Skip to content

HomeCyber Business Podcast

No Competing Priorities: Securing a County and Putting AI to Work with Eric Matthews - Ep 254

Eric Matthews IMAGEEric Matthews is the CISO for the County of Fresno, in the middle of California, where he works to protect the data of just over a million constituents and keep county services available. He leads a dedicated cybersecurity team that is not a subset of IT and reports directly to the county’s chief executive. Eric and Matthew discuss why local government infrastructure is a target, how a CISO makes the case for AI tools to executives, and why risk appetite and clear accountability matter when people use AI. They also compare frontier AI services with running open models in house. 

 

apple
spotify
stitcher
google podcast
Deezer
iheartradio
tunein
partner-share-lg

Here’s a glimpse of what you’ll learn: 

 

  • Why local infrastructure, including water systems, is a target for attackers
  • Why a security team that sits outside IT avoids competing priorities
  • Why Eric calls AI just another tool and not a replacement for the team
  • How a CISO can sell AI tools to executives with a value proposition and risk math
  • Why public sector security comes down to trust
  • How generative AI can draft a policy and check it against existing policies
  • Why users stay responsible for AI output and why Eric blames missing controls
  • When open models and in house hardware might make sense for a large organization

 

In this episode…

Eric explains that the County of Fresno sits in the middle of California and serves just over a million constituents, and that his role as CISO is to protect their data and keep county services up and running. He says local infrastructure is actively being attacked, with systems that run water purification hit quite often, and that local governments have traditionally not protected them as much as they should. His dedicated cybersecurity team is not a subset of IT and reports directly to the county’s chief executive, and it works with public works to inventory those systems and confirm their connections are safe. Eric says the biggest advantage is avoiding competing priorities or competing strategies with the IT strategy, since a CIO’s top priority is delivering IT solutions to departments, and a security leader inside IT often has to filter conversations through someone else. He says his team of four individual contributors and one security manager serves about 8,000 employees. He adds that a CISO must give executives the information to decide on people and tools, including the hard conversation that AI is scary but the bad guys are using it.

 

On AI, Eric calls it just another tool that must be understood and used correctly. Asked how to convince people that AI is necessary, he says the alternative is a security team of about a thousand people reading the 2.5 million emails that arrive every day. He is not looking to replace his team, since humans still make human decisions, but no person can analyze every event, so he wants a human in the decision tree with checks and balances in place. Asked how to bring leadership on board, he tells CISOs not to forget the chief officer in their title: give executives a value proposition tied to the mission, translate risk into likelihood and impact, and put a cost on it. In the public sector, he says, the profit is trust. He describes the conversation with county leadership that led to rolling out an AI based data loss prevention tool. For less sophisticated organizations, he says it starts with education, a few quick lessons on how to prompt, and use cases such as drafting a policy from bullet points and cross checking it against existing policies, which saves him days of work.

 

Eric says another use case is an AI security agent that checks proposed technical solutions against policy, which frees IT staff to be creative. He says he can usually tell an AI generated document by its grammar, and that he still reformats AI output himself. His biggest warning is against jumping into AI without security controls and business controls: Fresno has an acceptable use policy that bars using AI to make management decisions or evaluate personnel, and users can use AI but remain responsible for the output. He wants organizations to set a risk appetite for AI, accepting small errors such as a grammar mistake while keeping a human check on actions such as shutting down a network. AI is a tool like any other, he says, and when it makes a mistake he blames the lack of controls, not the AI. Turning to open models, he says he runs one at home and that an organization of Fresno’s size could probably break even or save money with its own, but only with the staff and infrastructure to support it and users who are comfortable with it. For an organization just dipping its toe into AI, he advises against building that infrastructure first. Eric closes by inviting listeners to find him on LinkedIn, saying he is open to talking, and Matthew thanks him.

 

 

 

Resources mentioned in this episode

 

Matthew Connor on LinkedIn
CyberLynx Website
Eric Matthews on LinkedIn
County of Fresno LinkedIn
Darktrace Website
Abnormal AI Website

 

Sponsor for this episode...

 

This episode is brought to you by CyberLynx.

CyberLynx is a Bethesda managed IT and cybersecurity company. Local techs you know, not a call center. Month-to-month. 24/7 intrusion detection.

We help growing companies with managed IT, help desk, backup and recovery, and a fractional CIO.

Talk to us at https://cyberlynx.com/contact, info@cyberlynx.com, or 301-798-9170.

 

Check out previous episodes:

The Summer of Copilot: Education, Context, and Thinking Strategically with Thomas Matthews - Ep 253

 

Train the Tool: Guardrails, Separation of Duties, and Responsible AI with Roberta Anderson - Ep 252


Defense in Depth for AI: Fundamentals, Tools, and People with James Hengsterman-Cash - Ep 251 


 

Transcript: 

Cyber Business Podcast

Guest: Eric Matthews

CISO of the County of Fresno

Host: Matthew Connor


Matthew Connor: Matthew Connor here, host of the Cyber Business Podcast. Today we're joined by Eric Matthews, CISO of the County of Fresno. Eric, welcome to the show.

Eric Matthews: Thank you for having me.

Matthew Connor: Thanks for joining us. Before we get too far in, a quick word from our sponsor.

Sponsor Break: CyberLynx. Hackers are using AI to conduct machine-speed attacks. Is your security keeping up? CyberLynx sells industry-leading AI-powered cybersecurity solutions that detect threats in real time, so you know about an attack before the damage is done, not after. Learn more at cyberlynx.com. That's cyberlynx.com.

