Cameron Gaudet: Welcome back to the Aware AI Brief. this is our monthly video series looking to address kind of real life practical tangible questions or topics as it relates to AI governance. And that all comes from you, the listener. So as I start every one of these, one of these episodes, encouraging you, both those that have sent in questions and comments, and if you have not and you desire to do so, please. Please send them our way and and that's kind of what fuels fuels this discussion here. But as always, I'm joined by Walter Haydock, founder of Stack Aware. How are we doing today, Walter?
Walter Haydock: Good Cameron. Good to see ya.
Cameron Gaudet: Good, good, very good. Surviving the the summer summer months here. It's it's kind of wild to think that like we're last week of July already. I know we say that every time the calendar comes around, feel like a broken record, but but yeah, man, it seems like seems like the summer went quick.
Walter Haydock: It did. It did. Yeah. Yeah. We've still got August ahead, but in New Hampshire it starts getting cold in August.
Cameron Gaudet: Yeah. that's true, right? Yeah. Yeah. So it it changes quicker. Yeah. Where I'm at, we're we're in the heat for a little bit. So
Walter Haydock: I bet. I bet. Yeah.
Cameron Gaudet: but today I wanna dive into to some of the the questions and comments that have come come our way here this month. So let's get into let's without further ado, let's get into some of these questions that you the listener might have presented. so question number one here was How does that is the wrong you know what I think I'm in the wrong file here. I apologize.
Walter Haydock: Yeah, no worries. I set up the wrong live stream. I I I duplicated the live stream. So I'm going
Cameron Gaudet: This is
Walter Haydock: to yeah.
Cameron Gaudet: yeah. This is just to prove that that it's live. Yeah. This is this is not an act. Those would be that
Walter Haydock: This is real. This is real. We're we're doing this live. Yes. This is not you know, no.
Cameron Gaudet: would be a redo of is there that I was looking at June. Was there anything in June that you wanted to redo? I was on the on the wrong one. He's like, actually, edit. all right. Yeah, right. Right. See
Walter Haydock: I guess we could talk about it again, yeah.
Cameron Gaudet: if the make sure the answers line up. question number one from this month. Illinois has a new newish law, the Artificial Intelligence Safety Measures Act. I know I've seen you post about it, talk about it a little bit. so they define something as categ catastrophic risk. and so I the question that came in was more along the lines of does does defining risk that way, is that a measurable, you know, measurable? process and if so, does that improve accountability?
Walter Haydock: Overall, I think it's good to define risk and outcomes in quantitative terms. And I think that Illinois did do this in SB 315. They said that catastrophic risk includes anything that results in death or serious injury to 50 or more people, or a million dollars in property damage, which I mean, yeah, it would be a tragedy, but I wouldn't call that catastrophic. I mean, that happens every day. You know, it'd be it'd be bad if an AI system or AI model caused that level of damage and that would certainly be un unprecedented, but that's that's a relatively low threshold, I'd say. Although if if you're getting to that point with an AI system, then things are probably ramping up quickly. So so it might make sense to to trigger some
Cameron Gaudet: Sure. Yeah.
Walter Haydock: some some preventive or some some reactive or or yeah, some reactive controls. Some Yeah. So I overall I think that's a good, it's a good approach. The law has gotten some feedback, you know, from industry groups. They are not super thrilled about it for obvious reasons. It applies more regulation to them. It's state specific. And I think probably the biggest issue that is rate that some folks have raised is the kind of external audit requirement is not perfectly defined. There's no code of practice for. Auditing a frontier framework. So I think that's true. With that said, Stackware has proposed, you know, using ISO 402001 internal audit plus some additional controls focused on SP315 as a way to meet that requirement. And you know, I think it's certainly doable. I don't think it's that crazy a thing to get done. And you know, we're we're standing by to to assist with with that for any hyperscalers who need some assistance.
Cameron Gaudet: Yeah. Yeah. Yeah, like you said, it might maybe not maybe not that crazy of a lift. but exactly your point. If you feel like it is, reach out. Reach out for sure. the next go ahead.
Walter Haydock: One thing, one interesting thing I thought in there was that it did exempt from the definition of catastrophic risk love otherwise lawful activities of the US federal government. So if if the
Cameron Gaudet: Mm, okay.
Walter Haydock: US military has an AI weapon that does that level of damage, it's not covered under the law. Although I guess I'm not sure how
Cameron Gaudet: Sure. Not right. Hm. That is an interesting distinction to
Walter Haydock: that would be happening like in the state of Illinois. That would be
Cameron Gaudet: In Illinois. Yeah.
Walter Haydock: probably I probably don't want like a police AI that can wipe out fifty people at a time. That's that
Cameron Gaudet: Right.
Walter Haydock: that's that's even if it's under control, I I I still don't like that idea.
Cameron Gaudet: Right. Yeah. Yeah. The need for that application are already already raises major red flags. Yeah. Yeah,
Walter Haydock: Yeah. Yeah. Yeah, there are a lot of questions raised.
Cameron Gaudet: yeah. But you have all the questions, yeah.
Walter Haydock: Yeah.
Cameron Gaudet: Yep. No, that's good. That's good worth pointing out for sure. The the the next question is more more a discussion I've seen for from online about AI governance, and that's this idea of principles based governance. So some of our Just kind of discourse I've seen online. Some argue that principle-based governance gives organizations needed adaptability, easy for you to say, adaptability, as the technology evolves, while others say it becomes kind of a way to delay concrete controls. I don't know if you've seen this or heard it discussed among people th as far as to you or kind of getting your input on it, but where do you land on that term or just that general application.
