the most, right? Known vulnerabilities can move towards weaponization in a matter of hours or even minutes, potentially, rather than weeks. And then you can start taking what were once seen as minor weaknesses. They can now be chained together and can change the risk associated with vulnerabilities that historically one received lower priority and two were just left and weren't and weren't dealt with. And so when you think about this, the core conversation is about, well, can this be exploited?

And what is our exposure? And that changes the vulnerability remediation calculus and the speed that organizations have to respond, right? And so it's not just about finding more things, right? It really gets down to how do I decide to detect, contain, recover and validate much faster to stay ahead of the agents? And how do I build AI into my workflows?

So some of these vulnerabilities, Charlie, are sitting unknown until the agents discover them. It's not like they were already around and known for months and months,

Or even decades, right? Uh, the, you know, there was, there was a Linux vulnerability from 2003. I mean, that's a long, that's twenty three years ago, right? That's a long, long time ago. Um, right, that vulnerability would've been, you know, in the US, right? Of drinking age for a few years at this stage, right? They'd be probably working as an analyst in your company. That's a long time, right? And so you have to go all the way back and start cleaning those up now while also getting to the forward looking vulns that could be out there.

sure. And if the amount of time from discovery to exploit is compressing like this right down to basically, you know, hours from days or weeks or months. How does this mean that businesses need to be ready to respond? Like, what are the implications relative to the way that companies previously defended against these kinds of attacks?

Yeah, it's a great question, Sean, because, um, it's this is not just a technology change, it's an organizational change as well, right? Um, right now we have organizations that tend to look like pyramids, security organizations, right? We have a small leadership layer that's supporting a large base of analysts and engineers that are actually doing the day-to-day execution of solving vulnerabilities, finding them, et cetera. But now, with both the shift in speed of the attacks and the utility of AI, some of the repetitive work that used to be done by analysts and engineers can now be done by agents, right?

So number one, the pyramid starts to look like a different shape, more like a diamond, and we're calling it a diamond of our overseers. Um, because it's not just the shape of the organization, but the roles within the organization that are changing. There are fewer roles, um, dedicated to repetitive execution, and more people focus on orchestrating judgment, validation, exception handling. Um, the accountability doesn't shift, it still remains at the top of the pyramid or the diamond, with CISOs and function leaders still owning kind of the risk appetite, regulatory obligations, strategic decisions.

Um, but the middle of the organization is changing a lot. Middle, the middle layer needs to move from managing people to managing AI driven work. Um, and, you know, not move from doing some tasks themselves to not doing it as much themselves to increase the speed. Um, and then the final shift that goes along with that organizational change is really an operating model and decision making change. Because regardless of how your organization looks and who is doing what within the organization, if the decision making rights don't change, your speed of response is going to be slow, right?

And that's actually a very hard shift for organizations to make. If you're going to respond to vulnerabilities within hours rather than days or weeks, you need the um decision making power and the speed of decision making to lie with a lower level in the org organization. And that becomes actually a competitive variable. Organizations that can delegate decision making in an effective way become more competitive because they can respond much faster.

So the security teams really need to raise their metabolic rate, if you will. So does this mean that some of the individual contributors who were previously doing some of the repetitive work that you referred to are now becoming managers themselves, but managers of agents? And when you think about the organizational development implications, how do you train these people up so that they can make that transition to being the most effective managers of agents possible?

Yeah, it increases their the, uh, the burden on training those people, frankly, um, making sure that they um have the judgment to review what's being done by by an AI or an agent. It's actually a higher value task, and we, you know, they need to make sure that they're on top of it. And also, if we're delegating decision making power to them, there needs to be a definition of what those decisions can be be made at what level, and that's effective and safe.

So AI agents and AI more broadly, then would still boost productivity all the way to the top of the pyramid, if you will. But if you have agents operating at those higher levels, what are the implications there more broadly in in terms of how an organization might operate?

Well, we're seeing AI having productivity improvements throughout, right? Um, but the replacement of tasks end to end, I think, are at least today more on the repetitive task side of things. Um, and frankly, that productivity increase at the top layers is going to be needed to respond to the increased volume and the increased work that that there is. It's not gonna be wasted.