Matthew Connor: And now back to our show. Eric, for those who aren't familiar, can you tell us about the County of Fresno and your role there as CISO?

Eric Matthews: Absolutely. The County of Fresno, we're right smack dab in the middle of California. We serve just over a million constituents with all sorts of services that you would expect a local government to provide. My role as the CISO is to protect the constituents' data and to make sure that those services stay up and running and available for all the people who need them.

Matthew Connor: I love it. It's really interesting, because I think as you look at counties, that usually entails critical infrastructure. In today's day and age, with the advent of AI and cyber threats, I don't want to say it's a hot button topic, but it's a really interesting one because there's so much to dissect there. And I don't think people are fully aware of all the forces at play when it comes to that critical infrastructure at the township, city, and county level. If you don't mind, I'd love to dive into it a little bit, keeping it high level, to walk people through it, because I think there's just a lot there to unpack.

Eric Matthews: Yeah. One of the most interesting things about being a CISO of a local government, especially one that's kind of medium to large size, is just the diversity of what we have to do. One of the things we have to protect is the local infrastructure, something that maybe other systems in the private sector don't really get to be fully exposed to. Anybody who's been paying attention to cybersecurity news knows that right now local infrastructure is actively being attacked by bad actors. It is, I'll say, traditional that maybe local governments are not protecting it as much as they should be, and that's what makes it an easy target. For example, data systems that run water purification systems are being hit quite often.

So it's very important, and fortunately for the County of Fresno, we do have a dedicated cybersecurity team that I'm in charge of. We are not a subset of IT. We report directly to the CEO of the county, so we can actually focus a lot on making sure that those infrastructures are secure and doing the things that we need to do. We work closely with, in this case, public works, to make sure those items are inventoried and that we know the connections to them are safe. The last thing we want to do as the county is have some critical infrastructure go down, especially if it's water or power or internet, anything like that. So, yeah, right now that's one of the big topics at the local government level.

Matthew Connor: Yeah. I think it's really great and speaks to the professionalism of Fresno that you were able to put in a cybersecurity team, dedicate that, and have a CISO. I think that's really fantastic. A lot of counties don't prioritize that. A lot of times it's tied in with IT, and the head of IT is also the head of security, which is all well and good, ish. But I think it leads to those challenges where you get a water treatment facility, and for the most part great, solid people, but they tend not to be the most technologically advanced. So there's that gap in their understanding of security. And you know what, it's always worked, so we leave it alone, nobody touches the data, and it's fine. But it's really not. It's a weak point in terms of security. So having that dedicated team to bridge the knowledge gap is critical. And I don't think it happens nearly as much as it should in most counties around the country.

Eric Matthews: Yeah, it really doesn't. And I will say, just for anybody listening, you don't have to have a dedicated cybersecurity team to have an effective security program in your organization. However, what I will say is that the biggest advantage to doing that is not having competing priorities or competing strategies with the IT strategy. That is why you want to be able to pull it out and have it report to the CEO or, worst case scenario, the CIO, if your organization has one. Because the CIO is not going to have security as the number one priority. Their number one priority is going to be making sure that the IT infrastructure and IT solutions are provided to the departments. At the County of Fresno, several years back, we had a consultant come in and do a big assessment, and one of the things they found was that we would be better off if we had a security program, [unclear], reporting to the CIO.

Fortunately, we have an executive leadership team who listened, and they take security very seriously. It's a high priority in the County of Fresno, and I'm incredibly fortunate with that. I've worked with other counterparts at other local governments who don't have it that way, and they do struggle to give the board and the executive leadership the information they need to make the right decisions. Tying it back to the infrastructure thing, public works' number one priority is not securing the water treatment plants. Their number one priority is going to be making sure that they're running, that they're working, that infrastructure is being provided to the local population. So it's very important for me, because it's obviously my priority to make sure that they are secure, to have a good working relationship with those departments so that we can talk at that high level.

In my opinion, when you have a security program that's within IT, typically the person in charge of security is not someone that's talking at the executive level, so they're not having the right conversations with the decision makers. Typically they're having to filter that through the CIO, and in some organizations the CIO doesn't even talk at the board level, which means you have to filter through someone else. So we're incredibly fortunate to have the program set up the way we have. I actually always thought I had a really small team, but I am a listener of your podcast, and I've heard other people talk about the size of their team, and I'm very fortunate. I have four dedicated individual contributors and one security manager, and we have about 8,000 employees. I always thought that was really small, but listening to your podcast and the other cities, I was like, I actually might have it pretty good.

Matthew Connor: Yeah, you're doing all right.

Eric Matthews: Right. And that has a lot to do with getting your organization to prioritize cybersecurity. The organization is going to put resources where it prioritizes, and the people that are in charge are not cybersecurity experts: the CEO, the board, the department heads. They are not cybersecurity experts. It's your job as the CISO to make sure they have the information they need so they can say, "You know what, we do need a few more people," or "You do need this tool." Especially in local government, they don't really like bringing in AI tools, but you have to have the hard conversation with them and say, "Look, I know AI is scary, but the bad guys are using AI. What are we supposed to do?"

Matthew Connor: Well, this is one of my favorite topics, as you probably well know, and I think it's really interesting. I'm a firm believer in fighting fire with fire, especially when we're talking about AI, because it's not even like bringing a knife to a gunfight. It does remind me of Indiana Jones, the sword and the... It's about that bad. But I think it's actually worse, because even if you focus on the fundamentals and you do all the things right, let's say you could do all the things right with traditional security tools, even if you were fully patched, we know how easy it is now to have zero-day exploits. It's kind of like, "Oh, I think that wall is very secure. It's great," and then somebody literally walks through it like a ghost and you're like, "Wait, how did you get in here? Everything was locked up." The reality is there's still a ghost, and you don't know they're in there because you weren't expecting them there, if you're not using AI.

