Walter Haydock: Welcome to the Deploy Security Podcast, where we go deep on the privacy, security, and compliance implications of using artificial intelligence. I'm your host, Walter Hadock, the founder of StackAware. Today I'm joined by Mario Platt, who's the Chief Information Security Officer at LastPass. He's been a security leader for more than 20 years, spanning penetration testing, security operations, security engineering, governance, and executive leadership. As CISO at LastPass and a former consultant in medtech, fintech, and crypto, he brings relevant perspective on protecting sensitive data, managing security risk, and operationalizing governance in highly regulated environments. Mario, welcome to the show.
Mario Platt: Thank you very much for having me. Great to be here.
Walter Haydock: So can you for for those folks who aren't familiar with LastPass, can you talk a little bit more about the company and its mission?
Mario Platt: Yep. so last pass we started as as a password manager, like the the first password manager, both in B to C and B to C and B to B. and with the goal of really making work safe and simple. that's still true, but we now offer much more capabilities in terms of the the platform that we provide. so this week alone we Started a few months back getting onto the area of SaaS access protection. and this week we've released some new product features on AI protection and also SaaS protection to enable clients to to get policies on how they're accessing their AI and SaaS and SaaS applications overall. so we're not just about passwords anymore. and our goal is really to focus on the needs of largely the SMB and the mid-market, so teams that may not have the the privilege of of having big security teams like myself, to to make sure that they can manage risk in ways that that that they can afford and in ways that they can have their own IT teams or small security teams managed for them.
Walter Haydock: Awesome. Awesome. Thanks for that background. So as CISO, obviously every CISO needs to deal with AI these days. And in different organizations, leadership might be pushing speed of deployment over some other concerns related to security, privacy, controls, things like that. How are you balancing that at LastPass right now?
Mario Platt: Yep. so that also goes into my into my experience, especially with talking with other peers in the in the industry. so my I I think my first comment would be that generally a lot of companies are trying to govern what they can't see. and that's a challenge, right? kind of hope is not a strategy, that's wishful thinking. so I I think we took governance overall I think really needs to be seen as an evolutionary practice. Now I'm in the in in the in the benefit or I have the privilege that we started slow, right? so we actually started by by blocking access to all of AI applications whilst we figured out what controls we need, how do you manage the related threats. And then over time we started more opening up and enabling the business a lot more. so we actually have that that experience that we had internally from going very, very slow, which was not good for innovation for all of the benefits that that business is trying to get to AI. Where we're now kind of on the on the path of now that we understand the risks and my the my own security team has started to learn what's acceptable, not acceptable and the things that that do not pose such risk to the company. We're now in the process of accelerating all of that much more. and I think and ideally that's what That's not what a lot of teams out there have to deal with. A lot of my peers they woke up one day and everyone was using insection AI, and that's a much much more complicated problem to solve that that I'm pr again privileged that that's not what I what I had to deal with. but but but really what you what you said is absolutely correct. that we need to do to really think how about we're enabling the business. At the end of the day, we're doing this with pursuing a lot of the benefits that the technology is is warranting, and we need to be cognizant of that. of both what experimentation looks like such that we can enable our stakeholders to to test the benefits that it could provide and also having processes that are not overly prescriptive, long, etc. that can help support adoption at speed within the organization.
Walter Haydock: Yeah, I think the point about trying to govern what you can't see is really key because I always recommend starting with a with an asset inventory at the beginning to understand exactly what's in use. You know, that can take the form of, you know, looking at spreadsheets, can you can look at finance statements, see who you're paying for AI tools. It can take the form of of technical approaches and you know, happen to hear what LastPass is doing from from a technical perspective on AI discovery, if if you're open to sharing.
Mario Platt: Yep. so the from now on we're gonna be using a lot more of our features the of our product, etc. That's one of the things that we that we're we're gonna help companies with, but that's really new. so generally we have obviously things like web proxies, like our procurement systems, etc., that we first use to to get a sense of what was out there. and then part of our procurement process, we introduced routines to kind of understand if as contracts were Coming through for our biggest key risk suppliers, what was their use of AI? Because that's another part of the challenge, right? A lot of us are dealing with the fact that we wake up one morning and suddenly one of our vendors enabled their on their product without telling us it's just there now, right? And we had no opportunity to review to so that's the I think the reality that that that everyone can relate to. so it's really been about as you said, leveraging existing inventories and the data that we had on the systems. ideally cross-checking those with web web logs so we can see what our employees are doing from from their devices and and correlate with yeah with what what's actually in use and not just what we believe is on the procurement system which yep which is not does not always tell the same story and both relate to to effective governance and security.
Walter Haydock: Yeah. And the kind of you wake up and you find something is enabled line of events. I remember there was a a Slack feature push where they said AI is enabled for your full workspace. And there was like no notification. They wasn't in the admin release notes or anything like that. I know a bunch of bunch of CISOs that I always that I that I talk with, they say, Well, hold on, what's going here? yeah, I I didn't know I didn't know this was coming. So definitely definitely a real problem. Yeah, yeah.
Mario Platt: Quite a few vendors. Quite a few vendors and not all of them give you the opportunity to opt out, right? it depending on their strategy, so it becomes
Walter Haydock: Yes, yes, yes.
Mario Platt: complicated to help manage to help the business manage risk.
Walter Haydock: Yes. Yes. So on that note, we we mentioned kind of the the surprise AI feature side of things. What are what's the biggest blind spot that you see organizations run into when they're adopting generative AI? Is it, you know, shadow AI? We talked about that. Is it leakage? Is it model drift? What what what's the biggest risk you see or blind spot?
