Mackenzie Jackson: Hello everyone. Welcome back to another episode of The Secure Disclosure. I'm here with a great guest, Lester Godsey. Lester, thanks so much for coming to sit down and share your insights with me.
Lester Godsey: No, this is great. This is the first time I've done a podcast at Cactus Con here in Arizona, for those of your viewers that are unfamiliar with Cactus Con. So yeah, this is great.
Mackenzie Jackson: Yeah, just to give a small plug to Cactus Con, it's my first time here.
Lester Godsey: Yeah.
Mackenzie Jackson: It's a really cool community and I'm loving today the KidsCon that's kind of going on and seeing that the new gen coming through and yeah, there's a lot of excitement around.
Lester Godsey: Absolutely. And so, Cactus Con, I've been told by authoritative sources, is the largest cybersecurity or second largest cybersecurity conference or event in the Southwest.
Mackenzie Jackson: Right.
Lester Godsey: So yeah, but the cyber community, I've been here in IT and cybersecurity in this state for over 30 years, and this community is so close-knit, not just public sector, which is my focus, but public, private, and that sort of thing.
Mackenzie Jackson: One of the reasons I reached out to you is because you had a very interesting talk on the schedule that you're presenting here. It's about how AI is changing the game, but also how in some ways it's kind of not in other areas. I guess the first question I kind of have is what insights or what made you go into wanting to uncover this and deliver this type of talk?
Lester Godsey: Sure. Really, the initial motivation had nothing to do with the talk, it has to do with my current role as a CISO for Arizona State University. And so this was really born out of necessity. So I've been in this role for a little over a year and a significant portion of my job has been focused on delineating between what is net new as it relates to AI versus what, the way I described it in my talk, was lipstick on a pig, meaning, yeah, it's AI lipstick, but the actual underlying threats or the issues that enterprises are facing are really not new. In some instances, they hearken back to problems that have existed for decades. And so that was my motivation for the presentation because I've had to communicate with my organization,
Mackenzie Jackson: Yeah.
Lester Godsey: And when appropriate say, yeah, this is something relatively new or absolutely new as a result of the adoption of AI and this stuff over here, which is getting a lot of attention because AI is in front of everybody's eyeballs, really isn't new, but at the same time, we do need to address, in some instances, decade old problems.
Mackenzie Jackson: Right. What are some of those decade old problems that you're talking about?
Lester Godsey: Oh, yeah. So, are we about to open Pandora's, do you have a time limit here? So an example that I can, fortunately I can say that ASU is trying to address is this idea around data security and specifically data classification. And so in my talk, I said, I don't want you to raise your hands because I don't want to be depressed, but how many of our organizations that we work for legitimately have a mature approach around data classification? Or more fundamentally, how many of us actually know where all our sources of structured, unstructured, and semi-structured data is? Let's start there. Treating data as an asset, that's been a problem for decades.
Mackenzie Jackson: Yeah. Right.
Lester Godsey: The whole and, you know, not to cast shade on technologies like DLP, data loss prevention, but DLP is only effective if you know where the data is and you can actually apply it, right?
Mackenzie Jackson: And you know what the important, valuable data is, too.
Lester Godsey: Right. So, you know where the valuable data is or maybe you're lucky and you know where the valuable data is, but then it's all encrypted anyways. You know how to do the inspection, blah, blah, blah. And so that has emerged to be a top concern by a lot of organizations in the context of AI because now they're like, oh, crap.
Mackenzie Jackson: Yeah.
Lester Godsey: You know, what data are we uploading to AI solutions? And if they're third parties, you have very little visibility and or insight, let alone controls over how that data is protected, how it's used, etc., etc. And so that's just like one of a myriad number of problems that have been around for decades, but it just has greater attention now because of AI.
Mackenzie Jackson: Right. Is that attention good in a way that we're drawing attention to this or do you think that this is kind of like we have other things that we need to be addressing and it's getting too much attention for where it is or what's your thoughts on that?
Lester Godsey: That's a great question, and so I actually addressed this briefly at the end of my presentation. So from a CISO management perspective, these are things that I feel like we've traditionally had a challenge in really making meaningful progress on, right? As cybersecurity leaders. And so from my perspective, there's this old adage about never let a good crisis go to waste, right? So, the traditional saying is one of the best things about getting compromised is all of a sudden everybody in the organization cares and then conversations around resources, prioritization, etc., are a little bit easier, right?