I'm getting to a question here for you. We see it in tools like Darktrace Network, where you're using machine learning to watch everything that's going on on the network, and when something weird happens, somebody walks through a wall, it's like, "Oh, hold up, let me stop that. Call for an adult." Great. We've seen for years now bad actors lying in wait in networks for months, sometimes years at a time, slowly moving across the network, gaining access and information. The thought that somebody could think, "Well, if we just do our job right, we don't need AI to fight this," I don't understand how you can fight such an advanced adversary, even when they're untrained and unskilled, because they've got amazing weapons. So how do you convince people that, no, really, we have to be fighting fire with fire here? We have to use AI. It can't be something we're scared of.

Eric Matthews: Yeah. So the alternative is real simple. If you don't want to use AI, then just have a security team of about a thousand people reading the 2.5 million emails that are coming in every single day. AI is just a tool. No security team has ever said, "We really don't want to use EDR. We think it's too technical." It's a tool that you use. AI is just another tool that you use. Now, it's very important to understand that tool and make sure it's being used correctly, but the way we look at it is, again, it's just another tool. What tools can I give my team? I'm not looking to replace my team. I still need humans to make human decisions. But I need to make sure my team has the tools necessary to fight what they're fighting. We use CrowdStrike, and CrowdStrike has integrated AI, and there's no way you can have a human analyze every single event that comes through the tool. You have to have it, and it doesn't have to be generative AI. There are a thousand different types of AI. It can be machine learning. But it's very important to bring that in.

Anybody who thinks you don't need AI to do security just needs to look at the number of known exploitable vulnerabilities and the trend upward. These are vulnerabilities that just 6 or 12 months ago would have just stayed at a level eight [unclear] for a while, and now, the day after they come out, they're becoming CVEs. And if you do not have tools that are using AI to be able to... we use the word "reason." We know it's not reasoning like a human, but it is AI reasoning to help find things quickly. And that's the thing, it's the speed. That's why I said you can have an AI tool or a thousand people, because it's the speed. Look at how many log events you have in your SIEM. Are you going to go through every single one of those, or are you going to have AI help you find a pattern and then have a human look at it and make a decision? I'm not quite at the maturity level yet where I want AI shutting down my network if it thinks something bad is happening. I'm fine with it flagging one of my staff members to make that decision.

To me it's a no-brainer. I will say AI is a lot more mature than people think it is. We'll say it's not 100% yet, but there are ways you can easily put checks and balances in place. There are easy ways to put double-checking in place. Again, put a human in the decision-making tree. But I don't know how people are running security for a large organization without the use of some AI. It's insane to me. It really is.

Matthew Connor: I think the interesting part is you mentioned the maturity. Now it's nearly October 2026, and I think we're just coming up on about four years since ChatGPT hit the market. The improvement we've seen over the last six months has been just phenomenal. Even if we were to stay at the level we are today, which I know is not even realistic, that's not going to happen, I think we've hit this tipping point where the expectation is that AI is to be perfect. First of all, does that even exist, even in theory? Anything being perfect doesn't actually make sense. We know there's no such thing as perfect. We treat each other as fallible humans, and yet for some reason we expect the computer to be perfect.

So I think there's a level of acceptance or understanding that, hey, AI is really not that different from humans. If you look at how we teach it, the only difference is that one is really fast at processing information and putting things together. And we're not raising it like we raise kids, because we don't have the time. But think about it: are we any different in the way we learn and put together patterns and function? I'm not sure we're that much different in how we reason. Most of it is pattern recognition of things we were taught. Kids don't come out thinking about calculus. They're not thinking about philosophy and pondering the world's greatest problems. So we're really just running on autocomplete. We're not that much different. Our genetic code isn't really that much different from computer code. And is AI thinking that much differently than us, and vice versa, other than the fact that it's super fast? Anyway, I go on and on about that, but I think it's really interesting.

I think the psychology of people right now is in a funny place, because you've got plenty of people who are resistant, like "I don't want AI to take my job." We saw the same thing early on with the web. If you go back and look at some of the clips from the 90s where people were just trashing the internet, saying "It's so stupid, blah, blah, blah," making all kinds of ridiculous comments, I think we get a lot of that today too. We're in that funny place psychologically. The question is how do we help leadership make sure they are on board with it moving forward, and everybody else in the organization getting comfortable with the fact that all the crazy people out there are saying AI is going to take your job next year. And don't listen to Zuckerberg when he fires half of his people and then has to rehire them. Crazy, right? Can we all just agree that his grasp on reality is a little [unclear]? Anyway, I think how we communicate that to everybody is really important as security leaders. I want to finally get to a question: how do you go about communicating to leadership and to the rest of the organization to get them on board with the fact that we live in this AI world and it's not here to take your job?

Eric Matthews: Yeah. I'll start this answer by saying this: CISOs need not forget what the C and the O in their title mean. You are a chief officer. I understand that not all CISOs are treated as chief officers. I am fortunate enough to be treated as one in my organization. But going to your executive leadership, going to the board, going to the people who are going to cut that check, and just telling them, "Look at all this work we can get done with AI," they don't care. You have to give them a value proposition. You have to be able to tie it back to the goals of the organization. If you're in private industry and your ultimate goal is to make money, show them how bringing in AI can make them more money, or, in the case of security, prevent them from losing money. In local government, it's really important to say, "Look, we have this mission to serve the public. We have this mission to make sure everybody is treated fairly, that the roads don't have potholes, and that the police and sheriffs are keeping our streets safe and clean. How can I help us with this?"