Mario Platt: Yep. F from so again I'm gonna call out my bias. I believe that shadow AI I think is one of the biggest challenges, for a number of different reasons. and also this depends on the maturity of controls that people had, right? one of the things that I don't think is talked about enough in the industry is how in my opinion, people that were not mature in managing SAS application risk, it became so much more compounded with AI, right? because the the the channels of access are are essentially the same, right? And if we we know for instance from so one of the things that we've released a couple of weeks ago, LastPass, that I would invite everyone to to have a look is the state of AI and SAS security report 2026. Right. And so there are some some specific statistics there that I think are are scary, right? Or at least they scare me, right? And that that talks about this compounding of of risk between not managing the SAS. Component and how that exacerbates the AI exposure risks to a lot of us. So knowing that a lot of companies don't have the maturity to know if people are using a corporate or versus a personal ChatGPT account, for instance, or Anthropic or whatever you may decide to use. So we know from anonymized data from our platform, for instance, that almost 53% of users they reuse passwords across accounts. We know that about 21% of Users, they've got some of their accounts have been in non-breaches that they haven't necessarily changed. So if you compound that with the fact that a lot of users are using their personal accounts on ChatGPTs and Athropics to deal with with customer data in their businesses, if you combine these three things together, that's a pretty scary picture, right? and I think that's that's one element where yeah the the SAS and the AI risk has has compounded each other, right? companies that that didn't do that. And and again goes back to the how I started, right? if we we we can't govern what we can't see. So really having that that visibility of what people are doing, are they using the the appropriate accounts? do we understand what people are putting into these tools? which isn't necessarily about about blocking, right? And that's a challenge that we security industry have had for many, many years. we always think that the only way to to reduce risk is to block things. But I think what the industry and what executive teams are telling us, what industry at large is telling us is there's actually a lot of risk tolerance for for AI-related risks, right? People want to pursue the benefits, right? That's just what what every executive team right now has mandates on improving efficiencies, etc., through the use of AI. So the risk tolerance exists. Our job as security people then I think becomes less about how am I going to block all of these things, but but more but a question that That I don't see security teams asking enough is how do I nudge users to the types of behaviors that I would like to see, right? That are less risky behaviors, etc. Right. And I think the the this idea of of nudging users, which is part of how we've developed the AI protect product at last plus that we can decide to block if we want to, or just warn a user. So you're about to put some confidential information here, are you sure you want to do that? Right? Some because that may not, and even in prompting with Chat GPTs and the clouds of this world. having one thing that we've done internally at last pass was build our own custom plugins into an into cloud right that check our security policies and try to nudge users. Look if you do it this way then you're meeting all of the security policies, right? This type of nudging behavior is something that I I don't see most security teams see the value in what they are doing because in the the arguments that I usually get are valid, which is look, an adversarial a sufficiently adversarial person will be able to go around them. True, but most of your employees they're not adversaries, right? they're not necessarily trying to deliberately circumvent your policies. So if all we're doing is only solving for the problem of adversarial, then you actually got a business culture problem, not a security one. It affects security, right? But that's a bigger problem if everyone follows your employees are trying to do that deliberately. And that's not the case for most companies, right? Most humans they want to do the right thing, they also want to do the quickest thing. So if you figure out ways to introduce those nudges. at the point where people are interacting with these tools and are trying to achieve the benefits, I think we are helping the company reduce the risk, right? Because no no employee wants to be the one that was in the in the path of a data breach. And I think that's that you yep we need security teams need to be more willing to to adopt those types of strategies in the name of risk reduction and not risk elimination.
Walter Haydock: I think the point about nudging employees who are trying to do the right thing and especially nuanced circumstances is really important. You mentioned, you know, logging into a personal Chat GP account account versus a corporate one. And employees might not even know that they're logged in on their personal account while they start doing something, but there are real differences in terms of how that data is handled. You know, if you've got a Chat GPT enterprise account, you can, you know, you there's no training on that data, you can set retention periods, you can purge it if it's
Mario Platt: Yes.
Walter Haydock: if necessary. If it's a personal account, not so much. So controlling that login behavior is it's it's very nuanced and it's great to hear that that you're able to do that because that can have serious impacts from from a data confidentiality perspective.
Mario Platt: Mm-hmm. One hundred percent.
Walter Haydock: So on a somewhat higher level, can you talk about what an effective AI governance framework looks like in practice rather than on paper? So you mentioned policies that can be long and and wordy. you know, that's something I see commonly are are are are policies that have great aspirations, but maybe the organization has trouble meeting those requirements. How do you how do you make sure that you can operationalize AI governance?
Mario Platt: Yeah. That's a great question. It's actually something that we I'm currently in the process of of evolving. So one of the elements that in and again going being consistent with my earlier comment, right, that as not being able to govern what what we cannot see. one of the elements that that that that is the direction of travel that we're working on at the moment is really having a focus on what I'm calling a policy enforcement architecture, right? So we've got a bunch of places where we can we can introduce policies. at the network level, at the browser level. start with LastPass at the plug-in level, at the levels of before we add plugins to the market, are we considering prompt injection and and those other things that that the engineering teams or the the the business teams are doing right so considering all of these points where we've got some level of control where we could get policy even at runtime with behaviors that we can do with through cloud or or or chat gpt is really kind of first having a view of what the what what we care about right what our way that are ideally threat driven that drives so this requirement now applies because you answered yes to this type of thing that can pose a threat to the business and then having those in policy enforcement points being very clear of we apply this control over here in this interaction in the other interaction and make sure that we don't build policies where we cannot say this is where it's going to be enforced right so that's that's one of the the the elements of again going back to my previous point. I had the privilege of starting slow and because we now see okay, now we know where the all the controls are, we can now start being start moving the model to the policy enforcement points being something that we do from a technical point of view and and that technical par part of the enforcement and nudging the users at the points of interaction as opposed to again something that may live on a policy document that people acknowledge once a year that may or may not affect their behaviour on Yeah.
Walter Haydock: Yeah, if an administrative control is only administrative, then I would apply a a very heavy discount to its effective its effectiveness. So, you know, train training certainly helpful and I I would absolutely encourage folks to do that. But if you know it's all it is, like you said, is just one time a year acknowledging that you're not going to do XYZ, then it's pretty hard to really ensure any sort of enforcement. So if you can turn those policy statements into actionable, enforceable controls, that's where you really achieve the the mitigation. So, you know, like like you said,
Mario Platt: Yeah.
Walter Haydock: if you're if you're at the point of entering certain types of data into a certain interface, then, you know, that that's where that's where the rubber meets the road and that's where you can really make a difference. So so absolutely
Mario Platt: Yeah.
Walter Haydock: agree on that.
Mario Platt: And extend and even I would go even a bit further, which is extending that that component which from a risk point of view to understand what are things that really need to be enforced, you know, in people should not be able to bypass, in things that we should just say, look, be careful, minimize this data, type of warning. But look if you've got a business reason and etcetera and you say, Okay, I've I I tried, there's no other way to do this, then if that's within the risk tolerance of your business, it's not regulated data, those types of things that obviously companies need to to be much more careful about, then it's about making sure that the there is tolerance of the business gets translated into those policy enforcement points, right? And that we're still nudging the users in the be in the directions of the behaviors we want to see, but not necessarily kind of stopping something that would then generate a ticket that someone needs to go on the queue and see. So yeah, and have lengthy processes where humans need to to discuss and argue and takes a long time. So understanding that is key for security to be perceived by the business. as as a helping reduce the risk but also helping enable in the experimentation etc which is what all companies are being asked to asked to do.
Walter Haydock: I think if you have clear policies, then you avoid both of those failure modes that that you talked about. So one, you, you know, the folks won't be doing things that they shouldn't be doing, but also they won't be uncertain if they're doing something that is authorized by the business. There won't be any any uncertainty. So if they can go to a register and say, hey, I can use Claude Code to, you know, generate prototypes. I can, you know, do this. If I apply a human review to the outputs, then you know, go ahead, experiment. create create new new products, features, things like that that can accelerate the business while staying within the risk appetite of the organization.
Mario Platt: Yep. One hundred percent.
Walter Haydock: So on an external perspective, how are you at LastPass, but also how are you helping your customers adapt to regulatory developments like the EU AI Act, which obviously has gone through some changes recently, and also in the United States, the proliferation of the state level AI laws that in you know in some cases are more restrictive in in my view than the EU AI Act. But you know, it's a it's a very varied playing field here on this side of the Atlantic. So Can you talk about how you're addressing that?