And yeah, and I think, like, one of the spots here, and like, it's a good question is you start getting into those higher levels, that that's where you start seeing the operating model shift on the work agents can take on. Versus, right, what, when does a human need to step in? As MC said, right, like a lot of the repetitive you are seeing that feed into the dashboards and thinking about some of the decisions that are required there.

But ultimately, we still see these leaders setting the guardrails, they still have to retain accountability. And then, you know, the agents will increasingly execute within those those boundaries. We see the agents still running it, sort of the find problems and decide what matters, and then, like, that escalation goes up based off of the guardrails. And then you move into sort of this human checkpoint around the fixing, which I think moves up a bit with it. It still stays where it originally was within this overall pyramid.

We still have full sort of alignment within the industry on not deploying own fixes without human reviews, especially on critical applications and critical business services. And so, you know, th- this is allowing speed, right? But you can prioritize and give all the data faster, and that's where we're seeing up high. But it still is the repetitive tasks that are feeding into the reporting.

Right? So if we think about how quickly these agents are moving, you know, the human might be in a one hour meeting. And in that short time frame, something has changed massively. There may be a huge attack. Or, and I'm interested to know how the humans are staying on top of their agents when both the attackers and defenders are moving at this really increased metabolic rate.

Yeah, no, I mean, it's it's a fascinating question, right? And I think it goes back to what we heard coming out of Black Hat with Hugging Face, right? That it was where there were so many attacks and such a swarm and movement going on all at once that the signal became the noise. And so how do you, how do you handle that? And look, I think each individual person has to be narrow. You've got to be ruthlessly focused.

But I also think it's recognizing how you establish the guardrails on the use of your agents and what they're feeding and how they consolidate the information. And again, this comes back to what kind of agent are you deploying? And and we've talked and published on this too, right? Around what are the level of guardrails you need to have? What are the identity requirements for each of the agents? How do you register them? How do you make sure you're not increasing the potential for an, uh, um, uh, an inadvertent insider risk moving through and building out these broader multi-agent systems?

But then creating a system that allows for not near constant updates unless the alerts need to be. And doing it in a structured manner so you can still have focus and thought as you're building and creating a more secure environment and not allowing yourself to to be overcome or overwhelmed by the amount of work that the agent is doing. Uh, cause this is a bit of a long game and we're not quite sure when you'll get to sort of equilibrium and comfort within that.

And so it's making sure you've got the right pace up front as you're going through. And what are those requirements? And maybe it's just one more agent to build that helps filter for you, right? And that, and that could be part of it, your post meeting agent.

And so it sounds like there are some pretty substantial operating model changes, along with colleagues lower in the organization that maybe had been doing repetitive tasks previously, they're now managing agents. We've talked about how they'll need training, but with these changes in the operating model and this metabolic rate, it seems that there could likely be some implications for budgets as well. So how are you seeing that play out, MC?

There's definitely an implication for B for for budgets, um, but it's not just increased budgets and everything will be fine. Um, actually, what matters mo most is where you're spending your budget in particular, as Charlie was talking about, uh, securing agents and how they can access your data is going to be a big focus in our analysis. Today, about four of cyber budgets are dedicating to agent, uh, agent security, and that could probably quadruple or so in the next three years as people are are focused more and more on securing those agents.

Um, in particular, a anything that has to do with identity and access management, governance, risk and compliance, and data security and privacy. Um, those are gonna be big focus areas. There's just gonna be more and more non-human identities trying to access data and more permissions to govern more sensitive data to protect. That's gonna be a big focus from a a spending perspective, hopefully, with some automation, and we talked about how automation and AI can actually help.

Also, that will moderate some of the spending growth that we would otherwise see because it is a big increase in workload. But with automation, hopefully, that's not necessarily a giant jump in budget size, it's more of a shift in where you spend.

Super? So I'm curious to ask, what impact, if any, might blockchain have on the traceability of all these agentic decisions that are getting, uh, that are getting made? And and how are you seeing that play out? And either Charlie or MC would love your perspective here.

I can jump in because we've had some debate back and forth I internally on the role of blockchain within security. Um, in fact, you know, our our tech trends article will highlight the role of blockchain in the traceability. Um, I haven't seen it fully deployed in the re in in in an effective way, to be frank, and I and I don't always feel fully comfortable within there. But I do think that there needs to be increased o focus on traceability around potential decisions, especially with high risk data and critical business processes.

