#122 - Doug Merritt: Cybersecurity Risk in the AI Era

29:54And with what we've done with cloud, Like, the internet is woven through cloud. Every critical service that you use for an Amazon, Google, or AWS is directly internet addressable. You know, when we go out to open AI and back, we're going across the internet. And all those systems are woven together to serve a corporation. So that is your, like, the walls are gone. And the walls are gone and good guys and good humans and bad humans can wander through that castle at any point. So you've got to really change your thinking.

30:32Doug Merritt:We talked earlier about the good side of the liberalization of information and data and the advancement of humans. But now we've shifted into the bad side of it where that same opportunity and openness can create significant issues. Yes. And you can't have – when you invent something, it's the good and the bad are two sides of the same coin. What is it? I'm sure that most of your listeners will have heard of Anthropics Mythos, the latest offering that they've held for a couple of weeks. And there's so much discussion right now about, like, how do we contain this? So, Mythos, it's just the latest model, and it's a huge step forward in capability versus their last model, which was just released in January, which is incredibly powerful.

31:29Opus 4.6 is amazing, and it has released Opus 4.7. But Mythos is way, way, way more powerful in many different dimensions, and one of those is its coding capability, its ability to understand the language of software development and coding. And there are two sides of that. What do we generally use it for? We use it to produce more code of higher quality, hopefully with less security vulnerabilities way more quickly. But the flip side of that, that's an inherent property of these engines, is it can find vulnerabilities within that code. It can find ways to abuse that code in one way or another.

32:08And it's extraordinarily difficult to put guardrails around that, like trying to understand what is a good coding behavior and a bad coding behavior. They try. The commercial engines try. There are some patterns that you can detect that would indicate that it's a bad set of activities. but it's an inherent property of these models. And Anthropic and OpenAI and others already have their next two or three models being queued up that are way more powerful than these current models. And the open source community, which is where the entire transformer paper was birthed, it was an open paper for the world to operate against.

32:45The open source community is progressing very rapidly. And just the gap between what Anthropic or OpenAI are doing with their models and what the open source models are doing continues to close because you've got thousands of people globally working on those open models as well. So we have this capability. It is here amongst us. And the negative side of all the incredible coding and software productivity that unleashes is it also unleashes the ability for less sophisticated people to find weaknesses and vulnerabilities in code. So we're taking an attacker base that once was a couple thousand. You had to be very sophisticated to do some of these difficult attacks.

33:30And we're expanding it very quickly. Like some of the vulnerabilities that Mythos found had zero human assistance. And they were found in hours. It costs compute cycles. So there's a cost. Like you got to pay tokens for that. But the sophistication on the attacker side to do really interesting things, we're lowering that bar pretty quickly so that non-sophisticated attackers can look like very sophisticated attackers now, which we'll go back to my earlier statement of if that's the case, if most of the software ever written, we're going to be able to find vulnerabilities that no one's even thought of across that code.

34:15and we've liberalized who has access these engines just through the open source arena, then we have to have a remediative strategy. And I've tried to attack this every way I can for a year and a half on what are, you know, there's many. But I think the most powerful and immediately impactful is you've got to contain the blast radius. Like the more that you can put guardrails around every workload, every piece of software, so you can limit the progression of these attackers, then that's how you can make your environment the safest.

34:52Doug Merritt:So most organizations tend to approach cybersecurity through compliance and governance. What does that approach get right and where does it fall short? It is a really important element of cyber. You've got to be able to, one, humans are the greatest weakness, a lot of compliance and governance is, are you doing the right things within your team? So you reduce the incidences of bad code or attacks being successful within your world. Governance and compliance is different than the actual act itself, though. So I've been a big governance and compliance person through my career. Many of the software products that my teams have written or that we roll out are very strong on compliance and governance, because you have to be able to see the landscape.

35:40You've got to understand, you know, what is the equation of my landscape? Where does it rate on a scale of zero to 10 as far as healthy or not healthy? And am I getting healthier or am I getting less healthy? And all that is really, really important to give teams and boards of directors and CEOs and customers confidence that I should be doing business with this entity or not. But you don't get to be cyber resilient through compliance and governance alone. You could be compliant with every standard out there and still have a breach, have a pretty cataclysmic breach.

