Generative AI: The Three Concerns Problem and A Tale of Three CISOs

thanks Damon um yeah I don't know why I got that reputation as the smartest person I I keep telling Bryson and others uh they need to meet more people but that said um I've had a chance to really um do some really interesting things in my career uh currently I'm now actually I left the ceso role and it's kind of actually part of the story here too The Tale of Three cesos um but I'm now a security Ambassador at juper 1 I used to be the ciso there um I used to be that Chief security scientist at B of America and while I was there I had a chance to create two things that uh

hopefully you guys have heard about in some respect or another if you were at the uh Chris Hoff's um keynote last year he actually mentioned these two he mentioned the Cyber defense Matrix which I brought copies of and I'm more than happy to give them away please come and see me if you want to get a copy as well as the DI Triad and I'm want to talk a little bit about those in context but um those are just two things that I'm generally known for um oh and and if I run out of books um there's a book signing that if you have one of those expensive uh business hall passes that you should really get for free um I'll

do a book signing over there as well all right now as I mentioned um I I don't really consider myself the smartest person in cyber security and I certainly would not put myself in the category for uh anything associated with this new found new founded new fangle technology by the way so I I hate calling it I hate using buzzword I prefer to just call it new fangled technology or nft and so if you hear me um just call it that just just make sure you understand why I'm calling it nft okay so in this uh new fangle technology I I'm I I think I've passed the peak of Mount stupid okay but I'm not that much further okay and so

I'm sure there's people who are way way smarter than I am more way more confident I am in in in this particular new space um but I try to think more deeply about what are the ramifications what are the things that we can look at um beyond the problem space that we might be facing and to kind of going and look at what uh Josh talked about yesterday he kind of gave that 10year recap so I'm looking for 10 years and that was kind of my role at Bank of America as the chief security scientist I said what kind of things should I be looking for 18 months out three years out uh so that we can be prepared to

tackle some of the challenges that we might face and so that's of the perspective I'm taking as well but I'm even though I said 10 years maybe it might take 10 years but I hope it's much sooner and for us to be prepared to be able to take on um the opportunities that come from this is what I would like to share here so as we look at um what's coming uh it seems like we're in this dichotomous uh time where it seems like it's the best of times and the worst of times and if you're in security wow it seems like we're in the worst of times we have uh a lot of challenges with how

employees use this technology how developers are building these uh Technologies and how attackers are weaponizing them okay and each of those poses um a real challenge for us and three you know Three core problems that we run into but I think I'm not not going to spend too much time on that because I'm sure um there's more than enough material on that okay and I'm not planning on sharing that piece as much but I'll give you a little quick snapshot of that just so you understand the context of how I'm thinking about it but the real question is how do we make this the best of times how do we take the opportunities that are presented

before us and really capitalize that uh for our careers for our industry and what we can do going forward okay so what are those opportunities and that's what I'm going to focus on primarily now as I talk about these best of time opportunities um you know it's it's somewhat Pro prognosticating okay as I mentioned I hope that these will come to pass and I have I've looked at this for a while to see the signs of it and I would say um well well you know I think one of the things I would look for is so key indicators that this is actually um trying to it's starting to happen and I've already started to see some of

these things happen so um we'll see if it turns out but uh I hope it's not well as I as I predict a future for you I hope it's not the too high fluting and high too Ivy Tower I think there's some real practical troops that can take away for some of these things all right so first let's talk about the worst of times and how employees use this new fangle technology well first of all is that we have a lot of uh fear uncertainty and doubt and as we have folks um using these Technologies we have people saying oh wait um it seems like it's spewing out the same intellectual property that we just put

in okay and I would say no no you got to understand this is not that's not how it works okay that's not how LM works and and the way I characterize it is that llms generate but they don't commemorate they generate new information they generate information but they don't commemorate uh information that's already in the system unless statistically it's it's highly prevalent uh Caleb Sima actually gave a great example if you type in um what is my Social Security number and you give it um the first five characters it won't be able to figure out the remaining five even if it was trained on your asso on a whole bunch of social security information okay so this perspective