Walter Haydock: I think you can have principles-based AI governance at a high level that is also not vague. So a big debate that I see is people saying, well, your policies should be high level. They shouldn't be super specific. I would say, yes, that's true, but they shouldn't be vague. And there's a difference here. You can be high level and you can say data owners are accountable for anything that a system does or does not do. That I would consider to be high level but specific. Because if something bad happens, we look at the data owner and say, hey, you know, what happened, what happened here? You know, there's no pointing
Cameron Gaudet: Right. Yeah.
Walter Haydock: a finger saying, well, you know, application AI applications should be supervised to make sure that human oversight is properly applied. You know, that that is high level and vague. That doesn't really tell you who to look to if something goes wrong. So I think. If you're using principles to guide your AI governance, that's a good idea. And I think you can have policies that are high level while at the same time not being vague. They should also specify and assign responsibility for developing specific procedures. So the CISO must develop a risk assessment process for AI systems. The data privacy officer must develop a privacy impact assessment process that takes into account AI system considerations, so on and so forth. And then Those people need to execute their responsibilities to make sure that they're managing risk effectively.
Cameron Gaudet: Yeah. Yeah, and I would say even even beyond the fact that your high level you know sh shouldn't be v vague or or can be specific, I would say that it probably should be specific in the sense of unless you're unless underneath that high level you get into it, the specifics and don't wanna don't want repetition. I've s w seen a lot of lot of good examples of that, a lot of not so great examples of that too, where the that high level terminology can is an area to be specific, especially when you as you alluded to the assigning of responsibilities, roles and responsibilities, to have direction. You know, the high level should not just be preamble. It should be it should be giving direction of some sort for you as well.
Walter Haydock: Yeah, and you can have a high level definition of covered personnel, for example. Is that employees? Is that contractors? If it's contractors, do the contractors know that they've signed up for this policy?
Cameron Gaudet: Right.
Walter Haydock: Did did they sign up for this policy? Because you can say until you're blue in the face that, you know, this policy applies to contractors, but if their contract doesn't say it does, then it doesn't. You know, it it might in your mind. So being yeah.
Cameron Gaudet: Right. Yeah, I was gonna say I'm one end of it. Yeah.
Walter Haydock: Yeah. It might yeah. It it applies it it applies in the in the in the least useful way in yeah, in your
Cameron Gaudet: Yeah, yeah.
Walter Haydock: mind. So yeah, being being you can be specific and high level at the same time.
Cameron Gaudet: Sure, absolutely. the next one was kind of a question that came up, I think, based primarily from the Google DeepMind CEO that had recently called for a new standards body kind of matching the the financial industry regulatory authority or or things with from the financial industry in order to be able to test model releases and kind of develop best practices. So I guess maybe just to enlighten the audience, what what problem is that trying to solve? And is that something that you think is a good idea? Or as we look at trying to be practical here on aware AI brief, like is that is that a real tangible way to solve the problem it's tried to solve?
Walter Haydock: There's been a lot of talk recently about developing national-level AI safety regulation. The Trump administration has put forward a variety of, in my opinion, conflicting ideas about exactly what this means. The director of the National Economic Council discussed an FDA for artificial intelligence. And then the special advisor to the president or special assistant to the president for AI and crypto. described an FDA for AI as fake news soon after that. So it's a little difficult to understand exactly what's going on here. Then the administration in June released an executive order creating a voluntary program for safety evaluation, specifically saying it did not create any sort of licensing arrangement as a FINRA would. Maybe it is because the FDA analogy has been so muddled and and kind of beaten up at this point that maybe we're just changing from pharmaceuticals to to financial services as our as our
Cameron Gaudet: Yeah. Yeah.
Walter Haydock: reference set. So I mean, I think in general it makes sense to have standards at the national level. I've always advocated for a national AI governance law that takes into account things like AI safety that consolidates all these requirements, that takes laws like SP three fifteen, preempts them and says, hey, here's the national standard for AI safety. So that Companies don't need to look at every single jurisdiction to decide what they're going to do. And FINRA, I don't necessarily think that's like the perfect way to do it. I'm always suspicious of when the government sets up private organizations to enforce standards and
Cameron Gaudet: Yeah.
Walter Haydock: to do things like that. I view that as kicking the can by the government. That's policymakers saying, we don't really want to get too involved in this. We want to outsource that.
Cameron Gaudet: Yeah.
Walter Haydock: And then it removes political accountability from. Whatever standards are being put out there. I mean, you know, what you want to get really kind of existential, we could talk about the Federal Reserve. Like, you know, is that part of the government? Depends
Cameron Gaudet: Yeah.
Walter Haydock: whom you ask, depends what context. You know, the president appoints the governors and but he can't remove them. And it's it's just it's all kind of very strange
Cameron Gaudet: Yeah.
Walter Haydock: to me. Whereas the FDA, the commissioner of the FDA is an officer appointed by the president and confirmed by the Senate.
Cameron Gaudet: Yeah. Yeah.
Walter Haydock: No questions there. You know, president doesn't like what he's doing, boom, he's gone. Actually, the FDA commissioner did just resign recently, a couple couple of weeks ago.
Cameron Gaudet: Right. Right.
Walter Haydock: so yeah, I think that's the right move. And there's a lot of talk about, you know, undue political influence and things like that in independent agencies. I mean, this is more of a political opinion on my part, but you know, I I worked on Capitol Hill. I've served in two out of three branches of government. My feeling is that, you know. The concept of an independent agency is unconstitutional, which actually the Supreme Court just said it it was. So, you know, the current the current prevailing law is is on my side here. but also just as citizens, we don't really want these autonomous bodies that we can't really get rid of, you know, being in charge of commerce and and development. You know, I want somebody who a senator can call in front of. a hearing and say, hey, you know, why
Cameron Gaudet: Yeah.
Walter Haydock: are you doing this? I want somebody that the president can fire if the president doesn't like what he or she is doing. You know, sounds sounds harsh, but but otherwise, you know, if if the president and Congress aren't in charge, who is in charge?