36:15Doug Merritt:And given that, how should boards think about cyber insurance, personal exposure, D &O considerations without becoming overly reactive? For most organizations, I think it's now mandatory, which has been a boon to the cyber insurance industry, that one, you've got cyber insurance and that you've got health ineffective D &O insurance. Given what we're seeing in the cyber landscape, I think that the cyber insurers really, really, really need to take a hard look at the products they're offering and how much exposure they actually have. Because the attack surface is growing both by the amount of code that's being added, but by the progression of the light of the LLMs and the amount of attack surface that's out there.

36:59And I would advocate super strongly to cyber insurance companies that they really start to understand containment as a philosophy and start to bake containment into part of the policies that they adopt. The number one thing that I advocate to boards and CEOs is I believe strongly that the chief information security officer should be elevated in the org, not because I serve that group, but because understanding, it's got to be the right CISO, it's called CISO, chief information security officer. You have to have the right to understand business and business risk and is able to translate the deep technical understanding they have into what does it mean for the company.

37:40Nothing is risk-free. But the closer that is to the CEO, I think the more impact they can have across the organization. And again, you can't be secure at the cost of business. You've got to drive your business, but you've got to have business that is secure as you drive it. and the D &O insurance has to extend to the CISO. The personal liability, some of these lawsuits have come out in the past four, five, six years that tags the CISO with both personal financial liability and personal criminal liability. I get the orientation around that. There's a healthy balance. They don't control everything, but they need to both have a feeling that their voice matters.

38:26There's obligation on them to be able to translate their voice into business terms, and they need to feel that they're protected as they're moving forward to try and protect the organization.

38:36Doug Merritt:At a high level, what are the main categories of cyber attacks today, and how should non-technical leaders think about them? Yes, another complexity in this landscape is there are a lot, there's so much technology, and there are many different avenues for attackers to utilize. The way that I think about it to advise cybersecurity teams and executives is, if you assume breach, then the most important element is, how do I stop a breach? And there are three, and it's called a runtime control in this weird tech speak, which is what can I invoke that actually kills the progression? And at the 10 ,000-foot level, our tech industry is really simple.

39:29You have actors. I would have said humans. But now you've got humans, agents. You've got entities that are consuming compute talking to entities consuming compute. So the three runtime controls, like every CISO that really wants to try and have a foundational level of cyber effectiveness are make sure you've got a strong and pervasive identity, both authorization but control capability. I've got to be everything that is acting has to have an identity and I have to have a way to kill that identity if I see something weird happening. That's one control. The second is everyone's not as pervasive of what's called endpoint, of an endpoint strategy as possible.

40:18Those are agents that sit around the compute. There are agents I can put onto my laptop or there are agents I can insert into what's called stateful workloads in a data center. Because if I see my iPhone, if someone sees my iPhone behaving weird, I can shut down my iPhone. The third that tends to be underplayed right now is network. Our entire world is connected and communicating. That's a critical runtime control. If I fail on the identity, someone is doing something aberrant that I have not caught, then my final element is cut off the channel, which goes back to that containment. Like if I can get into my entryway of my house, if I've got a way of keeping them in the entryway of my house, that's another runtime or remediative control.

41:08And I think going back to kind of first principles and foundational elements, like you have to have visibility, monitoring, controls, and confidence around those three core capabilities to then be able to build other elements of your cyber strategy.

41:26Doug Merritt:And for individuals, how should people think about identity theft differently than they do today? Hopefully. So one, be really thoughtful before you accept a network connection from anything. That is the number one path in or out to cause any type of harm. And people just blindly click it. I'm okay to accept this message or the network is actually pretty foundational to each of us as individuals. always have multi-factor. Everyone talks about complex passwords, but you have to have multiple different ways before, which is friction and a pain in the butt to try and protect the activities that you have.

42:13And everywhere that I possibly can, not every service allows this, but rather than use something like your cell phone as a multi-factor element, you try and use a third-party authenticator app. It is not that difficult for cyber criminals. All of our cell phones are on the dark web. All of our identity is on the dark web, sadly. And it is not as difficult as people think to spoof and pretend like you, like a third party, actually controls your cell. So when you get a text code sent over your cell, bad guy can intercept that, and it goes to them instead of you. So, you know, difficult passwords for sure.