that uh llms generate and not commemorate is one of the uh misconceptions or at least people thinking that uh we can um that it'll spin out spit out a whole bunch of information about intellectual property and so we're seeing a bunch of things happen in the industry for this I was going to put a bunch of logos of companies U but as you probably know the the curve for the number of companies being created around is like vertical okay so it's it's hard to keep up I said you know what it's not worth trying to capture that but there are tons of those Technologies and um I was also part of a a group that helped uh uh produce a

policy around how should we look at these this as a concern okay however that said okay I'm not going to uh hit that much because you guys can I'm sure you're already well versed in a lot of that where you can hear a lot of talks on it but I think there's an opportunity for us to again uh change the role so change how we look at our role and the opportunities that we have so how can we Elevate the ceso role in this new um new this new environment and so let me give you an analogy consider what a CFO does what is the CFO in charge of they are um they govern the wise and appropriate use of

money they allow businesses to essentially uh spend money to basically make more money right so that's the whole point of businesses right they don't actually make money too by the way they don't generate money I guess they can do fundraising but that's not really creating new money and if if a CFO said you can't spend any money you might as well fire them right they're not a really good CFO in that regard but what if the opportunity is for us to become the CFO for intellectual property so we have a bunch of intellectual property going out and we have these concerns around them but it's sort of like what if what if again it's a form of currency and we

spend currency to make better currency so what is what's the role then we govern the wise and appropriate use of intellectual property uh we allow them to spend this IP and this IP by the way has different you know I don't know if I want to put a dollar value to it I'm really deep well versed in the uh cyber risk quantification space and you know there's a whole bunch of stuff around that but I don't want to get into that let's just talk about it in the context of like are we talking about low denomination bills or high denomination bills okay are we talking like if I went to a CFO and asked hey I want to spend

$25 on something they will come to me and say what what are you asking me for I'm like just go swipe your credit card and go on what what is a low denomination intellectual property for us might I suggest for example our source code is low value denomination bills okay do I really need to ask for permission to transmit Snippets of source code now a lot of low denomination bills may add up to a high denomination bill so to speak and so we have to be you know figure out what what that sort of threshold is but the perspective is is we have a lot of intellectual property and if I said as a as a say so you cannot spend any of that

intellectual property well guess what you should probably fire me right and so what what if uh we change our role and the opportunity is for us to consider what what it looks like to be a CFO for intellectual property now with that comes actually some other interesting implications or interesting uh Concepts and with Finance and Accounting it's a very mature practice as you as you well know and in that context there is um lots of different things that we can borrow from that as well okay and we don't have those tools here but the these practices are things that we can probably figure out how we can adapt them to uh security as well and so I just pulled up a bunch

of terms in generally accepted accounting practices I'm going to call them security practices and uh here's an example impairment impairment is a financial term okay and it did some slight tweaking just to uh uh remove some of the finance specific words but if you read it it sounds like what we can do in security what is impairment it's when some sort of resource is impaired okay it's it's deated as impaired when I can no longer have any sort of of assurance that it can be fixed within a certain time frame okay sound like something we deal with on a regular basis right should we call it maybe impaired okay versus you know vulnerable or whatever else is yeah it pro maybe

okay and guess what when you have the term impairment there's also all these calculations that come into how we think about um like how do you calculate impairment cost assets we call them assets right it just seems natural except assets on a balance sheet is on the positive side of The Ledger but most of the assets that we deal in security are actually more liabilities okay do we actually have it on the right side of the Ledger uh and maybe we can start representing it as a as a liability and not as a quot to asset all of a sudden the business will see it as wait why are we carrying these liabilities not doing something about it

maybe we should try to get rid of these liabilities okay so again just slight wording change but you can see how that helps as well um survival metric so if you're in a startup or if you um if you're any Venture funded company uh there's whole bunch of metrics that we rely upon to see how much um time we have before we die as a company so they're called survival metrics uh what's your burn rate what's your Runway what's your turn and in the context of assets and how we look at liabilities I should say uh and and the way that we look at the resources that we have to be able to uh run a company um what H how

how does how does an impairment reduce our Runway how does an impairment increase churn and I'm not talking about churn in customers I'm talking about churn in other uh ways that we think about um how the these digital assets work as well um later on have a uh later today I have a talk on Double Entry accounting so what is double entry accounting it's a simple way to be able to have two ledgers two different systems provide a check against each other okay and I'll share some examples of that later today uh during a talk specifically on that so I'm not going to spend too much more time on that and then um EAA there's a