Cameron Gaudet: Mm-hmm.
Walter Haydock: That's my that's my thinking. So I don't I don't necessarily love the FINRA type of of approach. I think actually the FDA is a better is a better approach for for what it's worth, even though that that concept has gotten beaten up. probably not as oppressive as the FDA has become, because I think we've gotten some, you know, we've we've lost out on some innovation in the biotech space based on, you the FDA's approval process and things like that. So maybe, maybe a lighter weight FDA approach could could make sense. and only in the most risky cases would there be government intervention or preclearance required.
Cameron Gaudet: Mm. Yeah, I think I mean you you bring up some great points as far as that kind of that model and and trying to figure out I I do think even an FDA type of regulation probably I mean it I think it took the FDA a bit to get to the things that that we might look at as far as flaws. So there's there's a way to kind of kind of rewrite even that structure. But I as you were talking, I was just thinking like the The biggest thing about this whole idea or concept is that accountability structure, right? Like it is, I think that that could be a glaring well, it it could just be a very important piece of the puzzle when it comes to some sort of new standards body or what that looks like specifically is is what you were saying. The ability to be able to have someone stand before a Senate hearing, to the ability to be able to to keep accountability actionable, like to be able to actually hold someone accountable and and to have that structure in place. I think if the this a this is a little bit hypothetical, you know, like where where could we go from here, but but I think that's something that has to be part of the consideration. okay, this question came in as well. The people writing governance policy are often not the people who understand the underlying processes and control controls. Is that something that you see as a potential issue for organization? And how does that disconnect or undermine governance efforts and practice? So kind of the people writing the policy, maybe not the people that are quite as hands-on. And is that an issue? And
Walter Haydock: It's a huge problem if those who are writing policies do not understand how controls would be applied in the real world. And I'll give you an example in the realm of software development and product development. There used to be the role, or there still is, but less so these days. There's there was the role of product manager who would understand the business requirements, translate them into technical requirements for developers who would code up the application. Increasingly, we're seeing those two roles compress because of the speed of AI development. So a product manager can go from idea to working prototype very quickly. He or she doesn't need to write specifications and then, you know, have the development team whip it up and then check it and make sure it fits. That can all happen in the course of an afternoon. So I think we're gonna see the same thing with GRC engineering, governance, risk, and compliance engineering. The What I call GRC flat professionals, those who don't have the engineering capabilities or chops, are really going to need to learn how to develop those in order to implement these controls in an actionable way. Because the idea of having a pie in the sky policy where no one knows how to implement it, even the person writing the policy doesn't know how to implement it. I think that's going to be less and less valuable, if it was ever valuable to begin with. And You know, I think at best those types of policies are used as a retroactive enforcement mechanism to say, hey, you didn't follow policy, we're gonna discipline you
Cameron Gaudet: Sure, sure.
Walter Haydock: because you you didn't do XYZ. That is not the best way to to approach things. You know, in in a in a high functioning organization, somebody makes a mistake, you immediately issue a correction, you block, you know, you block a a an an inappropriate action as it's happening using technical controls. You know, if that person keeps making the same mistake, then maybe it's not a mistake, maybe it's a pattern. Then you know you're logging all these events, and then eventually you move into the disciplinary realm. So I think that's the right way to implement controls is to have a spectrum and you need a technical enforcement mechanism for that. So increasingly we're going to see GRC teams become true GRC engineers, and it's going to be It's gonna be an elite profession and there are gonna be fewer of them, but I think it's gonna be more effective for organizations going forward.
Cameron Gaudet: Sure. Sure. Yeah, and definitely effective and and to your point necessary, right? Like for to keep up with keep up with kind of the the rapid the rapid scaling for sure. if you could fix one problem or process in typical AI governance, we'll say tomorrow, 'cause a couple hours left of the day. We won't put that pressure on you. But if you could fix it by tomorrow, what would it be and why?
Walter Haydock: Asset inventory would be the first thing that I would fix when it comes to AI governance because even the most highly functioning organizations don't have great asset inventory. All their data scattered across various different sources. They don't know what applications are being mothballed. It there's no single place you can go in. The CISO can go in and say, okay, you know, how many endpoints do I have exposed to the internet right now? How many virtual machines are spinning, how many bedrock instances are going. It's very challenging to get that single source of truth for even small organizations. And I get it, it's difficult. I I I understand. I personally maintain the stack aware asset inventory for a very small company. And it it can be challenging sometimes, but it is possible to do. And that would be the number one thing that I would fix tomorrow.
Cameron Gaudet: I like. So yeah, let's do it. Let's let's fix it just like that. For sure. For sure.
Walter Haydock: Yeah, yeah, easy. Easy. Hey, Stack Aware, Stack Aware helps with asset inventory. We we plug into GRC platforms like Vanta or Drata. We plug into your cloud providers like AWS, like Azure, and we can pull down your assets from there and have a standardized you know list. And then also we can bounce the risks against the assets, which actually I think is a key thing that's missing from a lot of the governance risk compliance platforms. You look at these example risks and it's like, You know, somebody accidentally releases sensitive data, that's the risk. You know, what's the asset
Cameron Gaudet: Mm-hmm.
Walter Haydock: impacted? The sensitive data. Like we need to we need to be more specific than that. You know, is it
Cameron Gaudet: Yeah, yeah, yeah.
Walter Haydock: this database that has this protected health information in it that is going to be targeted with this vulnerability? Like that's the level specificity you need to get into to have an actionable risk register.
Cameron Gaudet: Yeah. 'cause that's that's how that's what it's gonna look like. It's not gonna look like an alert that says somebody took this and d yeah. It's gonna be more much, much more specific. Yeah.
Walter Haydock: Right. Yeah. Yeah. Hackers are not asking Chat GPT to write an exploit to like make a
Cameron Gaudet: Right.