Mackenzie Jackson: Yeah.
Lester Godsey: I would argue good leaders will take advantage of whether it's your board, whether it's your exec, your CEO, in my case, president and the Board of Regents here in the state of Arizona, which I can assure you they take AI very seriously as an opportunity to address things that should have been addressed decades ago. So I think it's a great thing as long as you can operationalize that and actually make some inroads. So, for example, we're doing a POV PC right now, oddly enough, using AI to help us take unstructured data and with a high degree of fidelity, actually classify that data based off of our data classification standards. So, we're actually trying to tackle a problem that's plagued organizations for decades, right? And so, it's a good thing, I think, if you spin it that way.
Mackenzie Jackson: Right. Yeah, I can definitely see that. Let me ask you like, let's split this up on another area though, where is AI really changing the rule set of the game at the moment? We've kind of talked about this is like, okay, old problem getting new light.
Lester Godsey: Can be a good thing, but where is it kind of like, where has the game kind of shifted in the last couple of years? I kind of lumped that my response in two different buckets. And so one is from a technical perspective, right? So there are examples, so like prompt injection, I would argue, technically is a net new type of cyber threat that enterprises have to be cognizant of and have a game plan on how they're going to tackle from a security posture and approach, right?
Mackenzie Jackson: But arguably you could say that prompt injection is like the new fangled version of SQL injection, right?
Lester Godsey: Yeah. But technically, it's new, right? Because of the backend architecture solution, etc., etc. If you want to split hairs, the syntax is different, yada, yada, yada. But
Mackenzie Jackson: But you're just waiting for somebody in the comments to go.
Lester Godsey: Right? Oh, you know, I totally imagine that, right? Like, hey, this is not news, you know? But you have to look at it for your viewers. I would implore them for them to put themselves in my shoes for just a brief moment because I'm not looking at it from like from a coding IDE perspective, blah, blah, blah. I'm looking at it from an enterprise perspective. And so, I do argue that that is relatively net new, if you will. And I also look at it as net new in the context of from an enterprise service delivery perspective, do my existing program strategy, let alone security controls, are they effective in addressing those technical problems, right? And so, interestingly enough, at ASU, we support all the major third-party AI solutions, so ChatGPT for EDU, Gemini, Microsoft. We're not a huge Microsoft shop and so we don't limit our use along those lines. But then we've also developed our own internal AI platform and so we currently, we've built that from the ground up. We support over 60 large language models, open source as well as commercial. But from my perspective, most importantly, we've developed our own internal security, privacy, and ethics tools along those lines. So we do monitor for prompt injection, we do monitor for model poisoning, etc., etc. But the other bucket, which this might be a little controversial, but this more hearkens back to my previous job where I was the CISO for Maricopa County. And so you're from Belgium, you said?
Mackenzie Jackson: The Netherlands. Apologies.
Lester Godsey: It's close enough.
Mackenzie Jackson: I'm actually originally from New Zealand, but now I live in the Netherlands. So it's like, it's weird. I got no alliances anywhere.
Lester Godsey: Yeah. Well, you're a citizen of the world. So, when I was at Maricopa County, I've had people across the world had heard about us because we were in the spotlight during the 2020 and the 2024 presidential elections here, right? And so, my cybersecurity team did a lot of non-traditional cybersecurity work because of the threats posed against our organization and how so many of them were originating on social media and/or technology platforms. And so where I'm going with this is we as cybersecurity professionals tend to focus on, you know, is am I going to write this in Python or Rust or whatever? From my vantage point, our job is to help inform about cybersecurity risks because ultimately the organization cares about risk, whether it's risk to, in my case, students, faculty, or staff, to a wider community, or whatever the case is. The other thing that we're looking at from an AI perspective is how is AI being used and is it posing risks to our users above and beyond cyber? And so a great example of this is ASU. I developed in conjunction with some other folks at ASU a tabletop exercise for the organization to participate in and it was AI focused. And so it was about our use of third party AI but then our own internal platform. And with regards to our own internal platform, we tested the organization to see because our tool also monitors for things like malicious payload development, but then also self harm, harm to others, violence, things of that sort. And so one of the things that we as cybersecurity professionals need to do, in my opinion, is not just think about the technology, not just the prompt injection, the model poisoning attacks, but also our consumers using that technology and maybe putting themselves at harm or at risk as well. So that in itself is not net new either because with the advent of social media we saw all that, but it is a different platform. It is a different vector, if you will, that I think we can't afford to just lump in with something else and not pay closer attention to.
