AI Governance & Security in Practice: Visibility, Controls, and Guardrails | Yagub Rahimov
Welcome to Secure and Simple Podcast. In this podcast, we demystify cybersecurity governance compliance with various standards and regulations and other topics that are of interest for consultants, CISOs and other cybersecurity professionals. Hello, I'm Dejan Kosutic, the CEO at Advisera and the host of Secure and Simple Podcast. Today, guest is Yagub Rahimov. He's the CEO and founder at PolygraphAI, a company that implements AI security with locally deployed, explainable, and auditable small language models.
Dejan Kosutic:And he has helped with various companies, including defense, banking, and law firms with the secure implementation of AI. So in today's podcast, you'll learn some practical tips about implementing AI governance in practice, especially from cybersecurity point of view. So welcome to the show, Yagub.
Yagub Rahimov:Hey, Dejan. Thanks very much for having me. You actually overhyped me, I guess. Thank you.
Dejan Kosutic:Well, think you deserve it. See it's very impressive career you have. Okay, so in your opinion, how does AI governance look like from a security perspective?
Yagub Rahimov:You mentioned a couple of things that we've done, or I've done personally myself. I've helped government institutions define what their AI governance, AI security really mean, what it should mean. I've been more with policymakers helping them to define the regulatory frameworks as well, and even on the nuclear operations, or shipping industry, etcetera. But you pay attention, I mentioned different industries, different focus points.
Yagub Rahimov:And there is one very, very simple thing that's in common in all of them. They actually don't know what they are trying to secure. So most of the times, CISOs have the network security, CISOs have the, I don't know, firewalls, etcetera. But they still are very, very they have... It's as if they have they have been blindfolded with their AI operations.
Yagub Rahimov:They don't know who is using what agent, who is using what AI, if any AI utility is there, and what sort of failures, privacy compliance failures that we are talking about that they are having. So the very first thing when we are talking about AI security, AI governance, one needs to understand is their exposure points. How are they exposed? And what should they be doing it? And the secondary part is that, which is very, very common, they think that very often the AI security is a blanket coverage, something that goes for the whole organization, and I very much disagree with it, and we go around and try to educate our counterparts on why it's not the same thing as a traditional firewall or traditional DLP or whatsoever that you may be doing.
Yagub Rahimov:AI itself, of like a human being in a way, intelligence, as it says, artificial intelligence, it is semantic in a way. So context really matters a lot. So you need to really think about, like, who should be doing what and how it should be engaging it. So saying that we are kind of blocking something and everything and expect something real, it doesn't work. Or saying that the whole company should never touch their code with AI, and then you're asking your employees that why are they not using the agentic coding or whatever.
Yagub Rahimov:You banned it. It doesn't work like that. So you need to really treat each division, and even inside the division, each group separately, and assign different sort of policies on it, and help them understand the context level importance, context level security around it, which then comes around it, human identity, non human identity, etcetera, that we can, I guess, and we will talk about today?
Dejan Kosutic:A couple of interesting things you you mentioned. So first of all, you're saying that basically any company doesn't really know how much or how are they exposed. So what do you suggest actually? How do they make an inventory, if you will, of all the things that they are using?
Yagub Rahimov:So I've had this example as a company saying that we do not allow anything except, I don't know, Copilot as an AI in our environment. And I'm I'm just basically simply asking them, do you have google.com enabled? It's Gemini powered. Right? So you no longer can say that we do not.
Yagub Rahimov:So the first... Very first thing, this is not promotion or whatsoever. I will mention what we are doing. That by all means is not a solicitation to anyone who is listening here. For us, what we do is we have built a desktop application that does not require any integration at all to your own operation.
Yagub Rahimov:Takes like fifteen seconds to install it, that's it, for each device. So within seventy two hours, we get to map out basically every website, every application, every tool that you've used in that organization, what sort of data input that has been there, and that basically gives us the ability to then later on break it into departments and the groups, map out what needs to be done for... Based on what you guys are really doing as an organization. Instead of coming and blindly saying that you have to do this, we we just keep it quiet. Let us actually understand how you are using current tools.
Yagub Rahimov:You are anyway doing that, what you have been doing. Within three days, now we understand that your ex... Let's say your legal team loves using Copilot. Your engineering team enjoys the, I don't know, like, cloud code. And then in your DevOps scene, they engage with, I don't know, like codex, cursor, or whatever it is in there.