And thinking about the potential role that blockchain and the ledgers could play within there actually does help, um, a little bit. I still think it's at its nascent stage and and haven't fully seen this deployed outside of maybe some of the more advanced organizations.

Okay, and I'll I. I'd also like to just let our listeners know that we'll make sure to put a link in the show notes to that tech trends article that you just mentioned, Charlie. So this, you know, all of this is not just an issue for the organizations themselves. Most organizations are also dependent on their supply chains, and the increased risks that we've been discussing would also apply throughout one's supply chain. So what do you do if your key supplier goes down as a result of a breach?

How should organizations be thinking about ensuring their security and stability, not only of themselves, but of the companies that supply them?

Like, look, it's no longer third party risk is no longer compliance or check the box, um, element. And if you're only thinking about it that way, you're probably exposing an organization to a much broader set of risks, right? And you really should be asking the question, can we continue serving our customers, operating as a business if something fails? Two or three layers below that vendor, or what is the broader ecosystem connection we have because of that vendor, and how should we think about their security?

And this really does go back to my fa- you know, my favorite story on ice cream sandwiches and the impact that, you know, and and again, this goes back to we s. We did a war game, a crisis war game example, uh, at the an International Dairy Farmer Uh Association digital conference. And as part of this, we got to travel to, um, a Blue Bunny factory to see how they make all of their various products, from ice cream cakes to bomb pops to, uh, ice cream sandwiches.

And it was honestly one of the most fascinating experiences of of my life. As someone who operates in the logical world a lot, I love seeing the physical side and how they get tied together. And and one of the things we talked about was just how simple the products were for ice cream sandwiches, but how right? And again, an ice cream sandwich is really a wrapper, two wafers and ice cream in the center, and the ice cream gets hard because the wafer absorbs the moisture and the wafer then gets soft because of it.

And that's why you get to enjoy, right? But there aren't many vendors of and makers of those wafers because there are very few requirements that go in there. And so if one of those wafer makers goes down, you lose your ability to provide ice cream sandwiches. Now, the middle of February, that matters much less than in the middle of July, so timing matters. But ultimately, how have you thought about resilience in your infrastructure for your wafers?

If you make ice cream sandwiches, or for processing payments, um, if you are a bank, or for delivering your core sets of paper and reams for schools when they start, right? And so how do you think about all of these matter? And so traditional vendor risk management is always just thought about that one component, and they think about the direct supplier, and so they try to spread out at that third party layer and create sort of vendor diversification.

But that doesn't necessarily create dependency diversification if you're all relying on the same fourth party or fifth party within there. And so what we have found is that several apparently independent vendors can may depend on the same cloud provider, identity service, data platform, or some other foundational infrastructure. That creates these hidden single points of failure that you need to assess and understand. You know, the potential you have for correlated failures across these suppliers.

Charlie, that's a really good point about this notion of shared points of failure or risk, even among a diverse set of suppliers. So how do you help your clients track all those dependencies that, you know, the fourth or fifth party links way down the chain, and and how do they mitigate them? I mean, this seems like a really big undertaking.

You know, there is an uncomfortable truth here, too, because I'm sure people are thinking, Well, how am I gonna figure that out, right? And it, it's, it's actually a bit uncomfortable. You may actually never have full visibility into every one of your vendors vendor, and then chasing that perfect visibility is not the goal that you want, right? It is actually the wrong goal. You want to just really sort of understand within your. What are my critical business services?

What is my minimum viable company, minimum viable business? What do I depend on for my own internal operations? And then how do I think about shifting into? Where do I go? What am I looking? Maybe from a region standpoint, what are the potential impacts there? And then how do I think about artificial intelligence and the role of artificial intelligence now? Models? Are we bringing in open weight models? What are the types of data sources and what is the resilience that they've built in there, too?

If they have an outage and their operations r rely on a lot of agents, right? How do you get the agents to show up to work if there's an outage in there, right? People can still potentially go to manual can agents, yes or no? And that's sort of a core question that you should have and start thinking about broader dependencies. It's not an easily solvable problem, but it is a problem to at least identify where that some of that core risk could come from.

And for you to start controlling against what you known, what you know, um, and just recognize that you'll never have a perfect answer.

Thanks, Charlie.