whole bunch there the industry the finance industry has been reshaped a lot because of the simple concept of iida and we can think about like what does it mean to have income in the context of cyber security we know what the term technical debt means right but how do we translate that into how it offsets this notion of income and the value that's being created by these assets that or these assets liability that we have see even I I still keep throwing myself off from these things but that's it the the premise here is that we have a set of practices that we can now uh try to cify within how we do cyber security and I

think the reason why this is particularly important is because I've been fairly deeply concerned and if you haven't been deeply concerned you should let you should be concerned about how the government is deciding some of these things for us okay and whether you agree or disagree with uh de Joe Sullivan's verdict um I I could put myself in his shoes and there are many T many cases where I would have probably done similar things as he would have done maybe a few things I wouldn't have done but nonetheless I can see how many of us can have fall into the same traps that he did and then Tim Brown with the wells notice that he got served I mean there's

there's a center of practice that they're assuming that we can't ever achieve okay so when we think about um when we think about for example some of these practices let's I I'll I'll talk about this in double during accounting how precise is accounting okay or how accurate are are the books how much variance do CFOs allow and guess what they do allow some variance okay it's not a perfect I mean the this the ledgers don't always match and yet they don't get sued well if it's a huge various they might but um within acceptable amounts they don't get sued they don't get fined they don't get these issues and it seems like that's not the same for us in cyber security

and so we have an opportunity to to well we have a couple opportunities one is to redefine our role not as a as a person that tries to secure as a technical weenie that tries to secure all these little things but rather in the sort of governance role of how we um manage and govern intellectual property and the institutional knowledge of the organization and we have these tools to help us well we need to come up with these tools and in doing so we end up with ways that we can potentially provide um guidance for the government and for us as a practice so that we can not uh deal with these in the future as

well all right so that's the first um best of times opportunity how can we become the CFO for intellectual property the second challenge that we deal with the second problem is developers building and um I'm I'm sure pretty much every company out there is now and you know using these new fangle technologies to try to do something with uh llms or generative AI um here's a a diagram that you can barely see because it's the wrong contrast um that uh that Andre and heret sent out it's a reference architecture for how you build llm uh applications and don't worry about the detail at this point but I do want to point out a couple things that uh in the context of

uh what we've always learned in security there are a couple inviable rules right um one you know never getet it into a land war with Asia land war in Asia but the one that's prob more important or well you know just slightly well less known but nonetheless important is to never trust user input and fundamentally one of the issues with uh one of the fundamental flaws with llms potentially fundamental flaws is that we can't separate out the control plane from the data plane all right so we know that I mean this is like a such a well-known uh uh principle in cyber security and yet when I go back to that chart that reference architecture everywhere that

I've highlighted in blue is user input unsanitized user input okay and it's pretty much everywhere right so okay um are we going to it seems like it's going to be a pretty bad thing if we build against this reference architecture that doesn't necessarily capture this uh core principle of um not trusting user input so uh how do we deal with this again there's there's tons of things out there I'm not going to uh spend too much time on them but just for reference you have things like Barryville institutes um Barryville Institute of machine learning U their taxonomy attacks I I love this particular one because it's the most straightforwards uh structured way to think about uh attacks against

machine learning of course many of youall seen the OAS top 10 there's mider Atlas so again I'm not going to spend too much time on those things but the but the perspective is that there's a lot of work that we're trying to do to to to address this problem which is all these places where we have um unsanitized user input and all these attack surfaces but I think there's an opportunity and to be able to explain that opportunity uh let me let me talk about Safety and Security something that actually Josh talks about often as well and by the way there's a a vendor out there integrity that's giving out this really big uh poster that says Safety