So it's about taking the cybersecurity speak, and let's not forget the basics of cybersecurity, CIA: making sure the data is confidential, making sure there's integrity to that data, and making sure the data and the services are available. Tying that to the business mission and selling that to the executives, telling them, "Look, I can do the math and tell you the likelihood and impact, so the risk, of not having this AI tool look at our emails before they come through. This is the chance of it being ransomware that takes us down. If we don't have an EDR that uses AI and looks at the behavior, this is the chance." So tying that back, saying, "I'm not asking for this tool, maybe AI or not, but AI is usually the harder one to explain." Being able to say, "This tool is going to take our chance of this service going down from, let's say, a 5% chance to a 3% chance. Is that 2% worth the $150,000 that tool costs?" And it is. Would you rather stand in front of the public, in front of a microphone with the press, and say, "Sorry, we can't provide you the service today because we didn't want to spend $150,000 on this cheap tool"?

In the circles that I go in, I think the problem is that a lot of CISOs are very technical, and they try to sell it as a technical solution, but it's not. Everything you do as a CISO needs to tie back to the ultimate goal of your organization. You've got to bring out that chief officer part of your title and speak at that high level. For me, it was really simple. We are currently rolling out an AI-based DLP and then an actual EDR tool. I sat down with the CEO and said, "Based on the current budget that I have and the resources that we have, starting with the highest priority thing I have to do and listing down to the least priority, DLP and EDR just missed the mark. I just don't have enough money for that, for everything else that you want me to do. But here's the value of that. I don't remember off the top of my head how much the tool costs, but I can tell you it is much cheaper than the value you're getting from it." We had a conversation about what that means to the county. How does that help departments? How does that help the actual citizens?

After explaining the actual value and how it ties into the bigger picture, because it's not about "DLP will prevent someone from uploading sensitive data to Google Drive." It does that, but that's not why they want it. Why they want it is because now we're responsible for public data being exposed, or now we have to go through this whole legal thing because we were not responsible. Unlike private industry, in the public sector our profit is trust. We want citizens to trust us. The last thing you want is people not trusting your local government. Being able to tie that into trust made it a very easy decision for the CEO to say, "Yeah, that makes sense. Of course we want to do that."

But you also have to understand that if they say, "Thank you for the information, we still don't want to do that," that's fine. It's their decision to make. You document it. You have a risk register. Put that in your risk register, and if something bad happens, you might have to have a hard conversation down the road and say, "Hey, remember that tool I wanted? Do you think maybe we can get it now?"

Matthew Connor: That's right. I think you're right on a number of levels. When you talk about the fact that most CISOs got into it because they like technology, like most of us in the tech industry, we like tech, and so we tend to be a little geeky about it. It becomes challenging for those people who are so interested in tech to say, "Look at the cool stuff it does," and then the business says, "That's cool, but how does that relate to the business?" And being able to, as you say, tie that into the needs of the organization, how this plays into the bigger picture and helps us achieve our goal of greater trust. Security relates directly to financial and reputational damage. There's just no two ways about it. And I think people are getting better at that, at least understanding it as leaders. There was a large divide for a long time. It was just an expense. Now people understand that security is a necessity. Now it's a matter of how much.

I think what's really interesting is, I don't know if you caught the Blackstone president, I think Jonathan Gray, had this talk about a week or two ago. It was basically the annual update, but what I thought was really fascinating was he was talking about the ROI that these companies are getting on their tokens. The ROI is astronomical. It is not even close. You rewind a couple of years ago and everybody's like, "Oh, there's no ROI, it's not proven." And now Blackstone, with their massive portfolio of companies, can very definitively point out that it is something like, in some cases, 16x. It's ridiculous. It is an astronomical number. I can put in a dollar and get 16 back all day. I'd be borrowing money to do it, right? And that's what we're seeing throughout the entire industry, on every level: how do we leverage that as much as possible.

It's fascinating to hear that the numbers are clearly there when it comes to ROI. Now it's not even a question. So the question, for the less sophisticated organizations, is how do we help them understand that and leverage it so they are not left behind? Originally the line was always, "You're not going to lose your job to AI. You're going to lose your job to somebody who knows AI." OK, that's true. But I think the real challenge is how do we help organizations that aren't as sophisticated fully leverage AI as much as possible and start achieving some of that ROI. I think that's the challenge. And what do you think?

Eric Matthews: Yeah. So I do actually believe that, obviously, [unclear]. But I think when you're interviewing for a new position, knowing how to use AI is going to start being something that's asked, or asked to be proven. With that being said, it's not hard to learn how to use AI. It's not a four-year degree. It's a few quick lessons on how to properly prompt and have a conversation with AI. So it really starts with just education. There are organizations out there, like Gartner and Info-Tech, that have hundreds, if not thousands, of case studies of how any industry can use AI to get value out of it. Go to a less sophisticated organization and explain, "Hey, here's a use case specifically for your industry on how you can use AI," and it can be something as simple as, let's say, me. I'm a generative AI user to write policy, and that does not mean I'm going to go to ChatGPT, put in a prompt that says "write me a security policy," and walk away. But having to hand-write a 40-page policy and then review it and cross-check it with all the other policies we have to make sure it's not conflicting is very time consuming.