Mario Platt: Yep. So I think there's first I think the the key thing to have is really good partnerships between the security and privacy functions, right? I I think that is key in terms of figuring out what's the best way to to deal with those. Having said that I I think the the same advice that would apply to any other security thing kind of still applies. If we if we build our governance frameworks around I think there are for me there are two kind of good standards to to kind of consider which yep is yeah obviously Stack Aware you helped us with building our our initial one and we're about to do some more work to to continue that evolution. But I think basing those on ISO forty two thousand and one in the NIST AIRMF. Right in RMF, that's the best thing that we can do. Because then all of the level of of specific laws, etc., particularly for companies that are not part of what many of those regulations call high risk systems, right? So if you're not dealing with credit scoring, with employment, with with all of those types of use cases that obviously have different types of of considerations, then it's a good baseline to then expand as as those regulations pop up like mushrooms at times to kind of r adapt some of your practices to make sure that you're adhering to to those but if you don't have that that strong baseline based on a standard that you can evidence that that you're basing off of good practice then it becomes really complicated and you run the risk of always be chasing the the the the regulation right and as you said particularly in the US it's a mishmash of of approaches even right so I think Colorado is a bit more like GDPR but then you've got Texas that it's much more about the the harm and incitation and deep fakes and so cuss consumer protection type stuff right so they're all taking very different lenses into how how they even look at what needs to be regulated in the context of AI. so it's definitely a complicated landscape and it's not gonna get easier, right?
Walter Haydock: I think the reference to ISO 40 2001 is really important there because if you're doing it correctly, it gives you a way to consider all these regulatory legislative requirements like the EU AI Act, the state level laws like Colorado, Texas, what have you. And then drive setting of objectives, creating an AI policy, doing risk assessments. All this stuff is driven by the external context of the organization. And It's it's kind of easy to try to check the box on that and say, like, you know, regulatory concerns are an external issue, which, you know, yes, I agree that those are those are external issues. But if you get much more specific, then you can say, okay, we're using this AI system in this use case. Let's evaluate that against EU AI Act, Colorado SB 189, you know, California CCPA ADMT regulations. Okay, well, it triggers. These requirements, but not those. We need to apply these controls, but not those. It gives you a very nuanced way to attack those types of requirements.
Mario Platt: Yeah. And again if you don't have that that strong standards based approach then okay now we need to do this for this one and that for that one. You don't have the basis where you can bolt on and extend as as need needs arise.
Walter Haydock: Yep. So obviously you're helping customers deal with AI security risk management. You're dealing with it internally. But then your customers are probably looking at you saying, okay, well, now you're using AI. Obviously, you know, you've you've described a comprehensive approach to it, how you started slow. But how are you responding to your customer scrutiny of LastPass regarding use of AI? What what is the state of sophistication you're seeing in the questions you get? you know, does it vary or is it pretty consistent? What are what are you seeing?
Mario Platt: So we see a bit of all of the above type of thing. so we see a lot of customers that aren't particularly asking questions, we see a lot of people for their own needs and they're they're asking many more questions. So we're seeing particularly coming through procurement process in in bigger deals, etcetera, references or asks relating to how we map our controls against again RMF, and if we are 42,000 certified, which is not something that we have yet. We're about to start a gap analysis that Stack Aware yourself are going to help us with, which I expect us to to do early next year. so we're currently in the process of making sure we already know and have the assurance that we built the program to begin with to use that as a baseline, right? Already because we were thinking about about eventually getting to certification, like we have 2701, 27701 right now. and yep, and now it's going to be that next level to to make Make sure that next year we can start providing some auditable credentials to to help facilitate the procurement process.
Walter Haydock: Awesome. Yeah, I've I've seen a wide variety of sophistication when it comes to customer inquiries about AI. I've seen, you know, hyper specific questions about, you know, how do you categorize yourself you know based on a certain law. you know, I've seen no questions asked. And then I've kind of seen a little bit of a a a muddled middle where folks kind of know that they should be concerned about. AI
Mario Platt: Yeah.
Walter Haydock: governance and security, but they're not exactly sure what they should be concerned about. So then you kind of get some some oddball questions in there that it's hard to answer. you know, I I I I've seen questions say, you know, confirm that there are no there's no bias in any of your AI systems, which which I would
Mario Platt: Ha ha ha.
Walter Haydock: advise, you know, okay, well, we'll hold on. We I mean the only way to confirm that is to just delete all the AI systems, then there's no bias because you know, that's how AI works is through a bias variance trade-off. there's there's no way around that. So, and then I've seen other kind of very, you know, very large enterprises try to lay down requirements on organizations. You know, one AI powered healthcare company I was working with, of a customer said, You're not allowed to use AI for healthcare. To which I would say, I I advise my customer, I'd say, You should just ask them, you know, what what should you be doing with with AI if you're if you're not doing that? Because you're an AI powered healthcare company. So Yeah, I see a a wide range of of sophistication in terms of customer concerns about about AI.
Mario Platt: Yeah, I've had the that other one as well. But so one thing for instance that I've been that I've been that also is driven by by risk and particularly reputational risk, in in my opinion. So there's a difference between a company providing using AI on their product, but but but it's something that they use on their back end, etcetera, that is kind of generally self contained and it's about optimization of the product features, right? That's one type of use case. one specific question that I as a CISO am asking some of potential suppliers That would particularly provide AI features that are customer-facing, that's a different thing, right? So one of the questions that I'm starting to ask, and I might maybe one of those let me gently call it obnoxious people that asks kind of random questions that seem that seem random, but is do you have an internal red team? Are you red teaming those those components, right? Because I think it's a very legitimate question because if it's something that's going to be that my customers will potentially be interacting with, I want to know how you are testing. that people can't say please ignore instructions and now do this instead etc because at the end of the day that's business reputation right so on any new supplier that i'm that yeah that potentially has the the the opportunity or the architecture to be customer facing, I'm being much more prescriptive in asking those more detailed questions because look, I wanna see those reports, I wanna see exactly how you're testing this, because yep. we don't want things that happen to other companies like Air Canada and that a others in the past, right? Where those questions were likely not asked and then you get yeah, you wake up one morning and you're having a type of impact you weren't expecting, right? So at least to do what we can from the diligence point of view to reduce to reduce that likelihood.
Walter Haydock: AI red teaming is a really great control because it shows, you know, what what a motivated and intelligent person can do when they're trying to force your AI system to misbehave. And you know, it's one thing to have design documents and you know, apply your your traditional security controls, but it's another thing to see in the wild what is your system doing when somebody who's who's motivated and an expert trying to trying to break it. So definitely agree.
Mario Platt: And if it's a in if it's a customer facing f system, people have all the time in the world to try and break it. Right. They're
Walter Haydock: Yep. Yep.
Mario Platt: not constrained by time. They can come back tomorrow and be another three hours at it. Right? that's yeah.
Walter Haydock: Yep. Exactly. Exactly. So in terms of the steps that you've taken, obviously we talked about identification, asset inventory, we talked about actionable controls. In terms of return on investment, what's the top thing that you've seen that you've implemented or you've seen other organizations implement? And this doesn't necessarily need to be the most impactful, but it's from an ROI perspective. What's the easiest switch you could flip? From a security and governance perspective.