Walter Haydock: person leak sensitive data by accident. They're being much more specific. They're saying, hey, write an exploit
Cameron Gaudet: Yeah. Yeah, much more specific.
Walter Haydock: for this CVE in this environment.
Cameron Gaudet: Yeah. Yeah. Yeah, that's that's really really good point. The next next little part here kinda to wrap up our episode today. you know I like to come up with these kind of strange segments, if you would. like I said, I blame my sports radio listening for for decades and decades for for the little little gambits, but this one I'm gonna call explain yourself. So Not necess that sounds way more accusatory than I than I desire
Walter Haydock: Mm-hmm.
Cameron Gaudet: it to be, but some things I've seen either a post online or maybe there were some people that had questions, kind of follow-up and comments, things like that. but may just just kind of more of a a verbal platform to to explain some some things and topics you've written on. So the first one is post that you had on AI pen testing. And one of the one of the first key items you had brought up was AI enabled authorization bypass. So maybe more of like a definition for for that, but but yeah, what do you got for that one?
Walter Haydock: AI enabled authorization bypass happens when you architect a feature in a way that accidentally overrides your intended authorization model. So one of the most common things that I see during AI pen tests is if there's a an application that's doing retrieval augmented generation against a data set, if you use the Service account with essentially unlimited permissions to pull the data from the database, the end user may not necessarily be authorized to see all the data that's being accessed through the rag process. So essentially you've you've jumped your authorization boundary by accident using AI. And this is just it's a very preventable thing. You need to make sure that you're retrieving the data using the permissions of the user who's eventually going to see them. It's very preventable, but I see it all the time. And I'm sure I will continue to see it.
Cameron Gaudet: Yeah. Okay. Okay. I definitely I I'd seen that one in that post and I was like that'd be a good one to to highlight and allow allow people and I say people myself too to to
Walter Haydock: Mm.
Cameron Gaudet: understand the full concept. It's for other people. It's definitely not for my own education. the next one was a post on GRC consulting. So you were talking about in that post the the GRC consulting service that just comes and leaves you with a two hundred page PDF and says good luck. And then you had you had listed on there, you know, allow Stack Aware to build the machine, right? So I guess build the machine, maybe, maybe exposition on that.
Walter Haydock: One of the worst things that a consultant can come in and do is just drop a PDF document that is meant to be read and consumed and implemented and then disappearing. And I bump into a lot of folks who've kind of been on the receiving end of that type of treatment from a consultant. And I take a different approach. What we do is we build repeatable systems for our customers to make sure that. They can run the machine when StackWare is gone. The whole point is not to build a reliance on a consultant who comes in every year and then kind of reissues the same recommendations in a different way. The point is to give our customers the ability to stand on their own and then supplement them with data, with expert advice once the implementation is through. And we strive to you know build on the customer's environment. We use their GRC platform, we use their, you know, however they want to present data, that type of things. We leverage that to the maximum degree possible. And that's what I meant by, you know, not avoiding the traditional GRC consultant approach.
Cameron Gaudet: Yeah. Yeah, building building machine. And and honestly kind of got into that a little bit talking about GRC engineering versus that kind of flat I think you called it kind of more flat GRC a
Walter Haydock: Mm-hmm.
Cameron Gaudet: little bit earlier. Just that something that something that is much more practical but lives and breathes with your organization as well. Yeah.
Walter Haydock: Yeah.
Cameron Gaudet: okay, last one for this, explain yourself. you had written your AI governance is a spreadsheet. That's the actual problem.
Walter Haydock: The way that I approach spreadsheets is they are essentially the front end for information. That's how people could consume them because a lot of people are used to seeing spreadsheets. But almost every spreadsheet that Stack Aware ever generates has behind it an API call that's pulling from a database where there's a single source of truth, everything is versioned, where there's you know linkage between different data sets. And if you're just using spreadsheets. As the source of truth, then you're gonna have problems with version control. They're gonna be brittle. You're gonna have, you know, version final, final three, copy, delete me, you know, being emailed
Cameron Gaudet: Mm-hmm.
Walter Haydock: to he emailed to the auditor, which is not a good look. so to the maximum extent possible, use systems of record. You can use spreadsheets, you know, sparingly. And then if you are going to use them, then use them as a layer on top of a source of truth. So, you know, all the stackware templates that we put out there. They are compatible with our back end. You know, if you're just downloading them for free, you're not going to get access to our our back end data and the updating. That that's something we reserve for our customers. But, you know, that's an incentive to to work with us.
Cameron Gaudet: Yeah, perfect. Perfect. And the the last comment I wanted to bring up from this past month, I really liked this one. somebody had had commented on you have a wonderful video about the complete ISO forty two thousand one certification roadmap video. I highly recommend it. it for anyone that is even remotely interested in in ISO forty two thousand one. It's great, great detail and and and delivering there. The comment was anywhere I can get more info. So maybe maybe directional directional advice there.
Walter Haydock: Yeah, the best place to go after watching the complete guide to ISO402001 in 2026 would be management system.ai, which is StackAware's five-day email course on ISO402001. You can get more information in a written format, and it's a good reference for you if you're building your AI management system.
Cameron Gaudet: Perfect. Perfect. All right. Well, I that's that's what I had for today as far as the questions that came in, Walter. And I appreciate you taking the time to to to sit here and and answer these and discuss with myself. thank you again to all that have have continued to or or in this last month have contributed to the questions and comments and topics. I was thinking about it at in preparation for this, that I think with it it's important not just to have questions but to be asking kind of new questions, to be having new conversation as as this technology continues to advance week in, week out. if we're asking the same questions and then talking about it the same way, we're probably falling behind a little bit. And and it's tough to keep up with everything. But my advice is always just ask keep asking new questions. Keep finding new topics to discuss when it comes to this that'll that'll only kind of help us grow and and self-educate as a as a collective really. So appreciate all the questions, appreciate everybody's input. And thank you for listening. We'll catch you again next month on Aware AI Brief.