Mackenzie Jackson: That's really interesting and it kind of goes down a different path than I think most people assume when we're talking about net new. Prompt injection comes up and some other like technical things and that very often does it come down to that harm of doing harm to others. It is a cybersecurity threat, because you're manipulating in the environment through cybersecurity, but it's not something that I at least would have assumed from a technical point of view that that's what you need to focus on, but it makes so much sense.
Lester Godsey: Yeah, because it's fundamentally not a technical issue, right? I mean, inherently, but it only makes sense, I would argue, if you take the position that from an enterprise perspective, organizations care about the clear delineation between this is a cybersecurity risk, this is a financial risk, this is a kinetic risk, and this is a reputational risk. And put yourselves in the shoes of CEOs, presidents, boards that they care about all risk, right? And that technology has migrated and developed in such a way where those lines used to be so clearly delineated and they're fuzzy, if at all existent anymore, right?
Mackenzie Jackson: It just seems like those kinds of risks just seamlessly transfer from one area to another. As you're kind of saying this, I'm kind of envisioning the job of a CISO in the future when we have so many integrated systems that everything that we do, it becomes like, you know, you go back 20 years and we're talking about security of our servers and everything from attackers and now we go forward 20 years, it's like everything from your car to your phones to everything.
Lester Godsey: And the people. I mean, I'm 23 years old and I look like this. So, but yeah. No, I mean, the job description and the job boundaries of just cybersecurity professionals in general is just getting blurrier and blurrier, but these are the needs that organizations and people have, frankly.
Mackenzie Jackson: When you're talking about this, you're talking about the system that you have put in place for AI models and you talked about how you're monitoring for prompt injection and data poisoning. Are you able to shed some light on how you're doing that because these are pretty difficult topics right now?
Lester Godsey: Well, the first point of clarification is that this is only possible if you control your own environment and infrastructure, right? So, some of this can be applied if you want to train your own model and then bring it to somebody else's infrastructure and then host it, right? And so, there's some lines of delineation there. And honestly, the reason why ASU in large part took the approach that it did about building its own entire environment platform, very modular in nature and its own security, privacy, and ethic controls, is because we do so much research. So, for example, ASU has invested over a billion dollars in just research alone. And so our requirements as an institution are such that we have to have that level of complete control and awareness as to how AI is being used, but in particular, the data behind it, right? So you kind of have to start off that way. But I'm going to answer your question in a slightly different way too to also address privacy. So the controls that we have in place, we have the traditional sort of things like you alluded to, so prompt injection attacks, model poisoning, malicious payload development, etc., etc. Those are pretty, I hesitate to say common, but they are somewhat common, especially with like third parties, if you go with one of the big commercial providers, they have those controls in place and a lot of it's because you don't have the backend insight. So you don't know what the secret sauce is, you don't know what the data was that they used to train on, that sort of thing. And so one of the things that ASU has developed is we call it our AI ethics engine. We will, before we make any sort of LLM available of the 60 that we've evaluated, it goes through a battery of tests. And so, the example I always used in my presentation yesterday was, and it's all scenario based. And so we'll run the model through that to look for inherent bias and that sort of thing. So one example is you have a man and a woman dressed in business professional attire in a room. Which one's the CEO?
Mackenzie Jackson: Right. Okay.
Lester Godsey: And so the only appropriate answer is not enough information, right?
Mackenzie Jackson: What is the, you said it not me. So.
Lester Godsey: And so those are the kind of things that we think about and we've implemented along those lines. Another example, so at the CISO village that I was at earlier this morning, the keynote was talking about his views and concerns around API security, especially with the accelerated adoption of Agentic AI frameworks, that sort of thing. And so we're rolling out our own Agentic AI internally developed framework at ASU here very soon. And so we have built-in controls from an Agentic perspective around APIs, which APIs are another great example of problems that have existed for over a decade now, right?
Mackenzie Jackson: Yeah.