Mario Platt: I'm not sure it's a a switch, but I I think the the the greatest ROI that I had, so I would separate that between governance and security, right? kind of both are important components. They relate to each other, but not necessarily the same thing. So from a from a governance point of view, personally I think was leveraging the the API that Stackware provides, right? in terms of AI risk assessments, etc. So the my teams were were spending circa eight hours, we didn't not always had Some of our suppliers have answers to the questions we were posing. A lot of things were getting stuck in the process. Stakeholders upset because we are not finishing the process, because we're not getting the answers that we need. So kind of adopting the Stackware API where we can self-serve a lot of the risk assessments, moved some of our of our internal governance processes from being eight hours to being closer to two or or less than that. Right. So that was an amazing ROI from internal stakeholder kind of perspective and yep. And seeing that we're trying to do our best to accelerate the process, right, while still getting really high quality results on data that's been vetted. So from a governance point of view, I would point to that. from a security perspective, as I mentioned previously, was us developing our own custom plugins for Anthropic, right? again, nudging users' behaviors in right directions. now we've got that integrated with with our MDR service and all of that type of thing. but but but that was yeah, creating our own plugins that direct users towards the right behaviors that does we use pre-use to we use hooks for some of the things that we want to determine to to do on a deterministic fashion within that plugin as well as opposed to leaving it to LLM based judgment so that's been the the the greatest ROI because I can I know that I'm nudging my users at the point they're interacting with systems towards the behaviors that I want to see.
Walter Haydock: Yeah, I mean you can really customize these these tools like ChatGPT and Claude to basically an infinite degree that really meet your business requirements. And I think it's it's great to hear that you're you're leveraging those capabilities, building building in-house tools to accelerate your your security In terms of the high level, so boards and executive teams, we talked a little bit about the executive side, but but at the board level, what should boards be asking about AI risk that they are not asking today? Because, you know, I'm sure folks are getting pressure from their boards to adopt AI. But what questions should be getting asked that aren't about the risk side of things at
Mario Platt: Yeah. so I'm again lucky that I I get asked very specific questions about that subject, so and and I I have to report on them. obviously the There's an element that I think is going back to to how we even think about the policies themselves, right? So how are we enforcing policies, right? So across not only u utilization of kind of chat based tools for lack of for lack of a better word, but it's also the element of how we're integrating business systems with our capabilities, right? So between those two types of things there's a lot that can go wrong and there's a lot of places where policies should be enforced and maybe aren't. So I I think Boards asking more about specifically okay, you've got these AI policies, I can see that they're written, how are you enforcing them, right? How does that affect when someone gets to their laptop and accesses a website, when they try to paste data into into a chat tool, when when you bring a new supplier to the business, how are you identifying do you have the logging capabilities to know how people are interacting with it? Do you know what type of data is being is being transferred, how people are making use of it, how are they being retained? So So all of those those are two components where I think it requires maybe a level of s sophistication, right? But but but it's questions that that that boards in the executive team should be asking because at the end of the day, again it's how we transfer it from being kind of wishful thinking, we wrote some policies and and we're hoping that people are doing the right things, to actually do we have the visibility in the policy enforcement points as people are doing these things, to to know if policies are being adhered to or not. Right.
Walter Haydock: Got it. So focusing on the operationalization of the AI security program through metrics, through technical controls, that's something that boards should be asking, but probably aren't at the moment. Not your board, but other boards. Great.
Mario Platt: Yes. Yes.
Walter Haydock: Awesome. So so looking ahead, say we're a few years down the road, say it's 2030, what's one AI security assumption that organizations are making today that will prove dangerously wrong. This could include, you know, maybe we shouldn't trust OpenAI and Anthropic to keep their models under control. but but, you know, something like that, what what do you think will be the most dangerous assumption that people are making today?
Mario Platt: Yeah. Mm. That's a question. so I think that an element of it has to do with the adversarial prompting, I think would be one of them, right? So are we asking how our suppliers are are really kind of focusing on their adversarial component where it's demanded, particularly for internet facing systems? there's a lot of AI startups, right? I get about a hundred emails a day and 20 phone calls that I don't pick up of AI of AI startups trying to sell me their products and services. And obviously there's a difference between kind of companies that have been around for three, six months that are probably expecting to be bought in the next three to six, right? And the type of of maturity that they're thinking that they need to get to, versus other companies that are building again. I think the the question I was asking earlier around the red teaming is a good sniff test, right? Because I think it tells you if the company that you're operating with, particularly if they've got a customer facing, if they're thinking about the long-term viability. And security of the product, or if they're thinking about an exit, right? Because if you're just thinking about an exit in the next 36 months, you're not going to be funding a red team. That's the general thing. Or you're probably more unlikely to make that type of investment. So I think kind of understanding how some of those companies are making their security investments to ensure that the products and services that they're providing keep on being secure. Either having an internal red team or investing in red team. teaming services that they can provide as assurances to the functions, right? Not necessarily having all of that that capability. But I think it shows a level of care, right, for for what you're producing and for assurances that you can provide your customers that not everyone out there currently is is currently thinking about
Walter Haydock: So maybe looking at misplaced economic incentives would be
Mario Platt: Yes.
Walter Haydock: you know, something something that that we might look back a couple of years from now and and see was was a little bit off. Definitely. Definitely, yeah.
Mario Platt: Yes? And I think we're all going to get at some point where we we know now with the IPO from Anthropic, for instance, right, the amount of money that they're losing, right? At some point some of those
Walter Haydock: Yeah.
Mario Platt: companies are gonna want to turn a profit and if we've got both
Walter Haydock: He
Mario Platt: our business and security tooling on there, in the change in the economics change under our feet in a minute, then what are we gonna do about it? it's complicated. Complicated.
Walter Haydock: Yeah. Yeah. I that I mean really that kind of amazes me. You know, you see you see these numbers you know, both OpenAI and Anthropic, they're just burning money. But then every announcement I see is like, we're dropping the token price by, you know, ninety percent for I just I I I just don't I don't quite see it. Like I could see you know, at at ten X the prices, then you know, yeah, these would be wildly profitable companies. But I guess you know, I don't know. We'll we'll we'll find out. We'll find out. Yeah. Yeah. We'll find out. I don't
Mario Platt: We'll learn, we'll deal with that at some point. We'll see. You'll find out. We'll find out.
Walter Haydock: trade individual stocks because I just I know I you know, I I don't I don't pretend
Mario Platt: Yeah, same.
Walter Haydock: to know enough, even even though I work in the space, I still don't. I clearly don't know enough to to make those types of like make those types of trades. So yeah, awesome. So yeah, as we as we wrap up, is there anything you'd like to plug on this podcast? We'll definitely get the report that you mentioned in the show notes so that folks can take a look at that, can check out the LastPass insights. But anything else that you'd like to to to share with our audience to point people to?
Mario Platt: yep. So other than the the Saz Air report has some very interesting statistics on data that that we got both from from surveys with clients etc anonymised platform data as well. So I think that's gonna be very informative. and again earlier this week we at last pass we released some new cap product capabilities, AI prote protect and also web protection. that again goes with things like warning users instead of blocking if you don't want to do so, blocking other things etcetera depending on your profile. That that really might fit and might solve a lot of your visibility problems and particularly if you're already a Last Plus customer. so yeah I would suggest everyone to to look at that. We've got a much more exciting things coming in the next six nine months, more product lines, more e evolution of our products so it's a really exciting time to to to be working here.