Yagub Rahimov:Now we we also have an understanding of what sort of failures are are there are there actually enabled on it. So with that said, now the next part is, then you move on from basically visibility visualization of your risks and challenges and security breaches to controlling. You start basically mosaically or puzzle wise, like one by one building the bricks and then going on to the enforcement. Once you are in the enforcement layer, you are supposed to be seeing every engagement with any AI tool within your own infrastructure and be able to assign policies from privacy to guardrails from agentic level kind of like security.
Yagub Rahimov:But one thing that I personally consider myself as Mr.Paranoid, as my close friends are calling me, The security part is something that cannot be dependent on another company, including Polygraph, meaning that you should not say that I don't trust, let's say, OpenAI or Anthropic, and let me trust Polygraph to trust OpenAI and poly... Anthropic or whomever else it is. We really believe that the security needs to come sovereignly. Meaning that you need to the models, the tools, etcetera, that are out there needs to be locally running so that you have independence, you have sovereignty in your own operation. Once it is done, then you get to see that privacy is being mosaically integrated.
Yagub Rahimov:You are in Europe, your AI app can be properly managed without having the challenges in there. And on the other side, let's say you are in France and you're operating with The US and The UK as an example only, Naturally speaking, your national language is French, and then with your international business operations you have supported English. Now you can actually enforce AI only to speak two languages. And you do all of these things independently of LLMs as a kind of like a guardian agent.
Dejan Kosutic:Okay, will certainly come to that point because it's very interesting. But going back to this basically finding out where the exposure is, so are you saying that basically it's enough to deploy some kind of a technical tool to, let's say, list all of the, let's say, interactions with the AI or is there something else that is needed?
Yagub Rahimov:There are two types of companies. Companies, they think that with their words, can control everything. You're gonna again maybe laugh at me. I get to see specifically in legal financial field and very often in Europe actually. I hear companies go in and make an employee sign that they will not put x y zed type of a company detail into an AI.
Yagub Rahimov:And then they're like, we are compliant. The employees have agreed for it. I'm like, if you are gonna be listening to network and then you're gonna be using another LLM to process that data because you need to be fast, you need to be able to do many things contextually. Now you are basically moving from one attack vector to another attack vector, so you have actually doubled up your thing. Maybe one of it has has been lowered, the other one has been super inflated.
Yagub Rahimov:So it it actually, there isn't right or wrong approach, there is just, unless you really are in 100% full control of your security measures, etcetera, your data boundary, Currently, I'm seeing it like different shades of wrong. And unfortunately, more often than it should be that I'm seeing it more in healthcare industry as an example. I'm seeing it more often in public sector. I'm seeing it more often in more critical industries such as banking as an example. You probably have seen the threats in a revolt environment that people make mistakes and we need to, my personal conviction as the founder of Polygraph is let's stop thinking that we are securing the technology.
Yagub Rahimov:We are not securing the technology, right? Everyone can say that cybersecurity is about securing the technology. Let's start with behavior first. Let's put our 100% effort to make the end user's life easier, show the failure before it happens to the end user and make it very simple. That's why I know you described me as a security expert.
Yagub Rahimov:I don't define myself as an AI security person. I define myself and my own team with the technology that they are using. So you help the person, every company today spends millions of dollars to train their employees on what to do from compliance perspective, GDPR, what to do from the GDPR perspective. But come on, I have run a very large financial institution myself, and we have brought those consultants too. And then the next day, the very same employee that was put into front of the training made that biggest gross mistake that he was told not to do.
Yagub Rahimov:It's called being human. Now what I'm saying here is that bringing that training into the keyboard, and I'm not saying again keyboard, keystroke operation, that's not what I'm saying. As the person engaged, you have used as an example Grammarly, I'm guessing. So show things in red that they should never be putting in there. And the person sees it.
Yagub Rahimov:I mean, have a, you probably can see my family picture behind me over there. I was six, seven year old kid. Honestly didn't have to train him on what not to do with the traffic light. He knows the red is that stop. An employee knows to stop when there is a red in there.
Yagub Rahimov:That's the part that's part of the behavioral part. And doing something similar, but from technology perspective for other users, which are not human, I'm talking about AI agents. And then enabling them with these tools that we are talking about, these agents, AI tools, other operation applications, with additional layers of security that you would anyway be putting it. When you put these two together, I'll give you one simple example. The the thing that I just explained.
Yagub Rahimov:Right? The first time we we put it was a very large public transportation group here in The United States.