And go buy and go buy some ice cream sandwiches.

I always love a good ice cream sandwich, and it's a nice way to highlight these dependencies. Thank you. So, as we dig deeper into these dependencies and and potential failure points, sooner or later, there is going to be a breach, right? And when that breach happens, one of the things that a CEO would probably wanna be able to think about is, you know, what's the resilience of their company? Are they set up to weather the storm?

So, MC, can you talk a little bit about response to breaches and what distinguishes the companies and organizations that are able to handle such intrusions well versus those that don't, and those that recover well, for example, from those that don't?

Yeah, there are a lot of things that organizations can do to make it so that, uh, a breach becomes a bad week rather than a a bad decade, right? Or defining crisis for their organization. And and you know, we see lots of examples of both in in in the world, and and we're gonna see more and more of them. Um, the leading organizations that react well essentially have just pre-made a lot of decisions. They've prepared well.

This is not news. I live in the Bay Area. Earthquake preparedness is critical. Everyone has a plan. This it's the same for this. We need to, we need to have a plan. We need to know who's gonna declare whether there's an incident, what the threshold is, who calls regulators, who calls customers, who talks to the press. Actually, a lot of the failures after a breach are not necessarily technology failures, um, they are communication failures that make those breaches a lot worse for customers and for employees.

A lot of times, organizations under communicate at first, then over communicate at exactly the wrong moment, and both of those patterns can erode trust really fast and faster than the actual incident. The companies that have mature incident response capabilities, they they plan ahead and they end up spending less on recovery. It's spending up front, um, to then spend a lot less when you actually have an incident. And, you know, and it essentially, if you can react within twenty four hours to a breach, you can prevent cost growth from being exponential and really kind of letting that breach define your company.

We, we really insist on advanced planning and, uh, hopefully everyone is working on kind of their their incident response plan as as we speak, or has them already.

Thanks, MC. And so now I'd like to switch gears a little bit and move from agents to quantum, which, you know, many have referred to as like a master key for encryption. Charlie, where really are we on quantum? You know, Q day, or when quantum actually happens, it could have massive implications for security. So how, how does one prepare for that in the broader context of what we're talking about today?

Yeah, I I mean, I think a lot of the foundations still matter a lot, right? And as Sean, you know, we've talked about quantum, quantum a bunch. And and I think there's a couple things right. And first, I just really wanna highlight that this is really like a long horizon, sort of balance sheet type risk here, right? It's. It's not an urgent fire, but it's something that gets more expensive to fix the longer you wait, right?

Or that, like, you don't, you don't prepare for, right? And you have to start thinking about it. I do wanna highlight, no one can actually give a precise sort of Q day, right? The physics in February said twenty thirty five, right? And like, I'm one of those people that says, like, you know, you can't fight physics. But since that that came out, it seems that maybe robots can, right, because both, you know, there there have been some companies that have highlighted readiness by 2029, right?

So six years before then, so somewhere within the next decade or so. Which for most of the security folks, they'll probably go, Charlie, like, you know, I've been doing security for 20 years. And it's been within the next decade for the past two. When is it? And I go well, it. I think we have a better number, a tighter number within there. And what we know now is that organizations shouldn't plan around that, right? We think it will happen sometime within the next five to ten years.

We think this is going to happen one from the physics, but two, we've actually seen accelerated investment both into quantum technology, but also within quantum security. We are seeing increased standards development within the United States government, within, uh, the EU and broader regulatory movement here. So they're, you know, the really smart minds are seeing something that is trying to create an impetus and force a bit of a move. You know that in some of the large institutions, as well as governments, are moving towards post quantum cry- cryptography.

So we think of a few things here, right? So what, what do you need to do? One is obviously there's sort of the harvest now, decrypt later perspective. And so what does this mean? That encrypted information stolen today could remain valuable long enough to be decrypted later, right? So get better data security, understand your data, protect it, and if data is breached, understand what was breached. Does it have that long ten year lifespan, or does it shift?

And you don't necessarily need it, right? Is it a product that was is already in market that you've already lost the, you know, the patent on? Or we already passed that? Doesn't matter, et cetera, et cetera, right? And so now you have to start taking, we know that as a risk. What is gonna happen once you get to this? And it's how do you think about transitioning? And so one, large cryptographic migrations can take years, right?