Walter Haydock: Awesome. Well, definitely check out LastPass's products if you are in the in the need or in the market for AI identification and and enforcement. So thanks for that. All right. Well, Mario, thank you very much for coming on the show. It's been great speaking with you as always and to our audience. Thank you very much for tuning in. This has been the Deploy Security Podcast, and we will see you next time.
Mario Platt: Thank you very much for having me. Great to be here.
Walter Haydock: So can you for for those folks who aren't familiar with LastPass, can you talk a little bit more about the company and its mission?
Mario Platt: Yep. so last pass we started as as a password manager, like the the first password manager, both in B to C and B to C and B to B. and with the goal of really making work safe and simple. that's still true, but we now offer much more capabilities in terms of the the platform that we provide. so this week alone we Started a few months back getting onto the area of SaaS access protection. and this week we've released some new product features on AI protection and also SaaS protection to enable clients to to get policies on how they're accessing their AI and SaaS and SaaS applications overall. so we're not just about passwords anymore. and our goal is really to focus on the needs of largely the SMB and the mid-market, so teams that may not have the the privilege of of having big security teams like myself, to to make sure that they can manage risk in ways that that that they can afford and in ways that they can have their own IT teams or small security teams managed for them.
Walter Haydock: Awesome. Awesome. Thanks for that background. So as CISO, obviously every CISO needs to deal with AI these days. And in different organizations, leadership might be pushing speed of deployment over some other concerns related to security, privacy, controls, things like that. How are you balancing that at LastPass right now?
Mario Platt: Yep. so that also goes into my into my experience, especially with talking with other peers in the in the industry. so my I I think my first comment would be that generally a lot of companies are trying to govern what they can't see. and that's a challenge, right? kind of hope is not a strategy, that's wishful thinking. so I I think we took governance overall I think really needs to be seen as an evolutionary practice. Now I'm in the in in the in the benefit or I have the privilege that we started slow, right? so we actually started by by blocking access to all of AI applications whilst we figured out what controls we need, how do you manage the related threats. And then over time we started more opening up and enabling the business a lot more. so we actually have that that experience that we had internally from going very, very slow, which was not good for innovation for all of the benefits that that business is trying to get to AI. Where we're now kind of on the on the path of now that we understand the risks and my the my own security team has started to learn what's acceptable, not acceptable and the things that that do not pose such risk to the company. We're now in the process of accelerating all of that much more. and I think and ideally that's what That's not what a lot of teams out there have to deal with. A lot of my peers they woke up one day and everyone was using insection AI, and that's a much much more complicated problem to solve that that I'm pr again privileged that that's not what I what I had to deal with. but but but really what you what you said is absolutely correct. that we need to do to really think how about we're enabling the business. At the end of the day, we're doing this with pursuing a lot of the benefits that the technology is is warranting, and we need to be cognizant of that. of both what experimentation looks like such that we can enable our stakeholders to to test the benefits that it could provide and also having processes that are not overly prescriptive, long, etc. that can help support adoption at speed within the organization.
Walter Haydock: Yeah, I think the point about trying to govern what you can't see is really key because I always recommend starting with a with an asset inventory at the beginning to understand exactly what's in use. You know, that can take the form of, you know, looking at spreadsheets, can you can look at finance statements, see who you're paying for AI tools. It can take the form of of technical approaches and you know, happen to hear what LastPass is doing from from a technical perspective on AI discovery, if if you're open to sharing.
Mario Platt: Yep. so the from now on we're gonna be using a lot more of our features the of our product, etc. That's one of the things that we that we're we're gonna help companies with, but that's really new. so generally we have obviously things like web proxies, like our procurement systems, etc., that we first use to to get a sense of what was out there. and then part of our procurement process, we introduced routines to kind of understand if as contracts were Coming through for our biggest key risk suppliers, what was their use of AI? Because that's another part of the challenge, right? A lot of us are dealing with the fact that we wake up one morning and suddenly one of our vendors enabled their on their product without telling us it's just there now, right? And we had no opportunity to review to so that's the I think the reality that that that everyone can relate to. so it's really been about as you said, leveraging existing inventories and the data that we had on the systems. ideally cross-checking those with web web logs so we can see what our employees are doing from from their devices and and correlate with yeah with what what's actually in use and not just what we believe is on the procurement system which yep which is not does not always tell the same story and both relate to to effective governance and security.
Walter Haydock: Yeah. And the kind of you wake up and you find something is enabled line of events. I remember there was a a Slack feature push where they said AI is enabled for your full workspace. And there was like no notification. They wasn't in the admin release notes or anything like that. I know a bunch of bunch of CISOs that I always that I that I talk with, they say, Well, hold on, what's going here? yeah, I I didn't know I didn't know this was coming. So definitely definitely a real problem. Yeah, yeah.
Mario Platt: Quite a few vendors. Quite a few vendors and not all of them give you the opportunity to opt out, right? it depending on their strategy, so it becomes
Walter Haydock: Yes, yes, yes.
Mario Platt: complicated to help manage to help the business manage risk.
Walter Haydock: Yes. Yes. So on that note, we we mentioned kind of the the surprise AI feature side of things. What are what's the biggest blind spot that you see organizations run into when they're adopting generative AI? Is it, you know, shadow AI? We talked about that. Is it leakage? Is it model drift? What what what's the biggest risk you see or blind spot?
Mario Platt: Yep. F from so again I'm gonna call out my bias. I believe that shadow AI I think is one of the biggest challenges, for a number of different reasons. and also this depends on the maturity of controls that people had, right? one of the things that I don't think is talked about enough in the industry is how in my opinion, people that were not mature in managing SAS application risk, it became so much more compounded with AI, right? because the the the channels of access are are essentially the same, right? And if we we know for instance from so one of the things that we've released a couple of weeks ago, LastPass, that I would invite everyone to to have a look is the state of AI and SAS security report 2026. Right. And so there are some some specific statistics there that I think are are scary, right? Or at least they scare me, right? And that that talks about this compounding of of risk between not managing the SAS. Component and how that exacerbates the AI exposure risks to a lot of us. So knowing that a lot of companies don't have the maturity to know if people are using a corporate or versus a personal ChatGPT account, for instance, or Anthropic or whatever you may decide to use. So we know from anonymized data from our platform, for instance, that almost 53% of users they reuse passwords across accounts. We know that about 21% of Users, they've got some of their accounts have been in non-breaches that they haven't necessarily changed. So if you compound that with the fact that a lot of users are using their personal accounts on ChatGPTs and Athropics to deal with with customer data in their businesses, if you combine these three things together, that's a pretty scary picture, right? and I think that's that's one element where yeah the the SAS and the AI risk has has compounded each other, right? companies that that didn't do that. And and again goes back to the how I started, right? if we we we can't govern what we can't see. So really having that that visibility of what people are doing, are they using the the appropriate accounts? do we understand what people are putting into these tools? which isn't necessarily about about blocking, right? And that's a challenge that we security industry have had for many, many years. we always think that the only way to to reduce risk is to block things. But I think what the industry and what executive teams are telling us, what industry at large is telling us is there's actually a lot of risk tolerance for for AI-related risks, right? People want to pursue the benefits, right? That's just what what every executive team right now has mandates on improving efficiencies, etc., through the use of AI. So the risk tolerance exists. Our job as security people then I think becomes less about how am I going to block all of these things, but but more but a question that That I don't see security teams asking enough is how do I nudge users to the types of behaviors that I would like to see, right? That are less risky behaviors, etc. Right. And I think the the this idea of of nudging users, which is part of how we've developed the AI protect product at last plus that we can decide to block if we want to, or just warn a user. So you're about to put some confidential information here, are you sure you want to do that? Right? Some because that may not, and even in prompting with Chat GPTs and the clouds of this world. having one thing that we've done internally at last pass was build our own custom plugins into an into cloud right that check our security policies and try to nudge users. Look if you do it this way then you're meeting all of the security policies, right? This type of nudging behavior is something that I I don't see most security teams see the value in what they are doing because in the the arguments that I usually get are valid, which is look, an adversarial a sufficiently adversarial person will be able to go around them. True, but most of your employees they're not adversaries, right? they're not necessarily trying to deliberately circumvent your policies. So if all we're doing is only solving for the problem of adversarial, then you actually got a business culture problem, not a security one. It affects security, right? But that's a bigger problem if everyone follows your employees are trying to do that deliberately. And that's not the case for most companies, right? Most humans they want to do the right thing, they also want to do the quickest thing. So if you figure out ways to introduce those nudges. at the point where people are interacting with these tools and are trying to achieve the benefits, I think we are helping the company reduce the risk, right? Because no no employee wants to be the one that was in the in the path of a data breach. And I think that's that you yep we need security teams need to be more willing to to adopt those types of strategies in the name of risk reduction and not risk elimination.