First and I love that because they're using the word safety now if you're not sure why I think that matters um here's the thing okay so if if you if you know um if you know Spanish then the word for safety is security and the word for security is security D so in Spanish we have one word for the same thing so for two different things in English we have two words and in cyber security we have the same word again so why don't we I mean in English why don't we call it something different because we have two different things that we do one that's called cyber safety and that one that's called cyber security and if you want to

understand the diff if you want to get a sense of what that distinction is we can apply um other context so let's take food so when we talk about food safety what are we talking about we're talking about things like hygiene compliance inspections good practices uh bill of materials having a sense of personal responsibility and when we talk about security we're think talking about things like starvation or like where's the Ukrainian weed or the baby formula and when people talk about security or rather safe um compliance doesn't equal security might it be because compliance is safety and safety doesn't equal security okay let me give you another example so uh airplanes if I'm an engineer at Boeing or at Airbus

my job is to ensure that the airplane stays up in the air doesn't come crashing to the ground pretty simple right my job is not to dodge Russian and Chinese missiles that is somebody else's job to make sure that the air space is free and or that that we have airspace security okay which is to have the space uh free and clear of Chinese and Russian missiles it's not my job rather it's somebody else's job usually the private uh the public sector right but the perspective here is that there's a activity that we do that most of us actually do that's actually safety oriented most of us do safety work okay and there are still some of youall that

do security work but just be clear that we do cyber safety more than we do cyber security okay uh by the way just real quick aside um s years ago Equifax got hit by a Chinese missile okay three years ago uh solarin got by hit by Russian missile all right but that event seven years ago as time passes on that Russian missile starts to look like a bird strike and that bird strike now is something that I'm responsible for okay if I'm designing an aircraft I need to make sure I can survive a bird strike um but I shouldn't be able to survive there should be no expectations that I can survive a missile strike at least most

for most organizations if you're apple and you manufacture iPhones you're probably building the equivalent of f-16s and you better be able to survive Russian missiles because guess what you're going to get those shot at you but nonetheless just the perspective of over time a Russian missile uh shot or a Chinese missile shot seven years ago is going to start looking like a bird sh and what does that burst strike look like now things like a software building materials okay um solar winds got hit by a Russian missile three years ago what's that going to look like uh four or five years from now you better have your supply your software supply chain um really secure because that's going to be

a bird strike as well so now with this perspective of us focusing on safety the opportunity for the ceso role is to think about us as the chief AI safety officer so you think that cyber security is a problem today guess what um there's going to be a much much bigger issue coming up okay and I think that it's going to cause all the things that we're dealing with in cyber security to pale and comparison and if you don't if you're not um if you're not sure as to why I'm saying this then I would recommend two books for you all to read through one is um life 3.0 by Max tedmar and the other

one's called human compatible by Stuart Russell very wellknown own researchers in the AI space but they give a perspective of what the future holds um where AI systems we're not the concern is not that AI systems are malevolent the concerns are that the AI systems are competent very competent so competent that uh we ask it to do something and it does it EXT extraordinarily well despite um what may violate our own value system so for example I would say uh I'm running late P me to the airport as quickly as you can and the this autonomous car will take me to the airport as quickly as it possibly can I might not survive I may kill people

along the way minimally I may be very nauseated okay which doesn't necessarily uhhere to my value system so how do we design these systems so that they're built safe um safely and responsibly so if you've heard all the things around responsible AI or um even the I think what you heard earlier about the C nitive science aspect of things this is all centered around this notion of AI safety and who better to take on the role of a chief AI safety officer than somebody who's been doing digital safety for years and years so I think that the opportunity for us going forward is to say hey you guys need an AI safety officer that's us we've been doing this

for a long long time now now what what does that mean to be an AI safety officer there's a lot of uh principles that do apply in cyber security um and we've learned some of these principles as well and I and I have a whole talk on this but I'm just going to hit the highlights on this but we've tried to do we've applied um these different new Fango Technologies in security as well tied it with Automation and have resulted in us uh basically shooting ourselves in the foot many times over okay um and so we've taken we we want to take those lessons learn where we've violated principles of um safety in some of these systems and

say okay how can we apply this to AI systems as well and so I offer like six principles that I put together um that I used when I was at um major financial institutions to figure out how do we uh set these sort of guard rails for ourselves as we built the build these systems how do we ensure for example that we know exactly what sort of inputs that we're putting into it how do we have ai building materials um how do we have conditions that are not unbounded but very tightly bounded how do I know um how can I ensure that um ultimately we want to make the system as deterministic as possible and to make it