42:57Get a password managers that can be as complex and crazy as possible. Use authenticator apps as your multi-factor wherever possible. Those are very difficult for bad guys to commandeer. And then just be aware of like, how many websites are you on? How many apps do you have? Do you actually need those apps? And it's very hard for most of us to see what are all those things linked to? going back to containment. But I think if you really treat your interactions with difficult passwords and true multi-factor, I think that we are in much better shape. Yeah. Don't go around opening all the doors and leaving them open.

43:33Doug Merritt:Yeah, exactly. So AI is empowering both attackers and defenders. Over the next few years, who do you think has the edge? Ultimately, I mean, the difficulty with defense is all you need is one failure for an attacker to be successful, where the attackers just need one opening. So there's an asymmetry between the attackers and the defenders. I think that defenders can be successful long-term as long as, one, we aggressively adopt AI on the offensive side or the defender side, I guess. But we really, really have to understand our landscapes and how do you put as many gates as possible in to make it much more difficult for the attackers to do their work.

44:29And I'm a broken record on this that I've said too many times in this podcast already, but I think the network is the most important channel that people have overlooked. And it is one of the strongest tools that people have to try and remediate and contain the bad activities that the attackers have.

44:50Doug Merritt:If listeners could make one or two practical changes this week, either as executives or individuals, if there's any practical changes they can make to meaningfully reduce cyber risk, what would it be? I will go with the executive, just to count the business people. I think that the questions that all executives need to be asking their cyber teams is, do you have a concrete list that you have high confidence in of one of all the workloads that you're trying to manage across our technology estate? and which workloads have internet access? And are you clear and are you confident in what internet access those workloads have?

45:45If you can just get a confidence meter on, I know all the workloads in my estate. I know exactly which ones have direct access or indirect access to the internet. And for every workload that is talking to the internet, I've got high degree of confidence on exactly what sites they're allowed to talk to. And I know those sites are verifiable, certifiable, and necessary sites. I think you're way ahead. The easiest way to be secure is nothing talks to the internet. That doesn't work. Like our system, our way of life does not work. So then the pivot is, does it need to talk to the internet? And do you have a whitelist of the few sites that it really, really needs that it can reliably communicate with?

46:39I'll give you all an example of an attack that happened about three weeks ago. So there's a really important middleware, AI middleware capability that put out by this vendor called LightLLM. It's kind of the connectivity layer between everyone in your company that is querying AI and the AI that it has accessed. A group called Team PCP found a way to insert malware within the public version of that middleware. What was the point of that malware? it took all the credentials that this middleware had, which were very, very valuable credentials, and it encrypted them, and it put them out over an internet channel.

47:24Very normal. I got a detection capability would not have seen that as weird. But the channel I went out to was a bad guy site. That middleware should only be able to communicate with a handful of certified critical sites that that needs to do its job. Most people did not have that internet filtering, that whitelisting, that way it's called egress controls in our industry in place. And so their credentials were exfiltrated to bad guys. Those attacks do not have to be successful. That is a very basic hygiene, not as difficult to do. And you have to think from the start when you're creating these things, like what could ever happen?

48:09but the appropriate posture for any workload is denial. Like it shouldn't communicate to anyone and then add the things it should communicate to one at a time and you'll largely be safe. I mean, if they don't ever rode in, they don't ever rode out, it's hard to do a lot of damage.

48:29Doug Merritt:Yeah, rather than having all the doors open, they're closed and somebody knocks and then you check to make sure you want to let them in, right? Right, right. I'd like to ask you a few questions about your leadership philosophy, which I found very insightful. I know you've distilled it down to five core principles. Would you walk us through those? Yes. And it's taken a long time. I had a really long list that I think I got down to seven at Splunk and I got them down to five at Aviatrix. And they're prioritized in order because it's the way that I tend to go through them. The first is relentless curiosity.

49:06If going back to the beginning of our conversation, if you are not just innately curious, and you can screen for that, for people you bring into your company, I think it's very difficult to have that learning culture I talked about. The second is lead with empathy. If you're curious, but you're not willing to kind of step back and listen, and really try and digest the curiosity, then you're kind of failed on the curiosity piece. And the third is purpose before action, right?

49:35Doug Merritt:If you've been very curious and you're leading with empathy and you're really trying to put yourself in the shoes of the people that you're acting with, then you've got to take a few moments to really understand what is it that I'm trying to get done? Like, what are the first principles in the thing that I'm trying to attack? You know, what is the scale of the thing? And the fourth is radical accountability. Now, if you get those first three and you're convicted you want to do something, then hold yourself and others accountable to make sure that you, you know, are marching forward with incremental progress every day towards that goal.