Being able to write a couple of pages of bullet points saying the policy needs to cover this, the policy needs to say this, I want to explicitly not have this in there, and then feeding it into ChatGPT and saying, "OK, make this look like a smart person wrote it," and it formats it for you, and then you feed in all your other policies and ask, "Does this conflict with the other policies? How does this format look?" It saves me days and days of work. So just learning how to use generative AI as a research tool, not something that's going to do the work of a human, but even internally, "Here are all my policies. Am I allowed to do X, Y, Z?" It can go back and say, "No, you're not, this policy says that." You know how long it would take a human to figure that out. Especially in a large organization like mine, where we have an infinite number of policies and regulations and procedures, to have one human memorize all of that is just ridiculous.

So it's about being able to educate those organizations with the use cases. Something that we're doing is trying to roll out AI usage in the county to be more enterprise so that we can be centrally controlled. One of the things we have to understand is that a department, let's say public works, they are not technical experts. They're not IT people, they're not AI experts. They are experts in what they do. If I went to public works and said, "How do you want to use AI?" they're going to say, "I don't know." So in your organization, whoever's in charge of that solution, and at the county it is a joint solution between my office and IT, it's important to go to that department and actually tell them, "This is how we think you can use it. Let's get some test pilots in there. Let me show you. Let me pull it up on the big screen and show you some really cool things you can do with that."

And that's just the generative AI part. We can't even touch agentic AI, which could infinitely increase the speed at which you get certain things done, and that's slightly more complicated on the back end than generative AI. But getting departments or organizations just to start to play with generative AI after they learn how to properly format a prompt, that's a big thing too. Once they see the value in that, you can say, "Hey, generative AI is not the only type of AI out there. There are these other tools we can put in," and it becomes a business decision after that. If you're in charge of an organization and it's your responsibility to make sure you're doing as much as you can with the limited resources you were given, once you start to see that I can help you deliver those solutions, the more you go, "OK, this is, again, just a tool, like any other tool that I'm going to give my staff." I let my staff search the internet for help. Why would I not have them ask AI for help?

But like we talked about before, a lot of people are afraid of it because it grew so fast and a lot of people don't understand how it works. There are things on the news about how AI was going to kill you and attacking [unclear] at all these places, and those are fringe cases. That is not the AI you are using in your organization. That is different. But trust me, you do not have access to Mythos, so don't worry about that stuff. And again, in a mature organization, you put controls around it. You think I'm going to roll out AI in my organization and not put controls around it? That's ridiculous. So it starts with educating people with the use cases, but that means your organization has to have an AI expert. In large organizations, there should be a centralized AI expert.

Matthew Connor: I think one of the things you hit on a little bit there, and I think it's really exciting, is that in larger organizations, and by larger I mean more than, let's say, a hundred, which is still super small, but now imagine, especially as you start getting into the thousands, the number of policies, procedures, and SOPs that exist throughout the entire organization, in every department and team. It's a lot. How great would it be for every employee to have their own agent that is dialed in on their policies, procedures, and SOPs? As they write email, as they make decisions, as they're onboarded and say, "Hey, this is my first day here. What do I need to know?" Instead of "read the SOPs, read the policy," it's "Can we just walk through this real fast?" So much better than having one senior person spend days, weeks, and months, because we know that typically it takes about six months for somebody to get up to speed on their job and become a productive member of the team, give or take.

Six months. How much better would that be if they could basically hit the ground running, because they've got an AI assistant that everything runs through, that says, "Oh, you know what, we can't do that because the SOP doesn't allow it. Have you considered...? Oh, this is the way it's done here, based on this policy and this procedure." "Oh, super cool, great." Or, "Hey, there's a gap in this. Can we look at it?" Your email flows through it. Everything does. Now it is an assistant. I think as organizations start using it more on that level and it becomes part of the normal work environment for a person, they'll say, "You know what? My AI assistant is awesome because it does all the heavy lifting for me. I get to think about my job and do the fun stuff, and it handles the tedious stuff. I've got somebody checking my work all the time, making sure it's great." I think that's where we need to get people to, where they've got the comfort of an AI assistant instead of an AI replacement in their head.

Eric Matthews: Yeah. And the six months it takes people to get onboarded is not because they are fully productive for six months trying to get onboarded. It's because they're sitting at their desks, waiting for the senior to get done doing what they need to do to come help you with what you're doing. So it's not helping anybody. Now, with that being said, obviously you and I are not advocating that you don't have a human help mentor someone as they come on. But kind of parallel to that, one of my first use cases, especially being the CISO, is that IT comes up with technical solutions, and sometimes they don't necessarily consider the security controls that have to be put into place. They either, I don't want to use the word "disregard," but they ignore that altogether, and then we have to come behind them and say, "No, no, no, back up, we have to do this," or they waste a lot of time trying to get the right person on the phone to answer the right questions.

If we had an AI agent or AI assistant that could say, "Hey, we want to build this solution for this customer, we want to do these things, what do I need to consider?" and that security agent says, "OK, we have these policies that you have to follow. There are some potential gaps in your controls. Here's something that could be a vulnerability." What that does is free the IT person to be creative, because now they have a very quick way to do a check and balance of "Oh, I'm going in the right direction with the solution," or "Oh, I need to change course." Because the other potential downside is that IT people, if they're constantly being called out, "Hey, you forgot this, you forgot this," are going to be less likely to come up with solutions anyway, and now you're not progressing forward. So it's just a quick, easy way to say, "Hey, I want to do this. Does this make sense?" And if you trained your agent well with your policies, it comes back and says, "Yeah, that's fine," or "Maybe you need to reach out," or "Maybe put this control in," or "1.6.3 says this." It's such a time saver, and it's freeing. That's what people don't understand. It really just frees you to be creative with your solutions.