Walter Haydock: I think the point about nudging employees who are trying to do the right thing and especially nuanced circumstances is really important. You mentioned, you know, logging into a personal Chat GP account account versus a corporate one. And employees might not even know that they're logged in on their personal account while they start doing something, but there are real differences in terms of how that data is handled. You know, if you've got a Chat GPT enterprise account, you can, you know, you there's no training on that data, you can set retention periods, you can purge it if it's
Mario Platt: Yes.
Walter Haydock: if necessary. If it's a personal account, not so much. So controlling that login behavior is it's it's very nuanced and it's great to hear that that you're able to do that because that can have serious impacts from from a data confidentiality perspective.
Mario Platt: Mm-hmm. One hundred percent.
Walter Haydock: So on a somewhat higher level, can you talk about what an effective AI governance framework looks like in practice rather than on paper? So you mentioned policies that can be long and and wordy. you know, that's something I see commonly are are are are policies that have great aspirations, but maybe the organization has trouble meeting those requirements. How do you how do you make sure that you can operationalize AI governance?
Mario Platt: Yeah. That's a great question. It's actually something that we I'm currently in the process of of evolving. So one of the elements that in and again going being consistent with my earlier comment, right, that as not being able to govern what what we cannot see. one of the elements that that that that is the direction of travel that we're working on at the moment is really having a focus on what I'm calling a policy enforcement architecture, right? So we've got a bunch of places where we can we can introduce policies. at the network level, at the browser level. start with LastPass at the plug-in level, at the levels of before we add plugins to the market, are we considering prompt injection and and those other things that that the engineering teams or the the the business teams are doing right so considering all of these points where we've got some level of control where we could get policy even at runtime with behaviors that we can do with through cloud or or or chat gpt is really kind of first having a view of what the what what we care about right what our way that are ideally threat driven that drives so this requirement now applies because you answered yes to this type of thing that can pose a threat to the business and then having those in policy enforcement points being very clear of we apply this control over here in this interaction in the other interaction and make sure that we don't build policies where we cannot say this is where it's going to be enforced right so that's that's one of the the the elements of again going back to my previous point. I had the privilege of starting slow and because we now see okay, now we know where the all the controls are, we can now start being start moving the model to the policy enforcement points being something that we do from a technical point of view and and that technical par part of the enforcement and nudging the users at the points of interaction as opposed to again something that may live on a policy document that people acknowledge once a year that may or may not affect their behaviour on Yeah.
Walter Haydock: Yeah, if an administrative control is only administrative, then I would apply a a very heavy discount to its effective its effectiveness. So, you know, train training certainly helpful and I I would absolutely encourage folks to do that. But if you know it's all it is, like you said, is just one time a year acknowledging that you're not going to do XYZ, then it's pretty hard to really ensure any sort of enforcement. So if you can turn those policy statements into actionable, enforceable controls, that's where you really achieve the the mitigation. So, you know, like like you said,
Mario Platt: Yeah.
Walter Haydock: if you're if you're at the point of entering certain types of data into a certain interface, then, you know, that that's where that's where the rubber meets the road and that's where you can really make a difference. So so absolutely
Mario Platt: Yeah.
Walter Haydock: agree on that.
Mario Platt: And extend and even I would go even a bit further, which is extending that that component which from a risk point of view to understand what are things that really need to be enforced, you know, in people should not be able to bypass, in things that we should just say, look, be careful, minimize this data, type of warning. But look if you've got a business reason and etcetera and you say, Okay, I've I I tried, there's no other way to do this, then if that's within the risk tolerance of your business, it's not regulated data, those types of things that obviously companies need to to be much more careful about, then it's about making sure that the there is tolerance of the business gets translated into those policy enforcement points, right? And that we're still nudging the users in the be in the directions of the behaviors we want to see, but not necessarily kind of stopping something that would then generate a ticket that someone needs to go on the queue and see. So yeah, and have lengthy processes where humans need to to discuss and argue and takes a long time. So understanding that is key for security to be perceived by the business. as as a helping reduce the risk but also helping enable in the experimentation etc which is what all companies are being asked to asked to do.
Walter Haydock: I think if you have clear policies, then you avoid both of those failure modes that that you talked about. So one, you, you know, the folks won't be doing things that they shouldn't be doing, but also they won't be uncertain if they're doing something that is authorized by the business. There won't be any any uncertainty. So if they can go to a register and say, hey, I can use Claude Code to, you know, generate prototypes. I can, you know, do this. If I apply a human review to the outputs, then you know, go ahead, experiment. create create new new products, features, things like that that can accelerate the business while staying within the risk appetite of the organization.
Mario Platt: Yep. One hundred percent.
Walter Haydock: So on an external perspective, how are you at LastPass, but also how are you helping your customers adapt to regulatory developments like the EU AI Act, which obviously has gone through some changes recently, and also in the United States, the proliferation of the state level AI laws that in you know in some cases are more restrictive in in my view than the EU AI Act. But you know, it's a it's a very varied playing field here on this side of the Atlantic. So Can you talk about how you're addressing that?