as deterministic as possible we try to make things as bounded as possible we want to have thresholds when we know um we need to either take action or not not take action and and by the way the the sort of uh the examples I I used in this um when I came up with these were things like when do you block an IP address automatically when do you have an orchestration system take action on your behalf these are questions that I'm sure some youall are dealing with and these are the kind of conditions that I put forth when I said okay we will fully automate this activity once this this an analytic system figures out what's going

on okay and so these are sort of the uh the guard rails that we came up with because we wanted to make sure that things didn't go out of control okay and so fundamentally these are AI safety sort of principles which you will learn just by using these Technologies within the within the context of security so anyway other things like just making sure that you have uh you understand the the the processes that um are anticipated by The Operators having a kill switch uh and by the way on that kill switch piece let me mention uh the book human compatible has a really great thesis on how to uh build an a kill switch for these systems definitely

worth taking a closer look at and then making sure that there's somebody who owns this um this system so that you have somebody who's accountable for making sure that it doesn't kill everybody all right okay so that's those are the guard rails and then lastly um the third one is the third uh challenge is attackers weaponizing so we're going to have a lot of different ways that attackers weaponize some of these are are pretty welln you're going to see presumably more convincing uh fishing attacks you'll see people perhaps generate or the novices generate more malware um but for the most part I think those are things that um you know we've already been tackling those problems and that's

not too terribly new however I think in the context of what these Technologies are going to offer is ways that we can uh that attackers can potentially accelerate their uh ability to find vulnerabilities within our systems as well as novel ways to leverage things like deep fakes and so on and so forth and so for these two uh new types of uh attack factors that the attackers may use to we when they weaponize these Technologies I think we should think about what those May imply and so let me talk about those two for a moment the first two again just keep doing what you're doing it's not going to make I don't how much much more of a difference

it's going to make but that's it I think the bottom two are ones that I would be a bit more concerned about so with the bottom two though there is uh some opportunities here and to be able to explain that opportunity let me explain um something called The diw Pyramid so diw stands for data information knowledge wisdom and the premise here is that each layer provides more value than the layer below so um and I think for the most part we are now into what we call what I would call the knowledge economy we've been at the data economy we've been at the uh information economy now we're at the knowledge economy to give you a sense of how you can look at

this like data might be let's say websites information would be Google Chat GPT is at the knowledge layer okay and the the premise here is that um llms is is opening up this new economy the cognitive economy the knowledge economy and it will um give us the ability to access institutional knowledge like we haven't had before okay uh to give you another sense of how this to think about this we've been wanting Enterprise search for a long time and it's failed because what we get back is a bunch of data and information not actual institutional knowledge right we get back a bunch of documents that you're supposed to process no I want to get you

to give me the answer I don't want you to just give me a whole bunch of documents that I have to go read through what uh llms have opened up is this knowledge economy and with that perspective I would offer a view that the our understanding of how to build secure systems is already known it's just not widely distributed it's not evenly distributed but what if there were a system that helped distribute this knowledge more uniform normally what if there were a mechanism by which we could have an engineer who had a business problem um ask this engine and it spews out this the instructions for how to build this system now the engineer will never will rarely ask how

do I build this system securely right that's not really part of their equation but we built that in as a part of the answer okay and by the way the answer not may not may not necessarily to build it more securely but as I frame it down here to have fewer security concerns okay fewer security concerns and the reason why I say that is because the DI Triad if you're not familiar with that it's a concept that I came up with called um that stands for distributed immutable ephemeral and the idea is if I build systems to be distributed immutable and Emeral it allows us to reduce the amount of security concerns I would have so

even if the system were vulnerable even if the system we're even compromised for that matter um my security burden is lowered okay how do I build systems to either be secure by default or even better to not require security at all and that sort of design pattern is actually known some of it's not known and some some maybe maybe narrow edge cases but across the industry we do know that and guess where we share it we share it in conferences like this okay we share it in um c um coupon and CNC F and all these other places where where uh people are sharing these design patterns it's just not widely disseminated it's it's not easily

accessible this knowledge sits in all these videos and I have to go and pick them out myself but what if there were a way to be able to tap that and pull that out and so at this point we have the mechanisms to build uh more secure or systems that don't require security at all it's just a question of how do you assemble this together and so this perspective of uh that we saw earlier in terms of the uh issue with vulnerabilities what if are able to build systems that have fewer vulnerabilities to begin with OR systems for which even if it had a vulnerability I didn't care and then the second piece is around