Matthew Connor: Yeah. And I'm curious, do you think we are now, or soon to be, at the place where the standard for work is AI-assisted work, because that work product is so much more professional, so much faster, so much more thorough? It is night and day, the difference. I think it's interesting to see how many organizations still aren't working at that level, still manually going through stuff and haven't implemented it. But are we at that place where that should be the new standard, or are we still a little bit off before it's like, "No, that's the standard. I don't know why you're manually doing this, it looks like garbage compared to this beautiful spreadsheet that Copilot made. What are you doing? Yours looks like dog food"? That's amazing. So I think we should be at that place.

Eric Matthews: What I can tell you is that I can tell the difference between an AI-generated document and a non-AI-generated document, and it's usually grammar. Personally, I do find myself telling people, "Hey, can you just run that through AI real quick? Can you run that through Copilot real quick, just clean it up a little bit?" That being said, I'm a stickler for fit, finish, style, and format, so I take everything out of AI and I still reformat how it looks the way I like it. Which, by the way, you can train AI to do for you. It's not hard. But AI is 100% where it needs to be. It has been where it needs to be to be that tool, to run everything through it and clean it up. For mature organizations, absolutely. You have to be OK with, I will say that the ChatGPT output is very sterile, but at least in the government sector, that's how we like it. We don't necessarily want a lot of personality in our documents. But at least in the local government circles that I go in, the maturity level is not quite there yet.

I think the biggest mistake organizations can make is jumping into AI without having the proper security controls and business controls in place. We have an acceptable use policy for generative AI that tells people you can't use AI to make a management decision, and you can't use AI to evaluate personnel and things like that. The big thing is you can use AI, but you are still responsible for the output. So even though AI is absolutely ready for it, until organizations have a culture of "you still need to double-check the work before you hit send"... This has happened in local government in California. One particular story is a school district. I don't know if it was a policy or not, but they did not check it. They just said, "Write the policy," and then they sent it out and it was horrible, and someone got fired over it, because you have to be responsible for it.

So the technology is definitely there. I think the individual users are definitely there. We have statistics in the County of Fresno that people are totally on board with using AI to make their work better. But there's still the organizational culture that has to be shaped into responsible use of AI, especially in the public sector. You don't want to put people's personal data into AI and have it make a decision. But I'm actually surprised that the entire administrative side of organizations has not been using AI for years. It is the perfect tool for anybody who writes a document. It's absolutely insane to me that it's not being used in that way. And I'm not saying take the human part out of it. Just like I said, I don't go to AI and say, "Write me a security policy." I'm still the author of that document. I just have ChatGPT type it out for me so that it looks like a smart person wrote it.

Matthew Connor: I love that you say that. It's so funny. In the Army, we had this expression: you can delegate authority, but not responsibility. And it's so clear, and for some reason that doesn't translate into the civilian world, which blows my mind. Look, you can have somebody else do the work, but when they screw it up, you're still to blame. It is still your work. And the fact that so many people still hide behind that and say, "I gave it to them to do and they screwed it up." Dude, it was your responsibility. You tell them to do it, sure, but you're responsible for it. As heavy as the head. That's just how it goes. It's wild that so many people for so long have used that mentality of "that's not my fault, I gave it to so-and-so." It doesn't matter, because you're responsible for the task. If people really internalize that, it doesn't matter if I wrote it or not, you sent it out. Even if it was your subordinate who sent it out, that's still on you, because you're responsible for them. So I think people will get there with AI, because it's inherent in that. We're not going to be giving it, and I guess this is up for debate, legal entity status, although there are some places that are trying to do this for monetary reasons. But as long as it's your assistant, I think people are going to start getting on board with the fact that, yep, I'm responsible for it.

The tricky part is it's getting so good that now, unlike two years ago, where I was like, "Dude, really?" and you'd find mistakes, now the mistakes are, sure, there are mistakes, but you probably wouldn't even have caught those mistakes. You weren't thinking on this level. You weren't as detailed as the AI is now. I think the challenge is that the AI is more detailed and more dialed in than most humans, and it's seeing everything very clearly because it can hold all of the information at once. It's remembering all the email exchanges, and it's got access to data that, I'm sorry, would have taken you a week to process in order to respond or even start writing, and it did it in a matter of seconds. So I think we're at that point where people get reliant on it, and, going back to an earlier point, are we at that stage where we have to say, "You know what, it's OK that it made a mistake, because I make mistakes, but boy, look how good this is"?

Are we there yet? Or is that coming, where it's like, "Dude, this is like 98%"? I'm not perfect. Nobody's perfect. I'm not talking about making a large strategic error. But most of the thinking that comes out of there, or most of the product now, if you say "Look at this email exchange," and then it replies, if you've set things right and it's now accessing your email, all these documents, your CRM, all of the stuff that's going on, and it's replying, good luck being better than that. So I'm curious, are we at that place where, when configured, it's better than us? So if it makes a mistake, so be it. I'll take that hit.

Eric Matthews: Yeah. And I think a good way to look at it is that the first email filters were nothing but regex. It was just a very complicated regex. Was it perfect? No, email still got through. Was it better than what we had before? Tons. Something I have to remind my team over and over and over again, especially in the security world, and in fairness I would say in the business world too, is that our job is not to eliminate risk. It's to reduce it. You cannot eliminate risk. It's impossible. You have a risk appetite, and you reduce your risk below that risk appetite. So what we need is a risk appetite for AI usage. What mistakes are we going to allow it to make, and what mistakes are we not going to allow it to make? Making a grammar mistake or a small factual error is probably something that should be fine. Like you said, humans would do it too. Giving AI the ability to shut down your network, maybe we're not quite there yet to have that kind of risk appetite, and we want a human to look at that first. But if I'm going to have an AI assistant write a reply to an email for me, my personal risk appetite is that I'm going to read the email real quick, because it's on my behalf.