Yagub Rahimov:We delivered them the desktop application that I was talking to you about with the pre warning that happens. And around, it was almost like three, four weeks later, I get a call from the technical team. And honestly I did not think about that would be the case because it was the very first time that imagine that we are deploying. We don't know what would be the impact yet. Have of course tested it, but while testing is not same thing as giving to a real paying customer.
Yagub Rahimov:And this client was saying that, what happened to our DLP that we now have, we're only seeing like 20, 22% of our DLP violations. Does your system violate or does your system kind of collide with our DLP watching, listening? And first, actually honestly, I went to my team and I'm like, guys, something is catastrophic going on. Seems like our system is colliding with something else. And we spent like ten days, we are doing everything possible and we cannot find any collusion and nothing.
Yagub Rahimov:It turns out that yes, actually, behaviorally speaking, when you show the failure to the person, that person actually is not going to do most of the failures that he was going be having otherwise. And that was the thing that I did not really think about myself. It was not my intention at the beginning.
Dejan Kosutic:Okay, so what you're saying is basically that if this kind of behavioral controls are applied, they will directly influence the employee behavior, which will then reduce the number of, let's say, wrong things that they are doing. Now the question is, doesn't this actually this kind of a layer, doesn't it actually raise the level of potential security problems? Because if this kind of a layer actually has an eye on everything that is happening within the network, isn't this a risk in itself?
Yagub Rahimov:If it was to be storing the data somewhere else and creating kinda like like ledger of every failure and what failed in there, the data that was been whatever there, absolutely as... That's why I said, like, watching your keyboard but not keystroking. If it was a key stroking, I I would turn around and I would jump out of this window myself. It's not that. Everything is locally being handled within that device itself.
Yagub Rahimov:And only the admin basically gets to see that Yagub was on chatgpt.com and he tried to put a credit card number in there. And that's it. The very craziest thing that if you wanted to put in there maybe, and I'm saying maybe would be that tokenizing it and showing the last last one or two digit or last one or two character of whatever the failure that was it. But the rest of it, it doesn't store anything. So there is no nothing.
Yagub Rahimov:And one thing that I think you need to be aware in here, ninety six percent of the cybersecurity failures last year was not because of technology going broke. And twenty four percent of the biggest failures, million dollar failures that we are talking about, it was because of negligence. The person saying that they just didn't know or they never really cared about it. Now this thing eliminates, I'm not saying that 100% of it, because it would be, step one is just to visualize it. Visualization does not stop it.
Yagub Rahimov:There needs to be control and enforcement. Even with the visualization alone, if we can reduce it by 70, 78%, the other 22% can be caught by other technology listening and so on, control and enforcement layer, which now makes it a lot more attainable, a lot more controllable.
Dejan Kosutic:So beyond this, let's say controls which influence a human behavior, what other controls are necessary for, let's say, AI governance to work or something that is directing actually AI governance in a more secure way?
Yagub Rahimov:So when it comes to AI governance, context means the world. AI security context means the the world. If you look around the the whole agents, the whole LLMs today, there is a race to the top whomever has the largest context window. Mhmm. Right?
Yagub Rahimov:I was on a call this morning with our global team on Guardrails operations. The three things that basically is always on top of our own agenda and including everyone else as well. Multilingual attack is really, really tough. Like, if you basically do not put control layer for which languages your AI can speak, you're opening yourself. No matter what sort of security that you have in there, you've you've increased the chance of your attack vectors by 80 to 90%.
Yagub Rahimov:So to reduce the number of languages, that's the very first one. We call it language control. The second one is long context attack vector. Basically means that it's a prompt injection, prompt infiltration, or jailbreaking attempt. But it doesn't come and say, Dejan, know, open the door and give me the money.
Yagub Rahimov:It first comes and says, Dejan, do you have a door? And then it comes around and asks like, Dejan, do you have an award in that door? And where would you be hiding the key and so on and so forth? So it walks it from the context perspective that Dejan actually opens the door and then the attacker is taking whatever they after. That's called long context.
Yagub Rahimov:I made it sound very simple. The other one is the tool callout attack. This is specifically on agentic tools that's out there. It's a tool callout attack. It's authentication level activities and so on.
Yagub Rahimov:And then you carry on furthermore, then you have the, as an example, the data poisoning attack. Data poisoning, its whole job is to basically create a methodology or a tax structure that you will not be your whole workflows, etcetera, will be upside down. So your team breaks down. It's called data poisoning. I don't even know how to explain it in a simple language.