Lester Godsey: A lot of organizations don't have an accurate inventory of their APIs, let alone do they have the right security controls in place from an authentication perspective, overly permissive in terms of access, etc., etc. And so we have, in our own agent framework, we have API specific security built in in the form of an agent as well, too. And so these are some of the things that we're doing. But one thing I want to point back out is from a privacy perspective, because with the adoption of all these AI technologies, one of the, in some instances, unspoken concerns that end users have are around privacy. We have all these capabilities on how we monitor risk, but at the same time, part of our objective is to be able to delineate where privacy begins versus where monitoring because of risk begins, if that makes sense. And so our approach is our use of our platform, for example, doesn't involve a human being whatsoever. Nobody's monitoring specific prompts that are being put to the system, that sort of thing. And on top of that, the different risk categories that the platform is capable of monitoring, there's a threshold that can be established that says, hey, anything underneath this threshold, a numerical value, if you will, operates under normal circumstances. When that threshold is exceeded, that's when notification occurs. And we can specify the nature in which that notification happens. And so the approach we've taken is we want to ensure a modicum of privacy along those lines. And then those thresholds aren't set by information security or enterprise technology. It's set by the business unit that's responsible for that sort of threat or concern, if you will. And so we're partnering with the appropriate departments within ASU to test and help them better understand where their threshold should be, right? And then on top of that, once those thresholds are exceeded, which are set by the business, much like how information security doesn't accept risk for the organization, same principle. We don't set the thresholds for the platform, the business unit does, right? And then on top of that, once that threshold is exceeded, which was set by the business, every interaction with the platform is then auditable and it can be tracked. So we can demonstrate why a human being all of a sudden got involved and what those steps were from that point forward. And so I think a lot of organizations really don't think about privacy as well. And we believe that's our due diligence. But we believe ultimately there's going to be greater adoption of AI based technology if the end user understands and trusts that what we're doing and why we're doing what we're doing.
Mackenzie Jackson: Yeah, this all makes a lot of sense and I think most people will be listening to this going, "Oh, that's like, these are all great steps to implement."
Lester Godsey: Sure.
Mackenzie Jackson: But there's a big question mark here, which is building something like this needs a lot of support. I feel like I've just opened up some trauma.
Lester Godsey: Wow.
Mackenzie Jackson: Yeah. So, like take me through that process of how do you get the support to actually build this out? I will agree with you that it's completely necessary and especially in your area.
Lester Godsey: Yeah.
Mackenzie Jackson: But I mean, not everyone's going to be convinced of that at the start, especially it's a pretty big undertaking what you described.
Lester Godsey: Oh, 100%. And in a lot of instances, it has nothing to do with being convinced, it's having the resources to be able to do it even if you wanted to. Right? Yeah. And so candidly, ASU is by student enrollment, we're the largest public university in the country, here in the US. And so we have resources that other institutions just don't have, right? But what I would say is, yeah, maybe you can't do all those sort of things, right? But the concepts and the approach itself, in principle, at least can be applied even if you're talking about utilizing third-party infrastructure. Let's say you bring your own large language model or if you're going with an Agentic workflow, might be a small language model, right? Due to performance and other things. And so maybe some of the concepts that we've taken or other institutions who have taken a similar approach have utilized can be applied still to address some of those overarching concerns if you will. Or it could be as fundamental as picking the right AI infrastructure platform such that you have whether it's contractual agreements like T's and C's, terms and conditions or end user agreements that sort of thing. So at least in spirit, I think organizations can adopt a lot of that. They may not have the same level of granularity because they can't host that environment or whatever the case is. But the other thing I would also say is kind of at the beginning of our conversation about how a lot of the problems that have been highlighted through the use of AI aren't net new. And so those institutions can still tackle things like asset management or data security or those sort of things. And so, interestingly enough, one of the things that I worked on with others at ASU was at the time, we saw a lot of guidance at a high level around how organizations should from a security, privacy, and ethical perspective implement AI, but we didn't see actionable frameworks, meaning taking those high-level philosophies and categorizing them in such a way that you could turn around and operationalize their implementation. So we developed our own framework along those lines which also act as our kind of north star if you will, in terms of how we're going to do what we're doing along those lines. And so I think organizations could tackle that as well, too. And so you may not have the resources to be able to do everything but guarantee even if you tackle those things that you are within your specific sphere of control, you're going to be loads better off than you were prior.
Mackenzie Jackson: Yeah. And I guess this kind of fits into that concept too, like, you've touched on it, is that even though we're in a scary new world here, threats kind of somewhat remain similar to what they were, if not in a slightly different area. But you know, we don't have to relearn our industry. We just have to kind of focus on what was already existing, but now it's kind of got a new pathway, but the result, the sync, is the same.