Walter Haydock: Thanks everyone.
Walter Haydock: Good Cameron. Good to see ya.
Cameron Gaudet: Good, good, very good. Surviving the the summer summer months here. It's it's kind of wild to think that like we're last week of July already. I know we say that every time the calendar comes around, feel like a broken record, but but yeah, man, it seems like seems like the summer went quick.
Walter Haydock: It did. It did. Yeah. Yeah. We've still got August ahead, but in New Hampshire it starts getting cold in August.
Cameron Gaudet: Yeah. that's true, right? Yeah. Yeah. So it it changes quicker. Yeah. Where I'm at, we're we're in the heat for a little bit. So
Walter Haydock: I bet. I bet. Yeah.
Cameron Gaudet: but today I wanna dive into to some of the the questions and comments that have come come our way here this month. So let's get into let's without further ado, let's get into some of these questions that you the listener might have presented. so question number one here was How does that is the wrong you know what I think I'm in the wrong file here. I apologize.
Walter Haydock: Yeah, no worries. I set up the wrong live stream. I I I duplicated the live stream. So I'm going
Cameron Gaudet: This is
Walter Haydock: to yeah.
Cameron Gaudet: yeah. This is just to prove that that it's live. Yeah. This is this is not an act. Those would be that
Walter Haydock: This is real. This is real. We're we're doing this live. Yes. This is not you know, no.
Cameron Gaudet: would be a redo of is there that I was looking at June. Was there anything in June that you wanted to redo? I was on the on the wrong one. He's like, actually, edit. all right. Yeah, right. Right. See
Walter Haydock: I guess we could talk about it again, yeah.
Cameron Gaudet: if the make sure the answers line up. question number one from this month. Illinois has a new newish law, the Artificial Intelligence Safety Measures Act. I know I've seen you post about it, talk about it a little bit. so they define something as categ catastrophic risk. and so I the question that came in was more along the lines of does does defining risk that way, is that a measurable, you know, measurable? process and if so, does that improve accountability?
Walter Haydock: Overall, I think it's good to define risk and outcomes in quantitative terms. And I think that Illinois did do this in SB 315. They said that catastrophic risk includes anything that results in death or serious injury to 50 or more people, or a million dollars in property damage, which I mean, yeah, it would be a tragedy, but I wouldn't call that catastrophic. I mean, that happens every day. You know, it'd be it'd be bad if an AI system or AI model caused that level of damage and that would certainly be un unprecedented, but that's that's a relatively low threshold, I'd say. Although if if you're getting to that point with an AI system, then things are probably ramping up quickly. So so it might make sense to to trigger some
Cameron Gaudet: Sure. Yeah.
Walter Haydock: some some preventive or some some reactive or or yeah, some reactive controls. Some Yeah. So I overall I think that's a good, it's a good approach. The law has gotten some feedback, you know, from industry groups. They are not super thrilled about it for obvious reasons. It applies more regulation to them. It's state specific. And I think probably the biggest issue that is rate that some folks have raised is the kind of external audit requirement is not perfectly defined. There's no code of practice for. Auditing a frontier framework. So I think that's true. With that said, Stackware has proposed, you know, using ISO 402001 internal audit plus some additional controls focused on SP315 as a way to meet that requirement. And you know, I think it's certainly doable. I don't think it's that crazy a thing to get done. And you know, we're we're standing by to to assist with with that for any hyperscalers who need some assistance.
Cameron Gaudet: Yeah. Yeah. Yeah, like you said, it might maybe not maybe not that crazy of a lift. but exactly your point. If you feel like it is, reach out. Reach out for sure. the next go ahead.
Walter Haydock: One thing, one interesting thing I thought in there was that it did exempt from the definition of catastrophic risk love otherwise lawful activities of the US federal government. So if if the
Cameron Gaudet: Mm, okay.
Walter Haydock: US military has an AI weapon that does that level of damage, it's not covered under the law. Although I guess I'm not sure how
Cameron Gaudet: Sure. Not right. Hm. That is an interesting distinction to
Walter Haydock: that would be happening like in the state of Illinois. That would be
Cameron Gaudet: In Illinois. Yeah.
Walter Haydock: probably I probably don't want like a police AI that can wipe out fifty people at a time. That's that
Cameron Gaudet: Right.
Walter Haydock: that's that's even if it's under control, I I I still don't like that idea.
Cameron Gaudet: Right. Yeah. Yeah. The need for that application are already already raises major red flags. Yeah. Yeah,
Walter Haydock: Yeah. Yeah. Yeah, there are a lot of questions raised.
Cameron Gaudet: yeah. But you have all the questions, yeah.
Walter Haydock: Yeah.
Cameron Gaudet: Yep. No, that's good. That's good worth pointing out for sure. The the the next question is more more a discussion I've seen for from online about AI governance, and that's this idea of principles based governance. So some of our Just kind of discourse I've seen online. Some argue that principle-based governance gives organizations needed adaptability, easy for you to say, adaptability, as the technology evolves, while others say it becomes kind of a way to delay concrete controls. I don't know if you've seen this or heard it discussed among people th as far as to you or kind of getting your input on it, but where do you land on that term or just that general application.
Walter Haydock: I think you can have principles-based AI governance at a high level that is also not vague. So a big debate that I see is people saying, well, your policies should be high level. They shouldn't be super specific. I would say, yes, that's true, but they shouldn't be vague. And there's a difference here. You can be high level and you can say data owners are accountable for anything that a system does or does not do. That I would consider to be high level but specific. Because if something bad happens, we look at the data owner and say, hey, you know, what happened, what happened here? You know, there's no pointing
Cameron Gaudet: Right. Yeah.