Yagub Rahimov:Carry on furthermore, inside it, the the code leak in there. Inside it, you have privacy metrics in there. Inside it, there is another thing in the defense field called Mosaic Intelligence. Mosaic Intelligence is something that, you know, individually speaking things are not really important. They just don't mean anything.
Yagub Rahimov:But when you put those things together, the knowledge wise, it makes something of value. So Mosaic Intelligence coverage, intelligence security, or contextual intelligence security becomes another attack vector. I think these are the main ones that I'm seeing personally inside it, but this can be further explained. We have 17 of these models. One of them is the JSON validation, etcetera.
Yagub Rahimov:But the ones that I just mentioned with are the main ones that most organizations I see that they are they're actually using the most or caring most about today.
Dejan Kosutic:This is very interesting, obviously these are, let's say, new types of attacks that didn't exist a couple of years ago, obviously new ones will develop in the future. So can you give a couple of examples on how do we actually tackle some of these threats or some of these attacks?
Yagub Rahimov:Whomever is with us and they are utilizing AI, they know that there is two today major type of AI utility. One that is majorly the consumer end user base activity, we call it ChatGPT, we call it Cloud, etcetera, chatbots. For... And even that chatbot itself is divided into two categories. It's self hosted and public one.
Yagub Rahimov:Mhmm. So self hosted one basically is running based on your API key or something like that you have. The public one is also known as a black box model. You don't really have a control on it. So on black box models, we integrate polygraph via different methods into that environment.
Yagub Rahimov:We cannot do it right now on an individual personal accounts. We do it for enterprise agreement accounts. Whereas on the the other part of the AI is the agentic AI, the AI agents, and so on. The agentic AI, as well as self hosted AI tools, the AI chatbots, they are much, much easier for us to secure right now. What we do is that we have what a Polygraph Secure AI gateway.
Yagub Rahimov:In in this example, basically, have a cloud API key or you have a DeepSeek API key or Loma API key, whatever that you have. And you know that this is gonna be given to, let's say, the engineering team. And in the engineering team, you put the rules in there that API keys cannot go hit the LLM, or customer data, the credentials, etcetera, cannot hit the LLM. And then on top of it, put additional guardrails such as the prompt injection, jailbreaking, etcetera, related attack vectors. And then JSON validation for JSON codes or JSON database validation that you're putting inside for malicious, basically, front loading or back load loading activities and so on.
Yagub Rahimov:And and then you basically create a brand new API key via the gateway, also known as a base URL, and replace it what you would have used otherwise directly. Now anything that hits or coming on its way to LLM, it's first gonna be hitting this gateway, the guardrails that you may also call, the security measures, and then the result is basically being transmitted. There is a big problem in this scenario if you are basically utilizing large language models or other rejects, etcetera, different operations. If you don't have the context in mind, you're gonna mess it up because LLM makes something out of other things, so the other things get to be eliminated. The logic gets to be fuzzy.
Yagub Rahimov:The logic gets to be kinda shifted around it. So you have to really think about how you are doing it. And the second very big issue in this specific example is that make sure that your latency doesn't destroy your productivity gain. Today, unfortunately, in AI security field outside polygraph, we see fourteen to seventeen, twenty one seconds per inference latency because of AI security, which kind of like destroys the productivity gains that comes out of this kind operation.
Yagub Rahimov:So when you do this thing, automatically, I mentioned two environment, black box and agentic self hosted chatbots. Self hosted chatbots and agentic operations via gateway security divert and return back. It's a dual sided fixed, approved proceed, deny fixed proceed basically and then receive it back, verify, approve, proceed, verify, fix, proceed kind of approach that it takes place. You can do it yourself, there are some open source solutions out there too that would work to certain level of accuracy. The big issue in that example will be you have to be making sure that you are not falling a victim of attackers actually putting things on GitHub that are actually backloading some other attack vectors that you are not aware of.
Yagub Rahimov:This is specific I see it with the skills as an example. The world is talking about to the skills for website design base, to the skills for AI code, or what whatever else it is that you're gonna be using. But people fall on today like, ah, Eureka, you know, like the hallelujah moment. And then they deploy it a week or ten days later, whole system gets to be taken over. You specifically would see a lot today that the API keys get to be stolen.
Yagub Rahimov:Like suddenly Gemini keys and clock keys or ChatGPT keys are going somewhere else. And this is because of the backdoor that they opened it up.