50:06The last is celebrate success. And I've said it over and over, that's more for me than everybody else. I am horrible at ever patting myself on the back or taking a moment and reflecting on anything positive I've done. I'm always on to the next thing. But I feel like that's fuel that's necessary to keep that flywheel going of you're making incremental progress.

50:30Doug Merritt:And every incremental win is something to be celebrated. Which of those principles do you think is least common among senior leaders today? My guess is there's probably more celebration than maybe you're personally used to, but I'm curious about the other four. What I have struggled with the most with people in teams is purpose before action. And I think, again, I think it goes back to all the conditioning we've had over the past 15 or 20 years. there's a premium on action and we're there's so much benefit that we get from from clicking and from doing it feels good and motion is really important in life like life is all about motion and purpose for action advocates stop for a second like before you just respond to something before you do the default really ask like five more questions on what is the driving force behind the thing that I'm advocating and go as far down the first principles as you can before you just jump in and take action.

51:41And I'll give an example that I'm concerned about right now. With Mythos coming out from Anthropic, the immediate recommendations and responses was the tried and true path that we've been on, which is use mythos to find as many vulnerabilities as you possibly can as quickly as you can and remediate those vulnerabilities as quickly as you can. And that's not a bad thing, but that shouldn't be the first thing to do. The first thing to do, if you think about the situation should be, let me analyze my containment architecture. Like how big is the blast radius on my estate right now? And assuming that someone gets in and then not gonna be able to find every vulnerability and close every vulnerability, there's physics involved in organizational dynamics and remediating a vulnerability.

52:32What else can I be doing? But like we've been taught for 15, 20 years, vulnerability detection, vulnerability remediation is the most important thing. And that's where everybody immediately gravitated to. It's like, slow down, slow down. And we've got new information. We've expanded the attack audience from thousands to millions. Like, really? Do you think you're going to outrun them? Do you think you're going to catch every vulnerability? Do you think you're going to be able to remediate in time before they exploit those vulnerabilities? Like, what else could you be doing? And that one seems to be something we've gravitated away from the past 10 to 15 years.

53:08Doug Merritt:Yeah, what you said there was, I think, really insightful. And one observation that I've had is there are certain leaders who have the unbelievable ability to zoom out and see the direction that everybody's working towards. Whereas most people tend to be zoomed in because they're busy doing whatever job they're tasked with. And you can feel like you're making a lot of progress because you're working hard. But if you were to zoom out, you may realize you're working hard, heading in the wrong direction. So you're actually not making progress. You may feel like you are when you're zoomed in, but if you zoom out, you realize you need to shift, you know, five degrees to the right, and now you're heading in the right direction and you'll actually realize progress.

53:50Doug Merritt:So I think having that purpose is another way of describing what I just said. Does that make sense? You said that in a far more engaging way than I did, Alex. Yes, 100%. That's why I was trying to kick. Yeah, I appreciate that. Well, one other aspect of this that I think is related to what you said earlier is, and I've heard you say this before, there's no failure, only learning. How do leaders create that mindset without lowering the bar? The two most controversial statements I have is that and release the objective, like the objective, stop focusing on the objective. And both of those are so counter to, again, the prevailing approach.

54:32I'll gate both of those. Like ultimately, do I have outcomes that I want? Do I have objectives I want? Yes. And if I don't achieve those objectives for long enough, should I be fired 100 %? It's like what it's a sequencing, like what do you put first? What's the most important element? And if you're over rotate on the outcome, I think you're going to be disappointed more often than not versus over rotated on the finding the right path. Now going back to your discussion and then understand, am I making incremental progress on that path? So the way that I try and drive that internally is if what you're doing is something that is super standard and we've done it a hundred times and you don't do it properly, that's a legitimate failure.

55:22Like you did something, you weren't paying attention, you were lazy, you were cutting corners. Like those are not failures I want. Where that, where I'm trying to apply, there is no failure. There's only learning is we're trying to build something that hasn't been built yet. And, and we don't know what the answer is. Like we all got, we've, hopefully we've studied it. We've done purpose for action and we're going hard in a direction, but we've got to be willing to pivot continuously because it's an experiment. And that is where, you know, there is no failure. Even my failures, if I'm really honest, though, like they were the best learnings.