Walter Haydock: a finger saying, well, you know, application AI applications should be supervised to make sure that human oversight is properly applied. You know, that that is high level and vague. That doesn't really tell you who to look to if something goes wrong. So I think. If you're using principles to guide your AI governance, that's a good idea. And I think you can have policies that are high level while at the same time not being vague. They should also specify and assign responsibility for developing specific procedures. So the CISO must develop a risk assessment process for AI systems. The data privacy officer must develop a privacy impact assessment process that takes into account AI system considerations, so on and so forth. And then Those people need to execute their responsibilities to make sure that they're managing risk effectively.
Cameron Gaudet: Yeah. Yeah, and I would say even even beyond the fact that your high level you know sh shouldn't be v vague or or can be specific, I would say that it probably should be specific in the sense of unless you're unless underneath that high level you get into it, the specifics and don't wanna don't want repetition. I've s w seen a lot of lot of good examples of that, a lot of not so great examples of that too, where the that high level terminology can is an area to be specific, especially when you as you alluded to the assigning of responsibilities, roles and responsibilities, to have direction. You know, the high level should not just be preamble. It should be it should be giving direction of some sort for you as well.
Walter Haydock: Yeah, and you can have a high level definition of covered personnel, for example. Is that employees? Is that contractors? If it's contractors, do the contractors know that they've signed up for this policy?
Cameron Gaudet: Right.
Walter Haydock: Did did they sign up for this policy? Because you can say until you're blue in the face that, you know, this policy applies to contractors, but if their contract doesn't say it does, then it doesn't. You know, it it might in your mind. So being yeah.
Cameron Gaudet: Right. Yeah, I was gonna say I'm one end of it. Yeah.
Walter Haydock: Yeah. It might yeah. It it applies it it applies in the in the in the least useful way in yeah, in your
Cameron Gaudet: Yeah, yeah.
Walter Haydock: mind. So yeah, being being you can be specific and high level at the same time.
Cameron Gaudet: Sure, absolutely. the next one was kind of a question that came up, I think, based primarily from the Google DeepMind CEO that had recently called for a new standards body kind of matching the the financial industry regulatory authority or or things with from the financial industry in order to be able to test model releases and kind of develop best practices. So I guess maybe just to enlighten the audience, what what problem is that trying to solve? And is that something that you think is a good idea? Or as we look at trying to be practical here on aware AI brief, like is that is that a real tangible way to solve the problem it's tried to solve?
Walter Haydock: There's been a lot of talk recently about developing national-level AI safety regulation. The Trump administration has put forward a variety of, in my opinion, conflicting ideas about exactly what this means. The director of the National Economic Council discussed an FDA for artificial intelligence. And then the special advisor to the president or special assistant to the president for AI and crypto. described an FDA for AI as fake news soon after that. So it's a little difficult to understand exactly what's going on here. Then the administration in June released an executive order creating a voluntary program for safety evaluation, specifically saying it did not create any sort of licensing arrangement as a FINRA would. Maybe it is because the FDA analogy has been so muddled and and kind of beaten up at this point that maybe we're just changing from pharmaceuticals to to financial services as our as our
Cameron Gaudet: Yeah. Yeah.
Walter Haydock: reference set. So I mean, I think in general it makes sense to have standards at the national level. I've always advocated for a national AI governance law that takes into account things like AI safety that consolidates all these requirements, that takes laws like SP three fifteen, preempts them and says, hey, here's the national standard for AI safety. So that Companies don't need to look at every single jurisdiction to decide what they're going to do. And FINRA, I don't necessarily think that's like the perfect way to do it. I'm always suspicious of when the government sets up private organizations to enforce standards and
Cameron Gaudet: Yeah.
Walter Haydock: to do things like that. I view that as kicking the can by the government. That's policymakers saying, we don't really want to get too involved in this. We want to outsource that.
Cameron Gaudet: Yeah.
Walter Haydock: And then it removes political accountability from. Whatever standards are being put out there. I mean, you know, what you want to get really kind of existential, we could talk about the Federal Reserve. Like, you know, is that part of the government? Depends
Cameron Gaudet: Yeah.
Walter Haydock: whom you ask, depends what context. You know, the president appoints the governors and but he can't remove them. And it's it's just it's all kind of very strange
Cameron Gaudet: Yeah.
Walter Haydock: to me. Whereas the FDA, the commissioner of the FDA is an officer appointed by the president and confirmed by the Senate.
Cameron Gaudet: Yeah. Yeah.
Walter Haydock: No questions there. You know, president doesn't like what he's doing, boom, he's gone. Actually, the FDA commissioner did just resign recently, a couple couple of weeks ago.
Cameron Gaudet: Right. Right.
Walter Haydock: so yeah, I think that's the right move. And there's a lot of talk about, you know, undue political influence and things like that in independent agencies. I mean, this is more of a political opinion on my part, but you know, I I worked on Capitol Hill. I've served in two out of three branches of government. My feeling is that, you know. The concept of an independent agency is unconstitutional, which actually the Supreme Court just said it it was. So, you know, the current the current prevailing law is is on my side here. but also just as citizens, we don't really want these autonomous bodies that we can't really get rid of, you know, being in charge of commerce and and development. You know, I want somebody who a senator can call in front of. a hearing and say, hey, you know, why
Cameron Gaudet: Yeah.
Walter Haydock: are you doing this? I want somebody that the president can fire if the president doesn't like what he or she is doing. You know, sounds sounds harsh, but but otherwise, you know, if if the president and Congress aren't in charge, who is in charge?
Cameron Gaudet: Mm-hmm.