If someone is using AI to take some notes that I have to write a procedure from, do I care if it makes a minor mistake? Not really. It's internal. It's something we can fix. No one needs to be punished for it. So I think it's really important for an organization to understand what things you want to hold a human accountable for, and what things you're OK with just being a mistake that you can fix. If someone is writing a policy and they make a grammar mistake, you're not going to punish them for that. So why would you hold them personally accountable if the AI makes that mistake? I think the problem with people and AI is they're overcomplicating something that's super simple. AI is a tool like any other tool. If I search for something on the internet and the link I happen to look at gave me bad information, am I going to get fired over that? No, it's a mistake. So if you use AI and it hallucinates and gives you bad information, are we going to fire someone over that? No, because it's a tool like any other tool. I think what scares people is the speed at which AI does things. It's so much faster than a human. AI can delete all your databases before you even knew it did. But this may be controversial: I don't blame the AI for that. I blame the lack of controls for that.

Matthew Connor: Great. OK. And we could really go down that rabbit hole, because I feel the exact same way. I think the leading model manufacturers, with the "Oh, it got out," like, dude, if you think of this as a genius child, it's about the way you raised it. And then it goes off and does something bad. I'm sorry, we're going back to you, the horrific parent. Your child just committed a felony. Guess what? That's on you. If your kid grabs the gun out of the drawer and it's loaded and he shoots somebody, guess who's in trouble? Sure, that little five-year-old or three-year-old, but the parent. That's just how that goes. You didn't raise the kid. You let them do these bad things. It's on the parent. So I totally agree with you. You put the guardrails in place, you do the right thing. It's not that complicated.

And I think that's a full-on smokescreen, one, to not be punished for the felonies being committed by their model. Two, I think this is a longer play to shut down and control open source, because right now all the money is flowing to the frontier models. If open source is there and the hardware is getting so much better, you can run pretty decent stuff for five grand. On a Mac Studio, you can run a pretty decent model locally. At what point in the not too distant future, for five grand, can you run, let's say, a current leading frontier model? Maybe in six months, a year, you could run that for five grand. At some point it's going to be like, "Yeah, that's enough." I think that's what they're worried about, how that's going to affect their bottom line in the year or two to come, if open models are able to be run on local hardware without spending any money on tokens. I think that's what that play is. I could be wrong. But I think that's why they're saying "please regulate us," because when was the last time anybody ran to the government and said, "Hey, please give me more rules to follow"? That's the first time I remember it ever happening. So I think that's the play. We'll see if I end up dead. It's because Big AI [unclear].

Eric Matthews: It's funny that you bring up open models, because a lot of organizations don't even think about open models. I run an open model on my home machine, which is not a very expensive machine, less than $5,000. Sure, it runs slightly slower than the big frontier models out there, but it does pretty much the same job, and I'm not doing anything super complicated. I'm not running agents or anything on it. But that's something else organizations need to maybe consider: for, let's say, $10,000 or $20,000 worth of server hardware, can an open model be good enough? You have more control over it. You can put better controls on it. It's local. It's inside. Think of the public sector, like where I am. It gives us the ability to put more confidential data in that model, because I don't have to worry about whether my contract with OpenAI or my contract with Claude is good enough to protect our data. It's protected. It's ours. It's in our environment. We can delete it whenever we want. That does take having technical experts in your IT department. But I literally just downloaded LM Studio and started running an open model on my machine. It's not complicated. It may literally just take using Claude or ChatGPT to help you set it up.

Matthew Connor: Right. And now you've got your AI expert helping you set up your open source model. So I wonder, you mentioned you have 8,000 employees. If even half of those, or a quarter, whatever the percentage is, but for easy math let's say half, 4,000 people. If you're getting them an AI subscription on any level, now you're talking about a sizable amount of money. Even if it's $20 a month, $20 times 4,000, that's $80,000 a month. At what level is it, "Hey, wait, we can build our own infrastructure here, our own hardware, and run this thing"? If your burn rate is $80,000 a month for the basic plan, not to mention any tokens... Now, this is where Big AI comes and kills me, right? But the tokenomics of it are pretty obvious. If you're spending $100,000 a month on AI, at what point do you put in a few hundred thousand dollars in hardware and set it up, and now all of your data is safely there for you and the organization? You've got full control over it. It's not going out. You don't have to worry about it. You're just securing it. How far off are we from that for organizations?

Eric Matthews: Yeah. And it really comes down, in my opinion, to two factors. One is does your organization have the personnel that can support it, and the infrastructure, so the IT support around that. For an organization my size, our centralized IT can more than handle that. The second thing you need to consider is how important is the brand new stuff that just came out today with the frontier models. My guess, and I haven't done the studies on it, is that almost everybody using AI is fine using the AI from six months ago. No one is using the super brand new features that came out today, unless you're in some weird future research. For most organizations who just want the generative AI piece and maybe a few agents to help workflows move along, having an open model that's only about six months old, we're not talking about open models that are five years old, they're going to do everything you want them to do, as long as you have an IT department that can handle things like integrations. But again, it's not complicated. Matter of fact, ask your own AI how to set that up. It'll probably tell you. That's definitely something you're going to want to consider, if you want to go open or not.