the Deep fakes and so let me for that I have to explain yet another concept um so again I mentioned the Cyber defense Matrix earlier I'm sorry famili folks are familiar with the U or heard about the um sisa's uh zero trust maturity model um there's five pillars identity devices networks applications and workloads and data there's five function there's five asset classes in the Cyber defense Matrix devices apps networks data and users now if you wonder why there's five and why those um asset classes align pretty closely there's a reason behind that but unfortunately some of it get slightly misinterpreted so let me clarify some of the misinterpretation first of all uh identity is not just users okay that's

how we people typically think of it they think of identity as being users and if you look at the logo that's a user right that's what CIS is conveying so users and identity as far as they're looking at it are kind of synonymous okay but they're not why because all these asset classes have an identity devices have identity applications have identity networks have identity data has an identity okay and you're like okay well I think I understand the first three but what does it mean to have a data identity uh and by the way before I go to answer that question I should mention uh oh yes so not all all all these H classes have identity not just users I

should mention I thought it was kind of funny um back in 2016 I I shared this briefing of like where does uh governance and analytics and um and automation orchestration go and they pretty much put it under the same same sort of view as well so um a lot of this I think pretty much inspired by some of the work I did but there's some just slight slight variations I wanted to correct but that's it okay let's go back to data identity what is data identity and why is that a why is that a problem well let's go back to deep fakes think about deep fakes fundamentally the Deep fake issue is a data identity problem okay we have a

situation where we have content data and I don't have a way to authenticate whether it's well I don't have a way to determine it's authentic right now we've had this problem before okay and it's the source of all the issues we saw earlier with this these lures right it's fishing emails right and we've had this trouble with email for a long long time because well um we it's an ecosystem problem and we need an ecosystem we need the will of all these organizations to say we got to fix this problem but email doesn't seem like we got the will to get that done okay there's been a lot of efforts to try to get email to be more authenticated but

it still sucks and that's why we still have all these fishing emails and so on that are effective but there's a bigger problem that's emerging a much bigger problem and it's one that is of societal concern meaning there are people who care about this Beyond this room okay people outside of the security community and because it's a bigger deeper concern I believe that the will to fix it is going to be coming up okay so what what this look like perhaps it's a situation where the ecosystem says hey you know what if you take a picture with a device that has a um certificate on it signed by a major manufacturer then I can uh

validate that this picture was taken by this piece of hardware and it's been unadulterated okay and you've seen already uh where places where people have tagged it to say this is AI generated but what I'm looking for is something that says no this is authentic right and I'm looking for authentic video I'm looking for authentic uh photos I'm looking for authentic audio now if I solve that problem what is email right I mean that's trivial email is Trivial compared to video audio images okay and if they can if we can find a way to solve that problem might we get rid of the bigger problem that we've had that we've had in security for such a long

time and so that's I think that's the perspective to look for in the future which is with the bottom one you have fewer systems uh systems built with fewer security concerns and then the bottom one being where even email becomes a trivial we can solve the the authenticated email problem which really gets rid of some of the problems above as well all right so to wrap up I think the best of times for number three is not elevating the role but eliminating the role how do we eliminate the role of the ciso because I think if we have a way to build systems with fewer security concerns if we have a way to solve some of the thorus problems

that we've had in security well you know my job as a ciso may not necessarily be as important anymore and so I would love to have this role shrink and what's interesting is uh Ryan McAn he's a uh he used to work on Netflix he's an advisor to a lot of startups and he states that there's a lot of security companies I not security there's a lot of uh new companies that don't even need a SEO because they've already um built systems that don't require that have a much lower reduced risk profile and that's partly because they build against the D Triad that you saw before but I think it's also because again they're building

against better design patterns and so I think the opportunity for us is to look at so as our job goes away we have other opportunities and one is to become the CFO for intellectual property and the other one is to be the chief AI safety officer so with that thank you very much [Applause]