So you need to start thinking about what do I s apply post quantum cryptography to? How do I think about readying myself for that? That seems to be the faster. And if I am in the type of organization that needs to reinvest and invest heavily within QKD or quantum key distribution, where am I going to do that? What is the lifespan of that technology, and is it worth investing now, or do I have to wait? And how do I prioritize all of that now?

And then start thinking about your broad credentials and all of your certificates and starting to map that and build the foundations that you should have anyway.

Okay, so does quantum just raise the bar then for the ways that companies need to execute their security, for example, just making sure that encrypted data stays within their organizations? Or is this really changing the fundamental landscape of security?

I think it raises the bar. I'm a firm believer in a few things, right? Number one is position improvement, right? And and this isn't about NIST scores anymore, right? I think we're past what is a maturity score. I think compliance is no longer the bare minimum. You probably have to be a bit more than that. It's about really one understanding, do I have the right control? One? Do I understand what my risks are, right? CIA confidentiality, integrity, availability is it are the core basics.

You throw in tech delivery on there too, because I think this matters and we don't think about it in cyber, but we should, right? So number one, it's how do I understand what are my risks? Two, what are the core capabilities that I need to have, and are my control objectives designed in a way that keeps me within my risk appetite? Do I have those controls deployed at the right level where I need them? So is my coverage correct?

And then how are my controls then collectively effective? And you should always be looking at continuous position improvement over time. And how am I continuing to mature my capabilities? How am I continuing to adapt to new technologies? And part of this is making sure you've got the foundations in place that allow you to adapt to the requirements that you may have within quantum. I don't think it fundamentally changes what you do, it changes how you encrypt your data, right?

It changes the requirements you need to have for managing the effectiveness of your data. And you should be ready for the potential harvest now, decrypt later and recognize what could go wrong, and those are foundations today. But this is just part of the continuous position improvement. And unlike what we started about at the beginning, where we talked about frontier AI, right, which happened almost like a tornado and just popped up, right? My view from like a s a storm or an earthquake that could just pop up, right?

My view is this is more like a hurricane. We know it's coming, we have a general idea that it's gonna hit during a certain window, and we have the time to start preparing. So make the preparations now to improve your position here. And that is the one nice thing. I don't see it as a as a fundamental shift, I see it as a shift that will occur. And we as security folks would be remiss if we didn't start thinking about planning and actually making the changes today.

Understood, thank you. So, given the risks we've been discussing and how they transcend industries and supply chains, what's the role of global cooperation, then, across industries or geographies, or even countries, to better address some of these risks? MC, you do a lot of work with financial institutions, for example. How are you? Are you seeing that happen there? Are are where industry associations maybe are trying to figure out how to protect against these risks and then sharing those insights with other organizations in their sector.

Or is this something that's still nascent?

I think I think collaboration is essential here. Um, and and frankly, you know, th- this is something. This is one of the risks that touches every single industry as a leader in risk and resilience. There's a lot of risks that are specific to industries and that we, you know, we need to be specialized in a sector, but this is one that touches everyone. We need everyone to be learning about this, um, and addressing it in concert, and I think learnings can be shared.

And this is one where talking to peers is incredibly valuable. Improving your position becomes a competitive advantage, but if you're sharing across industries, some of these advantages can be can be shared. And um, and improve security for for the world, right? In in the end, it's it's not just a business, uh, business question or business problem, it is a societal issue. Um, and we need everyone to be addressing it because otherwise customers are people and people will will see consequences.

Indeed, and as our listeners are thinking about their overall security strategy, then, and and you know how to approach this from an agentic perspective, how do you recommend that they think about which frontier lab to work with? Are there any frameworks that you've, um, developed in your work with clients to help them think about what model and with whom to work? Uh, to increase their security and and resilience, uh, against bad actors using the same kinds of technologies.

All right, I think like the core question is a little bit of all right. And I think it's one. What are the frontier labs that I work with to run scanning? Where do I start the scanning, right? But you also need to think about where do I roll with some potential open weights. Am I bringing in open weight models? How am I going to use them? How do I bring them? Because from a token standpoint, it can get very costly very quickly, right?