Walter Haydock: That's my that's my thinking. So I don't I don't necessarily love the FINRA type of of approach. I think actually the FDA is a better is a better approach for for what it's worth, even though that that concept has gotten beaten up. probably not as oppressive as the FDA has become, because I think we've gotten some, you know, we've we've lost out on some innovation in the biotech space based on, you the FDA's approval process and things like that. So maybe, maybe a lighter weight FDA approach could could make sense. and only in the most risky cases would there be government intervention or preclearance required.
Cameron Gaudet: Mm. Yeah, I think I mean you you bring up some great points as far as that kind of that model and and trying to figure out I I do think even an FDA type of regulation probably I mean it I think it took the FDA a bit to get to the things that that we might look at as far as flaws. So there's there's a way to kind of kind of rewrite even that structure. But I as you were talking, I was just thinking like the The biggest thing about this whole idea or concept is that accountability structure, right? Like it is, I think that that could be a glaring well, it it could just be a very important piece of the puzzle when it comes to some sort of new standards body or what that looks like specifically is is what you were saying. The ability to be able to have someone stand before a Senate hearing, to the ability to be able to to keep accountability actionable, like to be able to actually hold someone accountable and and to have that structure in place. I think if the this a this is a little bit hypothetical, you know, like where where could we go from here, but but I think that's something that has to be part of the consideration. okay, this question came in as well. The people writing governance policy are often not the people who understand the underlying processes and control controls. Is that something that you see as a potential issue for organization? And how does that disconnect or undermine governance efforts and practice? So kind of the people writing the policy, maybe not the people that are quite as hands-on. And is that an issue? And
Walter Haydock: It's a huge problem if those who are writing policies do not understand how controls would be applied in the real world. And I'll give you an example in the realm of software development and product development. There used to be the role, or there still is, but less so these days. There's there was the role of product manager who would understand the business requirements, translate them into technical requirements for developers who would code up the application. Increasingly, we're seeing those two roles compress because of the speed of AI development. So a product manager can go from idea to working prototype very quickly. He or she doesn't need to write specifications and then, you know, have the development team whip it up and then check it and make sure it fits. That can all happen in the course of an afternoon. So I think we're gonna see the same thing with GRC engineering, governance, risk, and compliance engineering. The What I call GRC flat professionals, those who don't have the engineering capabilities or chops, are really going to need to learn how to develop those in order to implement these controls in an actionable way. Because the idea of having a pie in the sky policy where no one knows how to implement it, even the person writing the policy doesn't know how to implement it. I think that's going to be less and less valuable, if it was ever valuable to begin with. And You know, I think at best those types of policies are used as a retroactive enforcement mechanism to say, hey, you didn't follow policy, we're gonna discipline you
Cameron Gaudet: Sure, sure.
Walter Haydock: because you you didn't do XYZ. That is not the best way to to approach things. You know, in in a in a high functioning organization, somebody makes a mistake, you immediately issue a correction, you block, you know, you block a a an an inappropriate action as it's happening using technical controls. You know, if that person keeps making the same mistake, then maybe it's not a mistake, maybe it's a pattern. Then you know you're logging all these events, and then eventually you move into the disciplinary realm. So I think that's the right way to implement controls is to have a spectrum and you need a technical enforcement mechanism for that. So increasingly we're going to see GRC teams become true GRC engineers, and it's going to be It's gonna be an elite profession and there are gonna be fewer of them, but I think it's gonna be more effective for organizations going forward.
Cameron Gaudet: Sure. Sure. Yeah, and definitely effective and and to your point necessary, right? Like for to keep up with keep up with kind of the the rapid the rapid scaling for sure. if you could fix one problem or process in typical AI governance, we'll say tomorrow, 'cause a couple hours left of the day. We won't put that pressure on you. But if you could fix it by tomorrow, what would it be and why?
Walter Haydock: Asset inventory would be the first thing that I would fix when it comes to AI governance because even the most highly functioning organizations don't have great asset inventory. All their data scattered across various different sources. They don't know what applications are being mothballed. It there's no single place you can go in. The CISO can go in and say, okay, you know, how many endpoints do I have exposed to the internet right now? How many virtual machines are spinning, how many bedrock instances are going. It's very challenging to get that single source of truth for even small organizations. And I get it, it's difficult. I I I understand. I personally maintain the stack aware asset inventory for a very small company. And it it can be challenging sometimes, but it is possible to do. And that would be the number one thing that I would fix tomorrow.
Cameron Gaudet: I like. So yeah, let's do it. Let's let's fix it just like that. For sure. For sure.
Walter Haydock: Yeah, yeah, easy. Easy. Hey, Stack Aware, Stack Aware helps with asset inventory. We we plug into GRC platforms like Vanta or Drata. We plug into your cloud providers like AWS, like Azure, and we can pull down your assets from there and have a standardized you know list. And then also we can bounce the risks against the assets, which actually I think is a key thing that's missing from a lot of the governance risk compliance platforms. You look at these example risks and it's like, You know, somebody accidentally releases sensitive data, that's the risk. You know, what's the asset
Cameron Gaudet: Mm-hmm.
Walter Haydock: impacted? The sensitive data. Like we need to we need to be more specific than that. You know, is it
Cameron Gaudet: Yeah, yeah, yeah.
Walter Haydock: this database that has this protected health information in it that is going to be targeted with this vulnerability? Like that's the level specificity you need to get into to have an actionable risk register.
Cameron Gaudet: Yeah. 'cause that's that's how that's what it's gonna look like. It's not gonna look like an alert that says somebody took this and d yeah. It's gonna be more much, much more specific. Yeah.
Walter Haydock: Right. Yeah. Yeah. Hackers are not asking Chat GPT to write an exploit to like make a
Cameron Gaudet: Right.
Walter Haydock: person leak sensitive data by accident. They're being much more specific. They're saying, hey, write an exploit
Cameron Gaudet: Yeah. Yeah, much more specific.
Walter Haydock: for this CVE in this environment.