take hello okay great uh we have time for questions thank you for the talk one last thing be a we're going to move this back away from that projector because it's got a bad leg and if you touch it it's going to fall over and go boom thank you for the talk I have a nuanced uh question in the spirit of deliberately feeding Roo's basilisk and your presentation seem to focus primarily on the business um user the engineer the technologist but my question is really for you on this room how how do you change or amend these principles with the dark sweet Embrace from the security Enterprise professional deliberately using llm for the purposes of

security and when we're making decisions around informations that's coming out of llm for security purpos purposes how do we need to be thinking about building our own tools and our own practices for infosec and not just thinking of that sort of Us and Them towards the uh user towards the business uh um aspect and towards the uh developer and engineer sure great question so let me answer it in two ways first is the the business alignment um you're a spot on in terms of the way I'm characterizing the future role is highly business aligned and that's been a flaw of how we thought about security for a long time that we didn't we weren't business aligned so

the real goal as I see it in the future is for us to be highly business aligned that said we are here today and we're dealing with um we're we're trying to leverage these Technologies to to deal with the present problems uh the problems that are present and so I didn't mention this as a way that we can think about uh how we leverage some of the Technologies and what sort of safeguards we can put in the other question there was another facet to your question is how are security teams is using llms today okay and I deliberately skipped that question because um I spend I I spend time I spend time with startups talking to startups about what

they do and I I haven't run into a single one that isn't using llm in some way to help address our security problem so um I I deliberately skipped that question because I I we're just I'm inundated by that I don't know if you guys are but nearly every startup I talk to has something associated with that so uh if you want to get more specific details I'm happy to share that because I've been I have a lot of it in my head I just didn't think it was worth spending time on here because really just go out to any vendor out there and they'll tell you good to see you again good to seeing

you um there's an industry effort um in the font industry monotype Adobe Microsoft Etc around authenticity and we're we're building a standard around it because it's a problem that needs to be solved for the benefit not only of just the font industry but uh larger Industries um I would encourage everybody in this room to research about standards emerging regarding authenticity and to start checking around the problem of how you validate authenticity independent of authentication they are separate problems that require independent Solutions and authentication is a path towards validating authenticity but not the only one and and it's I'm not just splitting hairs here the differences of Concepts matter and solving the problem of authenticity um is one of the biggest

problems we have to solve in order for our profession to thrive much like solving the issue of control or utility remain unsolved problems and undiscussed problems in our industry yeah I I don't know if I can uh understand the semantic difference I think I have some sense of the semantic difference between authenticity and authentication um so we can have a side off sideline discussion on that I think fundamentally what the premise I would still offer is that I I would I would push forward the notion that we know how to solve the problem we just haven't uh built up the will to solve the problem even if it's to say you know what we're just going to create

a deao standard to do this right and whether you're apple or Microsoft or Google they can they can do something that will drive that uh that sort of adoption even if we ended up with two three different ways to to do it yeah and it's it's in the works it is happening thanks your talk um so I'm more worried about um veracity as instead of authenticity um so you know I'm sure you're aware of the case lately where um uh AI generated a legal brief had eight different eight different uh case laws cided there were fake cases that made them out a whole cloth I've had conversations with it with where it tries to convince me of stuff that I

know personally didn't happen um what kind of guard rails do you envision around those kinds of things because I see us kind it's like it wants to appease us so it's going to give us the answers and the words that we want to hear yeah but there's not any real good way for us to authenticate the the the ver veracity of it right now nobody's even thinking about that problem I don't think are they um so I have a model for that oh good uh it's called the uh ew pyramid so llms are offer at the knowledge level and what sits above it wisdom right and so I think there's a and we don't have a machine that will go

and apply that wisdom so um that's still up to us to apply that wisdom on top of whatever anyone shares with us right so I'm I'm sharing to you knowledge I expect you to take your wisdom engine and discern whether or not what I'm telling you is um uh has the veracity that you're looking for right so that's something to consider um by the way on the on the topic of hallucinations uh there's a wonderful case that you should read about uh with this girl who was undergoing uh open brain surgery and she was conscious at the time the ne neurosurgeon zapped her brain she started to laugh this neurosurgeon asked why' you laugh we know the answer but

she looked based on the sensory input that she was getting her brain made a statistical guess as to what was the cause and the cause was oh it's because of what you're wearing or zap it's because of the photo on the on the wall or zap or it's because of something in other words her brain was hallucinating statistically fascinating story but anyway all right how do you see organizations addressing data compliance with AI because not only are you've dealing with the challenge uh now of users willingly giving up IP and information to public sources thinking that they're helping the company you're not just fighting Rogue actors that may be exfiltrating data so how do you see