And so when we work with clients, I'd say, you know, there's broadly sort of three elements that we bring in, right? So number one is how do? And this gets back to what MC said, how do I think about my core operating model and my processes? Agentifying a broken vulnerability management to patch remediation process is just going to give you increased costs on a broken process with limited risk reduction, right? And a faster response. And so how do I think about bringing all of, you know, getting those processes right, restructuring them in in a way that allows me to operate effectively?

The second way we then start working with folks is how do I actually then d deploy this? What is sort of that bounded workflow that I start working again? Small slices, not the entire program, but how do I start running a proof of value in a narrow spot across a vulnerability scan that allows me to then build out? This requires improved. Um, how do I think about my prompting? What is the effectiveness of my harness? And look, sometimes a really good harness and really strong prompting on a lower capable model will give you a better output than average harnesses and average prompting on the best, right?

And sometimes finding vulnerabilities is different than the need to find an exploit, and so you can operate on smaller. And then this gets to sort of the third part we think about on. This is the tokenomics, right? And structuring. How do I run the right scans? How do I use the right model? And what is the balance from a token cost that I need to bring in? And what is that structure? That helps set the foundation for when you think about the future of agents.

And how do I use it from threat intel feeds into my unified vulnerability program? How do I think about building an agentic sock? What do I do from a governance standpoint? And over time, you'll, you'll get there. But that upfront vulnerability scanning and that operating model change is where we're seeing the most impactful use across the majority of our clients. Some more advanced organizations, uh, the large banks, the technology companies, are way ahead of this, right?

But most organizations, the majority of organ organizations, that's where they should focus.

Thanks, Charlie and MC. Anything you'd like to add there?

And maybe just reinforce that this is not about selecting a technology vendor, right? Uh, this, I I think throughout this whole call, we've, we've talked about how this is not just a technology issue. Everything around the technology is changing as well. And so, you know, picking every technology vendor is is, you know, perhaps the right next step, uh. But then how you manage those technology vendors, how your organization works with them, how you shift decision rights, how you increase the s clock speed of your organization, all of these things need to change.

And the technology, a technology vendor will not, will not do all of these things for you. This is a business, business question and a business problem to solve.

Okay, so maybe Charlie, you could take us through a little bit of what it takes then to get this right, and what are you seeing out in the marketplace and with the clients that you're working with?

Yeah, and I I think it reinforces a bit of what MC said. But from like the very technical component, the we're having conversations, obviously on the frontier AI, but we're spending a lot of time too on the security foundations and the fundamentals, right? We highlighted, and I don't wanna spend time on third party and and vendor security, but this is coming up a lot. Identity and access management transformations, whether it is thinking about what is your identity governance program, how are you thinking about your privileged access, right?

And what is, um, just in time privilege? How are you thinking about restructuring your broad move, right? Like the privilege game is changing a lot. And this becomes core conversations, both, by the way, with CISOs and CIOs, but also actually with in investors. Our data protection and cloud security conversations are continuing too, right? And so we see some of these, you know, cloud workload protection gaps that were okay before, are now not okay before, but they weren't a big issue, are becoming much larger.

But in reality, for me, is getting into data protection plus IT asset management plus a CMDB, right? If you don't know where your assets are, it's really hard to secure them. If you don't understand what your criticality of your data is, and you go to do an AI rollout, that becomes very hard as well. Even if you've escalated all of the privileges, right, it becomes well, who's downloaded a spreadsheet onto the from a SharePoint? And what does this mean if they're using it?

How do you manage the change behind it? How do you track all of this, right? And problem is that becomes a big, like 100 million piece of of sort of that type of technology. If you think about all of the spreadsheets within the organization that agents are gonna use that you don't normally track. And then finally, look again, zero trust, United States government has been pushing on zero trust, right? That was where the majority of the conversation was over the past four years.

We see that even more now, right? And that being a driving force to some of these fundamentals up top. And then finally, a lot of the conversations is organizations want to push AI to the SOC, because they see it one is faster response, but two, cost cutting. And again, I come back is let's take a step back, let's make sure that your vulnerability process works first, and let's build that in and integrate it, and then let's build those skills within your tier one, tier two, uh, sock response.