55:58They were when I tried to accomplish something and I wound up not getting, not being successful. That benefited me way more than continuously doing something with success, which is the other element I'm trying to capture is if life is all about learning and growth, it's a full contact sport. Get in. You're going to fall on your face a ton of times. that's the glory of life. Like if you don't enjoy that, then again, you probably won't be super curious. Like there you're going back to the whole piece of, of, uh, it's just, it's life is fun. If, if you're super learning and growth oriented. Yeah.

56:40Doug Merritt:I think the way you describe that is to me is very clear. There are certain destinations that have been reached a million times and there's a clearly defined path to get there. You do A and then you do B and you do C. Whereas if you're trying to achieve a destination where there is no clearly defined path, I don't think it's correct to assume you know what that path is going to be. So recognize that you may be going down the wrong path and failure may be just realizing it's the wrong path. And then you learn from that and then you adjust and you keep adjusting and you're focused on what the destination is and recognize that you may take many detours to get there.

57:20Doug Merritt:And that may, and as you're learning the right path, and then now you've created that right path for others. And now you move on to the next thing. Yep. And the detours are part of it. They're part of it. That's,

57:34I personally, and I know so many CEOs that have tried to recreate something they have from scratch, because it's ugly. Like it's got tech debt and it's got so many things that you learn along the way. And I've never actually seen one of those go well because the, as ugly as the thing is, it's beautiful because of all the detours that are in the product. It's, if you do something fresh, do something new, right? Recreating the old thing in a more perfect state is a hard path for most people.

58:12Doug Merritt:Yeah, and it's probably less beneficial for society as a whole than discovering a new destination and finding a good path to get there. Yep. Well, Doug, this has been a fascinating conversation. It was a lot of fun for me and I hope our listeners enjoyed it as much as I did. Thank you for joining us. Thank you for having me on, Alex. I had a lot of fun too. I really appreciate the opportunity to talk. Thanks for listening. We hope you enjoyed this episode. Please visit our website at insightfulinvestor.org to access past shows and learn more about our podcast. If you have questions, feel free to email us at info at insightfulinvestor.org.

58:52Doug Merritt:And if you enjoyed the discussion, please subscribe to this podcast to ensure you don't miss future episodes. And don't forget to forward today's conversation to others you think would enjoy listening. Important information. This podcast is provided for informational purposes only and should not be considered legal, tax, investment, or business advice. It is not a solicitation, recommendation, or endorsement. All opinions expressed by participants are their own and do not necessarily reflect the views of the Evoque Advisors Division of MAI Capital Management, LLC, or Evoque, its affiliates, or any companies mentioned.

59:26Doug Merritt:Information shared has not been independently verified by MAI or its affiliates. MAI Capital Management LLC, or MAI, is registered with the U.S. Securities and Exchange Commission, SEC, which does not imply any particular level of skill or training. Certain information contained herein has been obtained from third-party sources, and such information has not been independently verified. No representation, warranty, or undertaking expressed or implied is given to the accuracy or completeness of such information by any person. While such resources are believed to be reliable, Evoke does not assume any responsibility for the accuracy or completeness of such information.

1:00:02Doug Merritt:Evoke does not undertake any obligation to update the information contained herein as of any future date. The content is intended for a general audience and does not constitute a recommendation to buy or sell securities or adopt any investment strategy. Any examples or scenarios discussed are illustrative only, involve risks and uncertainties, and do not guarantee future results. Non-traditional assets carry significant risks and may not be suitable for all investors. Decisions should be based on individual objectives, risk tolerance, and circumstances. Statements herein are general and may not reflect an individual's or entity's specific circumstances or applicable laws, which vary by jurisdiction.

1:00:41Doug Merritt:Further, speakers' views are personal and may differ from evoke and MAI recommendations and are not specific investment advice, and do not consider client objectives, risk tolerance, and diversification. Guests may have current or past relationships with Evoke and MAI, its affiliates, or the host, including as clients, service providers, or business partners. Participation does not constitute an endorsement or testimonial. No compensation has been paid or received for guest participation unless disclosed. MAI and its affiliates may have business relationships with entities mentioned in this podcast, which could create potential conflicts of interest.

1:01:15Doug Merritt:These relationships may include advisory services, investment management, or other arrangements. MAI seeks to manage such conflicts consistent with its fiduciary obligations and policies.