companies thinking about this differently because currently it's still a fight to get people to maintain compliance and think about data compliance um even industries that fall under SEC finra um you know FDA part 11 stuff those orgs really you know they they fight compliance on that as much as they can because they feel like it hinders operations in the business so just curious how you see IA CH or AI changing that or whether or not that's something you think is actually going to change or if it's going to be the same fight with a new thing yeah it it's fundamentally a different type of problem so let me explain why so data governance information governance

knowledge governance okay uh let's look at access controls knowledge uh data Access Control controls information access controls knowledge access controls data access controls can be done at a very discret level you have access to this document you have access to this database this row whatever yes or no very binary no information is a is more abstract it's confidential what what is what what do you mean by confid like what's the specific thing that's confidential right uh or secret or tops or whatever um this perspective of abstracted data is where information off uh operates and what is information governance at that scale it's a little bit more intangible right then you have knowledge which is even more abstract

and intangible um so and and we seen this already happen with llms that seemingly uh take copyright reproduce copyrighted material right by the way we in our conversations reproduce copyright material all the time and we don't site the sources it's it's called education okay and so how do we know what the sources are we can't actually because it's so from so many different pieces of uh from so many pieces of data and information that gets assimilated into this notion of knowledge so it is a it's a whole different problem um and not I don't think I can give you the answer because I don't know the answer but I can pose the the the entirely different

nature of the problem itself thank you um separating the AI for a moment your vision of a chief intellectual property know CFO alternative is nice for trade secrets and intellectual property but entire new branch of Officer the chief product security officer for the things you put into the world or in OT and IC environments it's availability and cyber physical safety roles for ciso so have you put some of these models to product security and in operational cyber physical system security their Futures and their evolutions or is this a fork in the tree of a system uh the view that I would have is whatever you're building um whatever you're building if there if the understanding

of how to build for example a uh more either more secure or a system that doesn't require security at all it system OT system IC System whatever else is um again I will postulate and maybe you can tell me if I'm wrong but that is known it's a known quantity is that a fair is that a fair statement it's less mature well I I don't disagree but somebody knows oh this is how you should build a system that is not just secure or doesn't require security at all but rather also works with the business it actually achieves the business goals okay how do we get that knowledge more widely disseminated write a book about it or

have something that allow that Taps that understanding right and that's what I think the future holds now that understanding has to be properly verified that is that is correct okay uh but nonetheless we have a we have a leg up now and being able to say okay here's a design that may actually be the better design thank you sir for doing this huge fan you know one of the criticisms I've heard in both the professional world and the education world is that llms will give false answers and my response is always well I get told the wrong thing every day uh you know without llms going back to your diw model I'm curious you know one of the things that I've seen is

where there's a masking of ignorance where suddenly an analyst who wasn't super smart yesterday as suddenly an expert in has filled where do you see organizations like like what are some safeties or guard rails or or ways to expose where we might have these false flag experts who really don't truly understand the material but because of this introduction of new knowledge and being able to basically have a tutor on demand to which I'm all for by the way but you know this this kind of hiding of ignorance and subterfuge of expertis if that makes sense yeah and I don't know if there's I don't know where the solution is going to emerge um perhaps it's sort of like uh if I if I wanna if

I want to teach um if I want to teach my kids well then am I going to fill them with a whole bunch will I give them access to the internet and YouTube and say go learn from there or will I say would I have a curated set of content that I know to be true right that this is verified by scientists by um people who've actually like you know there there's and there may be disagreement right not I'm not I'm not saying that there's disagreement things but the perspective of you know this is foundation knowledge that everyone should have okay train up on that right uh and that's what we would try to that's what we we kind of need to do it

goes back to an earlier point I made around the U the sources right so if you want to have um uh like this AI Bill and materials is something that we're actually trying to figure out like how do we establish that because I don't want that to include poison sure wonderful thank you so anyway hopefully that answers your question I don't know how we purge it yet but that's that's the premise I'm here to thank you for your talk and um say it's time to move on to the next one all right

cool