Cameron Gaudet: Yeah. Yeah. Yeah, that's that's really really good point. The next next little part here kinda to wrap up our episode today. you know I like to come up with these kind of strange segments, if you would. like I said, I blame my sports radio listening for for decades and decades for for the little little gambits, but this one I'm gonna call explain yourself. So Not necess that sounds way more accusatory than I than I desire
Walter Haydock: Mm-hmm.
Cameron Gaudet: it to be, but some things I've seen either a post online or maybe there were some people that had questions, kind of follow-up and comments, things like that. but may just just kind of more of a a verbal platform to to explain some some things and topics you've written on. So the first one is post that you had on AI pen testing. And one of the one of the first key items you had brought up was AI enabled authorization bypass. So maybe more of like a definition for for that, but but yeah, what do you got for that one?
Walter Haydock: AI enabled authorization bypass happens when you architect a feature in a way that accidentally overrides your intended authorization model. So one of the most common things that I see during AI pen tests is if there's a an application that's doing retrieval augmented generation against a data set, if you use the Service account with essentially unlimited permissions to pull the data from the database, the end user may not necessarily be authorized to see all the data that's being accessed through the rag process. So essentially you've you've jumped your authorization boundary by accident using AI. And this is just it's a very preventable thing. You need to make sure that you're retrieving the data using the permissions of the user who's eventually going to see them. It's very preventable, but I see it all the time. And I'm sure I will continue to see it.
Cameron Gaudet: Yeah. Okay. Okay. I definitely I I'd seen that one in that post and I was like that'd be a good one to to highlight and allow allow people and I say people myself too to to
Walter Haydock: Mm.
Cameron Gaudet: understand the full concept. It's for other people. It's definitely not for my own education. the next one was a post on GRC consulting. So you were talking about in that post the the GRC consulting service that just comes and leaves you with a two hundred page PDF and says good luck. And then you had you had listed on there, you know, allow Stack Aware to build the machine, right? So I guess build the machine, maybe, maybe exposition on that.
Walter Haydock: One of the worst things that a consultant can come in and do is just drop a PDF document that is meant to be read and consumed and implemented and then disappearing. And I bump into a lot of folks who've kind of been on the receiving end of that type of treatment from a consultant. And I take a different approach. What we do is we build repeatable systems for our customers to make sure that. They can run the machine when StackWare is gone. The whole point is not to build a reliance on a consultant who comes in every year and then kind of reissues the same recommendations in a different way. The point is to give our customers the ability to stand on their own and then supplement them with data, with expert advice once the implementation is through. And we strive to you know build on the customer's environment. We use their GRC platform, we use their, you know, however they want to present data, that type of things. We leverage that to the maximum degree possible. And that's what I meant by, you know, not avoiding the traditional GRC consultant approach.
Cameron Gaudet: Yeah. Yeah, building building machine. And and honestly kind of got into that a little bit talking about GRC engineering versus that kind of flat I think you called it kind of more flat GRC a
Walter Haydock: Mm-hmm.
Cameron Gaudet: little bit earlier. Just that something that something that is much more practical but lives and breathes with your organization as well. Yeah.
Walter Haydock: Yeah.
Cameron Gaudet: okay, last one for this, explain yourself. you had written your AI governance is a spreadsheet. That's the actual problem.
Walter Haydock: The way that I approach spreadsheets is they are essentially the front end for information. That's how people could consume them because a lot of people are used to seeing spreadsheets. But almost every spreadsheet that Stack Aware ever generates has behind it an API call that's pulling from a database where there's a single source of truth, everything is versioned, where there's you know linkage between different data sets. And if you're just using spreadsheets. As the source of truth, then you're gonna have problems with version control. They're gonna be brittle. You're gonna have, you know, version final, final three, copy, delete me, you know, being emailed
Cameron Gaudet: Mm-hmm.
Walter Haydock: to he emailed to the auditor, which is not a good look. so to the maximum extent possible, use systems of record. You can use spreadsheets, you know, sparingly. And then if you are going to use them, then use them as a layer on top of a source of truth. So, you know, all the stackware templates that we put out there. They are compatible with our back end. You know, if you're just downloading them for free, you're not going to get access to our our back end data and the updating. That that's something we reserve for our customers. But, you know, that's an incentive to to work with us.
Cameron Gaudet: Yeah, perfect. Perfect. And the the last comment I wanted to bring up from this past month, I really liked this one. somebody had had commented on you have a wonderful video about the complete ISO forty two thousand one certification roadmap video. I highly recommend it. it for anyone that is even remotely interested in in ISO forty two thousand one. It's great, great detail and and and delivering there. The comment was anywhere I can get more info. So maybe maybe directional directional advice there.
Walter Haydock: Yeah, the best place to go after watching the complete guide to ISO402001 in 2026 would be management system.ai, which is StackAware's five-day email course on ISO402001. You can get more information in a written format, and it's a good reference for you if you're building your AI management system.
Cameron Gaudet: Perfect. Perfect. All right. Well, I that's that's what I had for today as far as the questions that came in, Walter. And I appreciate you taking the time to to to sit here and and answer these and discuss with myself. thank you again to all that have have continued to or or in this last month have contributed to the questions and comments and topics. I was thinking about it at in preparation for this, that I think with it it's important not just to have questions but to be asking kind of new questions, to be having new conversation as as this technology continues to advance week in, week out. if we're asking the same questions and then talking about it the same way, we're probably falling behind a little bit. And and it's tough to keep up with everything. But my advice is always just ask keep asking new questions. Keep finding new topics to discuss when it comes to this that'll that'll only kind of help us grow and and self-educate as a as a collective really. So appreciate all the questions, appreciate everybody's input. And thank you for listening. We'll catch you again next month on Aware AI Brief.
Walter Haydock: Thanks everyone.