Mario Platt: Yep. So I think there's first I think the the key thing to have is really good partnerships between the security and privacy functions, right? I I think that is key in terms of figuring out what's the best way to to deal with those. Having said that I I think the the same advice that would apply to any other security thing kind of still applies. If we if we build our governance frameworks around I think there are for me there are two kind of good standards to to kind of consider which yep is yeah obviously Stack Aware you helped us with building our our initial one and we're about to do some more work to to continue that evolution. But I think basing those on ISO forty two thousand and one in the NIST AIRMF. Right in RMF, that's the best thing that we can do. Because then all of the level of of specific laws, etc., particularly for companies that are not part of what many of those regulations call high risk systems, right? So if you're not dealing with credit scoring, with employment, with with all of those types of use cases that obviously have different types of of considerations, then it's a good baseline to then expand as as those regulations pop up like mushrooms at times to kind of r adapt some of your practices to make sure that you're adhering to to those but if you don't have that that strong baseline based on a standard that you can evidence that that you're basing off of good practice then it becomes really complicated and you run the risk of always be chasing the the the the regulation right and as you said particularly in the US it's a mishmash of of approaches even right so I think Colorado is a bit more like GDPR but then you've got Texas that it's much more about the the harm and incitation and deep fakes and so cuss consumer protection type stuff right so they're all taking very different lenses into how how they even look at what needs to be regulated in the context of AI. so it's definitely a complicated landscape and it's not gonna get easier, right?
Walter Haydock: I think the reference to ISO 40 2001 is really important there because if you're doing it correctly, it gives you a way to consider all these regulatory legislative requirements like the EU AI Act, the state level laws like Colorado, Texas, what have you. And then drive setting of objectives, creating an AI policy, doing risk assessments. All this stuff is driven by the external context of the organization. And It's it's kind of easy to try to check the box on that and say, like, you know, regulatory concerns are an external issue, which, you know, yes, I agree that those are those are external issues. But if you get much more specific, then you can say, okay, we're using this AI system in this use case. Let's evaluate that against EU AI Act, Colorado SB 189, you know, California CCPA ADMT regulations. Okay, well, it triggers. These requirements, but not those. We need to apply these controls, but not those. It gives you a very nuanced way to attack those types of requirements.
Mario Platt: Yeah. And again if you don't have that that strong standards based approach then okay now we need to do this for this one and that for that one. You don't have the basis where you can bolt on and extend as as need needs arise.
Walter Haydock: Yep. So obviously you're helping customers deal with AI security risk management. You're dealing with it internally. But then your customers are probably looking at you saying, okay, well, now you're using AI. Obviously, you know, you've you've described a comprehensive approach to it, how you started slow. But how are you responding to your customer scrutiny of LastPass regarding use of AI? What what is the state of sophistication you're seeing in the questions you get? you know, does it vary or is it pretty consistent? What are what are you seeing?
Mario Platt: So we see a bit of all of the above type of thing. so we see a lot of customers that aren't particularly asking questions, we see a lot of people for their own needs and they're they're asking many more questions. So we're seeing particularly coming through procurement process in in bigger deals, etcetera, references or asks relating to how we map our controls against again RMF, and if we are 42,000 certified, which is not something that we have yet. We're about to start a gap analysis that Stack Aware yourself are going to help us with, which I expect us to to do early next year. so we're currently in the process of making sure we already know and have the assurance that we built the program to begin with to use that as a baseline, right? Already because we were thinking about about eventually getting to certification, like we have 2701, 27701 right now. and yep, and now it's going to be that next level to to make Make sure that next year we can start providing some auditable credentials to to help facilitate the procurement process.
Walter Haydock: Awesome. Yeah, I've I've seen a wide variety of sophistication when it comes to customer inquiries about AI. I've seen, you know, hyper specific questions about, you know, how do you categorize yourself you know based on a certain law. you know, I've seen no questions asked. And then I've kind of seen a little bit of a a a muddled middle where folks kind of know that they should be concerned about. AI
Mario Platt: Yeah.
Walter Haydock: governance and security, but they're not exactly sure what they should be concerned about. So then you kind of get some some oddball questions in there that it's hard to answer. you know, I I I I've seen questions say, you know, confirm that there are no there's no bias in any of your AI systems, which which I would
Mario Platt: Ha ha ha.
Walter Haydock: advise, you know, okay, well, we'll hold on. We I mean the only way to confirm that is to just delete all the AI systems, then there's no bias because you know, that's how AI works is through a bias variance trade-off. there's there's no way around that. So, and then I've seen other kind of very, you know, very large enterprises try to lay down requirements on organizations. You know, one AI powered healthcare company I was working with, of a customer said, You're not allowed to use AI for healthcare. To which I would say, I I advise my customer, I'd say, You should just ask them, you know, what what should you be doing with with AI if you're if you're not doing that? Because you're an AI powered healthcare company. So Yeah, I see a a wide range of of sophistication in terms of customer concerns about about AI.
Mario Platt: Yeah, I've had the that other one as well. But so one thing for instance that I've been that I've been that also is driven by by risk and particularly reputational risk, in in my opinion. So there's a difference between a company providing using AI on their product, but but but it's something that they use on their back end, etcetera, that is kind of generally self contained and it's about optimization of the product features, right? That's one type of use case. one specific question that I as a CISO am asking some of potential suppliers That would particularly provide AI features that are customer-facing, that's a different thing, right? So one of the questions that I'm starting to ask, and I might maybe one of those let me gently call it obnoxious people that asks kind of random questions that seem that seem random, but is do you have an internal red team? Are you red teaming those those components, right? Because I think it's a very legitimate question because if it's something that's going to be that my customers will potentially be interacting with, I want to know how you are testing. that people can't say please ignore instructions and now do this instead etc because at the end of the day that's business reputation right so on any new supplier that i'm that yeah that potentially has the the the opportunity or the architecture to be customer facing, I'm being much more prescriptive in asking those more detailed questions because look, I wanna see those reports, I wanna see exactly how you're testing this, because yep. we don't want things that happen to other companies like Air Canada and that a others in the past, right? Where those questions were likely not asked and then you get yeah, you wake up one morning and you're having a type of impact you weren't expecting, right? So at least to do what we can from the diligence point of view to reduce to reduce that likelihood.
Walter Haydock: AI red teaming is a really great control because it shows, you know, what what a motivated and intelligent person can do when they're trying to force your AI system to misbehave. And you know, it's one thing to have design documents and you know, apply your your traditional security controls, but it's another thing to see in the wild what is your system doing when somebody who's who's motivated and an expert trying to trying to break it. So definitely agree.
Mario Platt: And if it's a in if it's a customer facing f system, people have all the time in the world to try and break it. Right. They're
Walter Haydock: Yep. Yep.
Mario Platt: not constrained by time. They can come back tomorrow and be another three hours at it. Right? that's yeah.
Walter Haydock: Yep. Exactly. Exactly. So in terms of the steps that you've taken, obviously we talked about identification, asset inventory, we talked about actionable controls. In terms of return on investment, what's the top thing that you've seen that you've implemented or you've seen other organizations implement? And this doesn't necessarily need to be the most impactful, but it's from an ROI perspective. What's the easiest switch you could flip? From a security and governance perspective.