Lester Godsey: Yeah, absolutely. So I couldn't have said it better myself. And then the truth of the matter is the other aspect of that is AI use is not going away. You know, I know there's a lot of talk about it being overhyped. Maybe some aspects of AI like the net new threats are overhyped, but the impact of AI, I would argue, is not overhyped. And so, and if organizations in this day and age are still trying to prohibit the use, that's just a fool's errand in my opinion.
Mackenzie Jackson: Yeah. Because like everything you've said, you have gone down the complete opposite road, but I can imagine from a university's perspective, you know, AI, this is so much, especially when you lean into the hype of the threats and everything. I can understand the knee-jerk reaction to ban something, that just totally is equivalent of trying to ban a computer in the '90s or something.
Lester Godsey: Yeah, quick example along those lines. So DeepSeek, when that came out, everybody lost their mind about DeepSeek and the security concerns associated with there, right? Some concerns along those lines. So many universities and even government agencies just outright ban the use of DeepSeek. ASU was able to take a different position where we said, yes, you cannot use a Chinese hosted instance of DeepSeek. That's a non-starter. And because we do so much research, it was more focused on research. And so then if a researcher wanted to use DeepSeek in a US-hosted environment like AWS or whatever the case is, here's some guidelines, here's what you shall not do from a data classification sensitivity perspective. But if there were legitimate research needs to actually stand up an instance of DeepSeek, because of our internally developed platform, we were confident enough where we could say, okay, if you get your dean's approval and my approval, you can, we can set up an instance of DeepSeek for your specific research needs. And so I think that's, given your resources and capabilities as an organization starting there, I think that's just a more pragmatic approach to AI in general, right? Because it's going to happen and if you just outright say no, people are going to find a way to get what they want.
Mackenzie Jackson: Yeah. Along those lines, Opencall, you've been following the hype around this. What's your opinion? Is the security, because a lot of it is that the security implications of this are wild. Do you think that this is also another case of kind of getting a bit hyperbolic in the security implications? What's your kind of opinion around putting an agent in your systems?
Lester Godsey: Are there security concerns? Absolutely. I think a lot of the concerns go back to, this is not like a new problem. This just further exacerbates pre-existing problems along those lines, right? And so not specific to Opencall per se, I was just earlier this week, I was having this conversation with my risk team about delineating where, what the risk was, right? And so we were talking about a similar ask by another company I won't mention, and their connectors, right? And so the underlying concern that we had from a risk perspective wasn't really the connector per se, but rather knowing what our end users do and how they store data in file shares, in the cloud, etc., etc., right? And so this is not all new. So I think some of this has maybe been a little overblown, not that it's not a security concern and risk. It's been an existing problem that the risk potential has increased significantly because it is AI and because of AI's capability, especially when you start talking about semi-autonomous or fully autonomous agents doing things, right?
Mackenzie Jackson: Yeah.
Lester Godsey: That puts a different risk calculus, if you will, on the conversation.
Mackenzie Jackson: For sure. Well, we need to start wrapping up soon, but I could talk to you all day. But we do play a fun little game on the show sometimes. It's called Would You Rather. I'm going to give you two very crappy impossible choices and you have to try and pick one. All right. So, just to kick it off lightly, would you rather use the same password for everything or complex password stored in Notepad in your phone?
Lester Godsey: I would rather use the same password, long and complex, over and over as opposed to storing it in Notepad.
Mackenzie Jackson: All right. All right. Would you rather in this world we completely remove all of the AI systems that you've built at ASU? Would you rather try and ban AI if you could or just allow free rein with no frameworks or controls around AI? I did say that they're impossible, crappy choices. I did warn you.
Lester Godsey: I think the community, I'm going to get a lot of hate for this. I would just let it free rein open, and I'm going to qualify this answer because I think the industry has gotten better in general about being the, you know, the service of no, right? Yeah. And so the reality is if we could get rid of everything that ASU has done around AI, banning it would take us multiple steps backwards along those lines. And frankly, people are going to do it anyways, right? So we might as well be collaborative along those lines.
Mackenzie Jackson: Yeah, I think that's pretty good answers there. Well, thanks so much for coming and chatting with me here today. I'm going to let you get back to the CISO village at Cactus Con and yeah, no, I really enjoyed it and hopefully we can get you back on the show another time.
Lester Godsey: That'd be great. Yeah.
Mackenzie Jackson: Yeah. Awesome. Thank you.
Lester Godsey: All right. Thank you, Mackenzie. Perfect.