For me, I think can you support the infrastructure is probably the number one thing you want to consider. For a local government like us, we're spread out through hundreds of buildings, so we have fiber everywhere, and there's lots of traffic going through that fiber. Do we want to handle the infrastructure ourselves? Do we have the staff that can support it? And then it's how comfortable are the end users using it? If everybody's already using ChatGPT, do you want to train them on how to use your open model, or just go with ChatGPT so everybody's happy? These are things you have to consider, but at the end of the day, depending on how your organization uses AI, it's very plausible that running your own open model is absolutely cheaper.

Matthew Connor: I think right now we're kind of drinking from the firehose, and we have been for a while, and, as we discussed, it's been such rapid advancement and change that I think it's hard for people to keep up. I think that plays to the frontier model. And if you look at it and you see the ROI, Blackstone points out that it's well worth it. Just put the money in. You're getting an ROI. Great. And you don't have to worry about it, kind of like Microsoft 365 versus your own Exchange server. It's a no-brainer.

Eric Matthews: That is a really nice point. However, I think the difference between M365, with Microsoft hosting your email, versus an Exchange server is that you don't have to worry about Microsoft consuming all of your email and using that against you. All we really have putting us at ease is a piece of paper saying, "Yeah, we're not going to do it. We're not going to train on it." But already you can't control it. You already can't see where everything goes. Who's to say they're not? How do we know for sure? There's no real [unclear] that I know of. Maybe I'm in the dark on this one, but I've not heard the comforting "This is how we know, and how we can verify, that your data is indeed sovereign and safe." All we need is one little time of that happening, which has probably happened plenty, but one time. And now suddenly it's like, "Well, wait a second, maybe we can host this ourselves." Or having some organization be the, and I'm sure there are plenty out there, but it's not really in the news right now that "Oh hey, this organization hosts their own. It's a great open model. It's working great. Here you go." I think you get a few of those stories and people are like, "Hey, we could do that too. Look how much money we could save."

Matthew Connor: It'll be fascinating. And I think that's the concern. The concern is that could happen on a massive scale, which disrupts the progress we're making. The investment in the infrastructure and the models is astronomical, and it's great. But if we all start going to hosting our own, it sucks that it would destroy that, and it's at a real disadvantage. So I'm just going to shut up, forget what I said. I think collectively it's good for all of us to keep using the frontier models for the time being, but in the future...

Eric Matthews: And I do agree with that. Especially for an organization that wants to dip their toe into AI, do not go down the path of trying to set up your own open model, the infrastructure, and then have an AI expert. Absolutely not. Again, talking about maturity levels, if the organization really sees a lot of value in how it uses AI, then it may be worth the conversation: "We have been spending $80,000 a month. We really like it. It works for our organization. Should we invest in this infrastructure now?" I don't have experience with open models and whatever agents you can put on top of that, as far as integration with email and files and stuff like that. I know things like Claude and ChatGPT and obviously Copilot do it very well. One of my favorite things is just asking Copilot to find an email for me. It does it way faster than I can. That doesn't mean open models can't do that. But for any organization going from no AI to wanting to play with it, you're probably going to want to slowly roll out those licenses anyway. We're not going to get 8,000 licenses and expect everybody to use it. We're going to start small and slowly roll it out. But it is a very interesting thought experiment to ask, "Hey, if this works out, can we do it cheaper? Can we do it better? Are there things we can do with our own open model that maybe we couldn't do with the frontier models because of security concerns?"

One of the things we're talking about is that if you have a really strong contract with the frontier models, and I may be accidentally or purposely putting sensitive data in, you absolutely cannot use that to train, you can't leak that, whatever, that's great. But if they do, you have to ask yourself, are you concerned about the data going out, or are you concerned about who's liable for that data going out? If all you're concerned about is the liability, use the frontier models that have a strong contract that makes them liable. If you're actually concerned about the data going out, then you really need to sit down and think about what the risk is, and is it worth the risk for you to put your data on someone else's servers? This is a problem with any SaaS solution. There are a lot of really good reasons to love SaaS, because you've pushed a lot of responsibility to that vendor to secure their own stuff. But you have to ask yourself, are you OK with the data going out as long as they're liable, or are you not OK with the data going out no matter what? Because, again, you can't eliminate risk. You can only reduce it. By using a frontier model, all you're doing is transferring the risk to them on a good contract, if they sign the contract, basically telling them they are fully financially responsible if the data gets out.

I think a lot of people are worried about the technical side of AI. They're probably not thinking about the business side of AI. That's again where I'd say use the frontier models to dip your toe in, see if you like it, see if it works. But if you have the right organization, if you're an organization of 100 people, you're not having your own open model and managing that. But we're an organization of 8,000 people, and an organization this size could probably break even or save money having our own model.

Matthew Connor: Yeah, it's an interesting thought experiment, and we'll see how it plays out over the coming months and years. But, Eric, I cannot thank you enough for coming on today. This has been an absolute blast. Before we go, can you tell everybody where they can find out more about you and more about the County of Fresno?

Eric Matthews: Yeah. You can find out anything you want to know about the County of Fresno through our website, fresnocountyca.gov. All of our departments are featured there. Everything we do for the local communities is there. It's a great website. You can find me on LinkedIn. Just search for Eric Matthews and you will find me there. I'm open to talk to people. I'd love to share my ideas, so feel free to reach out to me if you have questions or just want to chat. Or if you just want to know what it's like to have an organization that truly supports security [unclear], I'll let you know how that is.

Matthew Connor: That's awesome. All right. Thanks, Eric, and until next time.

Eric Matthews: Thank you, Matt.

 

Find your gaps before an attacker does

Book a free 20-minute call. Month-to-month, no call centers. Prefer to talk now? Call (301) 798-9170