After we get it right in your vulnerability scanning,

Thank you, Charlie. And just as we wrap up, maybe you could share a high level perspective on where organizations go from here. MC

I think, you know, the things that you start with are some of the simple sounding, but very hard things, frankly. Making sure you have ownership and decision rights figured out, who owns AI and cyber risk, how you know defining how AI can act autonomously versus not establishing escalation thresholds, you know, making sure you have visibility and control across your entire environment. N- n- like Charlie said, know where your assets are data, identities, AI agents, um, you know, limits to access and addressing, you know, ability to address critical exposures, um, and think about how you redesign the way the people and AI work together, right?

And h- what that means for automatable tasks, uh, but also h- human review and training the people to, um, to actually do the human review. What would you add, Charlie?

you know, I think one of the things to get on top of, because I think you covered everything, MC, uh, that I would, I would typically say, I think there, in my mind, there's like, uh, probably two things that I see. The some of the biggest gaps right now from security is one is if you're a finance person, if someone from your CFO's office or your budget office is not sitting with you in your core meetings and you're planning from here.

I mean, you're in a budget cycle likely right now, but you need to have finance sit with you, right? Because everyone is looking to cut costs and they think AI can cut costs, but there has to be an increased spend. So the faster you get within in in there matters, and then flexibility on what comes next. We really don't know what's gonna come next. All we know is that offense is gonna be way ahead for a period of time, a year, maybe two years.

And we know that open weight models will start catching up to where, uh, the frontier labs are today. And so you need to be ready for that, right? But if you don't have the right funding, if you don't have the right talent and you're not having those conversations today, it'll be really hard to keep up.

Thank you both, really appreciate the conversation. And now I'd like to try and end us on a more positive note.

Good luck.

So, so you've talked a lot about the threats and how they're moving faster than ever and and what companies need to do about them. But let's spend a few minutes now just to talk a little bit about the opportunities. What opportunity are you each most excited about in terms of what these frontier labs potentially represent for a company that's being forward thinking about building in security to address this new context? Charlie, do you wanna go first?

Yeah, I mean, I mean, for me, it gets back to the same thing that makes me excited as a consultant. Like, we've got really great talent and they're able to do a lot more, a lot faster when they take advantage of artificial intelligence. And I think it's the exact same thing from a security program tier one response access reviews from an identity and access management. Like, those are all things that require a ton of time, are very labor intensive, but don't require a ton of skills or training.

And makes it very hard to retain talent who wants to do a lot more, right? So for me, AI can knock out some of those lower level tasks and then really build the security program that you want going forward. Um, in a way that has advanced talent that takes advantage of all of the tools and ultimately shifts the front from everyone being on their back foot, being ready for the offense coming in to starting to push out and really build that security perimeter and making it stronger than it ever has.

And that's what excites me. I I. I find this actually a great time to be S in security because you get to be creative, you get to push, and everyone is like, when was the last time people actually thought about security this much? I don't know, right? You know, it's always like, you know, so for me, that's that's what's exciting. That's how people. What I think people should take away is how do you get the excitement in the organization behind this?

Thanks, Charlie and MC. What are your thoughts here? What has you really excited about this?

I think I'd summarize it in, you know, never waste a good crisis, right? I think this is an opportunity for everyone to get better at their, at their game. Like Charlie said, cyber talent has been really hard to find, and and I think that's part of the reason why cybersecurity is is hard to, uh, maintain. Because because we just don't have enough of the amazing people that are needed to to do this. And so AI is an opportunity to get ahead of the game and to really kind of make the most of that critical cyber talent.

If we all jump on the opportunity, I think we can have a much better, uh, system in the end.

Awesome Charlie MC, thanks so much for taking the time with us today.

Thank you.

Thank you.

And thank you to all of our listeners for joining us. We hope you enjoyed the conversation and we welcome your feedback and ideas for future episodes. Our email is itsr at mckinsey dot com, which stands for inside the strategy room. You can also share your ratings and reviews on any podcast player, with many thanks to all who've already done so. We really appreciate the comments and feedback we receive every week and encourage you to please keep them coming.

And if you enjoyed this episode and you'd like to subscribe, you can easily follow our weekly series on any podcast player, where you can also access our entire library of more than 320 episodes. Finally, if you'd like to automatically receive our latest publications and insights, we encourage you to visit our risk and resilience practice at mckinsey dot com slash risk. You can also connect with us on LinkedIn, and we've included those links for you in the show notes today.

Thanks again for listening. We look forward to having you join us again next week inside the strategy room.