Mario Platt: I'm not sure it's a a switch, but I I think the the the greatest ROI that I had, so I would separate that between governance and security, right? kind of both are important components. They relate to each other, but not necessarily the same thing. So from a from a governance point of view, personally I think was leveraging the the API that Stackware provides, right? in terms of AI risk assessments, etc. So the my teams were were spending circa eight hours, we didn't not always had Some of our suppliers have answers to the questions we were posing. A lot of things were getting stuck in the process. Stakeholders upset because we are not finishing the process, because we're not getting the answers that we need. So kind of adopting the Stackware API where we can self-serve a lot of the risk assessments, moved some of our of our internal governance processes from being eight hours to being closer to two or or less than that. Right. So that was an amazing ROI from internal stakeholder kind of perspective and yep. And seeing that we're trying to do our best to accelerate the process, right, while still getting really high quality results on data that's been vetted. So from a governance point of view, I would point to that. from a security perspective, as I mentioned previously, was us developing our own custom plugins for Anthropic, right? again, nudging users' behaviors in right directions. now we've got that integrated with with our MDR service and all of that type of thing. but but but that was yeah, creating our own plugins that direct users towards the right behaviors that does we use pre-use to we use hooks for some of the things that we want to determine to to do on a deterministic fashion within that plugin as well as opposed to leaving it to LLM based judgment so that's been the the the greatest ROI because I can I know that I'm nudging my users at the point they're interacting with systems towards the behaviors that I want to see.
Walter Haydock: Yeah, I mean you can really customize these these tools like ChatGPT and Claude to basically an infinite degree that really meet your business requirements. And I think it's it's great to hear that you're you're leveraging those capabilities, building building in-house tools to accelerate your your security In terms of the high level, so boards and executive teams, we talked a little bit about the executive side, but but at the board level, what should boards be asking about AI risk that they are not asking today? Because, you know, I'm sure folks are getting pressure from their boards to adopt AI. But what questions should be getting asked that aren't about the risk side of things at
Mario Platt: Yeah. so I'm again lucky that I I get asked very specific questions about that subject, so and and I I have to report on them. obviously the There's an element that I think is going back to to how we even think about the policies themselves, right? So how are we enforcing policies, right? So across not only u utilization of kind of chat based tools for lack of for lack of a better word, but it's also the element of how we're integrating business systems with our capabilities, right? So between those two types of things there's a lot that can go wrong and there's a lot of places where policies should be enforced and maybe aren't. So I I think Boards asking more about specifically okay, you've got these AI policies, I can see that they're written, how are you enforcing them, right? How does that affect when someone gets to their laptop and accesses a website, when they try to paste data into into a chat tool, when when you bring a new supplier to the business, how are you identifying do you have the logging capabilities to know how people are interacting with it? Do you know what type of data is being is being transferred, how people are making use of it, how are they being retained? So So all of those those are two components where I think it requires maybe a level of s sophistication, right? But but but it's questions that that that boards in the executive team should be asking because at the end of the day, again it's how we transfer it from being kind of wishful thinking, we wrote some policies and and we're hoping that people are doing the right things, to actually do we have the visibility in the policy enforcement points as people are doing these things, to to know if policies are being adhered to or not. Right.
Walter Haydock: Got it. So focusing on the operationalization of the AI security program through metrics, through technical controls, that's something that boards should be asking, but probably aren't at the moment. Not your board, but other boards. Great.
Mario Platt: Yes. Yes.
Walter Haydock: Awesome. So so looking ahead, say we're a few years down the road, say it's 2030, what's one AI security assumption that organizations are making today that will prove dangerously wrong. This could include, you know, maybe we shouldn't trust OpenAI and Anthropic to keep their models under control. but but, you know, something like that, what what do you think will be the most dangerous assumption that people are making today?
Mario Platt: Yeah. Mm. That's a question. so I think that an element of it has to do with the adversarial prompting, I think would be one of them, right? So are we asking how our suppliers are are really kind of focusing on their adversarial component where it's demanded, particularly for internet facing systems? there's a lot of AI startups, right? I get about a hundred emails a day and 20 phone calls that I don't pick up of AI of AI startups trying to sell me their products and services. And obviously there's a difference between kind of companies that have been around for three, six months that are probably expecting to be bought in the next three to six, right? And the type of of maturity that they're thinking that they need to get to, versus other companies that are building again. I think the the question I was asking earlier around the red teaming is a good sniff test, right? Because I think it tells you if the company that you're operating with, particularly if they've got a customer facing, if they're thinking about the long-term viability. And security of the product, or if they're thinking about an exit, right? Because if you're just thinking about an exit in the next 36 months, you're not going to be funding a red team. That's the general thing. Or you're probably more unlikely to make that type of investment. So I think kind of understanding how some of those companies are making their security investments to ensure that the products and services that they're providing keep on being secure. Either having an internal red team or investing in red team. teaming services that they can provide as assurances to the functions, right? Not necessarily having all of that that capability. But I think it shows a level of care, right, for for what you're producing and for assurances that you can provide your customers that not everyone out there currently is is currently thinking about
Walter Haydock: So maybe looking at misplaced economic incentives would be
Mario Platt: Yes.
Walter Haydock: you know, something something that that we might look back a couple of years from now and and see was was a little bit off. Definitely. Definitely, yeah.
Mario Platt: Yes? And I think we're all going to get at some point where we we know now with the IPO from Anthropic, for instance, right, the amount of money that they're losing, right? At some point some of those
Walter Haydock: Yeah.
Mario Platt: companies are gonna want to turn a profit and if we've got both
Walter Haydock: He
Mario Platt: our business and security tooling on there, in the change in the economics change under our feet in a minute, then what are we gonna do about it? it's complicated. Complicated.
Walter Haydock: Yeah. Yeah. I that I mean really that kind of amazes me. You know, you see you see these numbers you know, both OpenAI and Anthropic, they're just burning money. But then every announcement I see is like, we're dropping the token price by, you know, ninety percent for I just I I I just don't I don't quite see it. Like I could see you know, at at ten X the prices, then you know, yeah, these would be wildly profitable companies. But I guess you know, I don't know. We'll we'll we'll find out. We'll find out. Yeah. Yeah. We'll find out. I don't
Mario Platt: We'll learn, we'll deal with that at some point. We'll see. You'll find out. We'll find out.
Walter Haydock: trade individual stocks because I just I know I you know, I I don't I don't pretend
Mario Platt: Yeah, same.
Walter Haydock: to know enough, even even though I work in the space, I still don't. I clearly don't know enough to to make those types of like make those types of trades. So yeah, awesome. So yeah, as we as we wrap up, is there anything you'd like to plug on this podcast? We'll definitely get the report that you mentioned in the show notes so that folks can take a look at that, can check out the LastPass insights. But anything else that you'd like to to to share with our audience to point people to?
Mario Platt: yep. So other than the the Saz Air report has some very interesting statistics on data that that we got both from from surveys with clients etc anonymised platform data as well. So I think that's gonna be very informative. and again earlier this week we at last pass we released some new cap product capabilities, AI prote protect and also web protection. that again goes with things like warning users instead of blocking if you don't want to do so, blocking other things etcetera depending on your profile. That that really might fit and might solve a lot of your visibility problems and particularly if you're already a Last Plus customer. so yeah I would suggest everyone to to look at that. We've got a much more exciting things coming in the next six nine months, more product lines, more e evolution of our products so it's a really exciting time to to to be working here.
Walter Haydock: Awesome. Well, definitely check out LastPass's products if you are in the in the need or in the market for AI identification and and enforcement. So thanks for that. All right. Well, Mario, thank you very much for coming on the show. It's been great speaking with you as always and to our audience. Thank you very much for tuning in. This has been the Deploy Security Podcast, and we will see you next time.