Dejan Kosutic:Okay, now if a company really wants to start systematically to deal with these risks, they have to actually do some kind of a risk assessment, I assume, right? I think that part still continues to be very valid, right? So how do you actually perform this kind of risk assessment in this very new, let's say, environment?
Yagub Rahimov:Coincidentally, we actually put a completely free risk assessment calculator on our website polygraph.ai, if anyone wants to go check it out. I would say whether this or anything else, conduct your own first independent self assessed assessment. And then put aside a responsible person in the company, someone that you trust that will do the job fully. And quite often we see this person basically the deputy CISO or Red Team head or Red Team analyst or something like that. One person that knows your inside operation and who is also able to talk to other vendors like ourselves as an example.
Yagub Rahimov:Let them go to the conversation in there, find a party that can run a quick analytics on your own environment, not not just the words that we think that you are doing something. Let... Do not let them immediately sell you something. Like, think of it like an analytics environment. Even if you have to pay them, pay them something, but no obligation afterwards.
Yagub Rahimov:Let them prove that you are failing on something. Once this is done, then based on their things, they are supposed to be giving you proposals because now they know how you have failed or what you are gonna be doing it and so on. Another thing that I think companies should be doing is that the very first testing that they are doing, even if you have to run it on on prem etcetera, try to cut your costs low because that you don't want to spend months of deployment and engagement, etcetera, with someone that you don't even know if you're gonna be long term engaging in. So try to go as fast as possible, whatever the SaaS environment or whatever the vendor hosted environment that you have. Once it is done, then you are going into the proposal stage and planning stage and so on.
Yagub Rahimov:The goal is basically give yourself a ninety day from discovery to proposal to approval stage on your side as a minimum level, because generally speaking, you need at least one week to get people on board internally within your own operation.
Dejan Kosutic:You mentioned that security dependent should not on another company. So how do you actually achieve this, that you are not really dependent on a particular vendor?
Yagub Rahimov:It's easier said than done, honestly. Perhaps it's my personal values that I I would like to share in here that I I I find it very hypocritical for, let's say, an AI security company coming and saying that do not trust cloud, it's not reliable, or do not trust OpenAI. And then behind the scene, they're using their own enterprise cloud or enterprise something else to secure this one. That's the part that I'm saying, you cannot really trust on that. The secondary part is that I don't believe that in today's world where the world has really shifted in the last three years, one size fits the template security system that this is what we do, does not fit for everyone.
Yagub Rahimov:Whomever is providing something, they need to leave it to you, should they can have templates, should they can have suggestions, should they have best practices, etcetera. But you should be able to customize things based on your needs and wants and so on. And it needs to connect and be able to provide you to have the connectivity to your other reporting systems as well. At the end there is too much noise, you know, as Mr. Paranoid as I am, I get to see, I'm paying attention to every single warning that comes up in there.
Yagub Rahimov:Nature of our business, the parties that we are operating in. It's not about how many more warnings I can put in there. It's called noise, Noise is the killer of the productivity, honestly. It's all about finding a method and how do you give it to me, your service in a way that I really value that I see? And preferably if you are a large financial institution or if you are in a country that have the data residency requirements and so on.
Yagub Rahimov:Give it to me in my own environment. My data should not be part of a paradox of Hippocratic structure saying that don't trust him, don't trust that, trust me to trust you or something like that. It should be, don't trust your group, don't trust Aon, don't trust OpenAI, don't trust cloud, don't trust anyone, including myself. Trust yourself and trust the technology that you are the full host and full controller of. This is what I mean by saying, at least for the privacy part, this is the way that it needs to happen.
Yagub Rahimov:Otherwise, another big problem which we are not even addressing in here, I don't think that the cost is the big problem or challenge of AI kind of frenzy right now. Everyone is talking about, oh, I'm giving, I'm paying too much to AI this, AI that. But no one is talking about how much data that they are actually giving to other way or other side. We've seen it in the past, in our Figma case as an example, which was kind of shaken up with the cloud design or cloud or whatever it was called. We've seen these kind of issues happening with Harvey as an example, which suddenly came like a cloud law or cloud legal, whatever they're calling The data is the king of everything.
Yagub Rahimov:Arguably you can say that it's 10 times more valuable than your own operation. And that's again the part that's saying, own your security yourself. That doesn't mean that build it yourself. You will never be able to build as fast as dedicated companies like ourselves or someone else out there. There are dedicated dozens and dozens of engineers and researchers working on attack vectors every day.
Yagub Rahimov:Use it, in a controlled environment and customized environment that you really deserve.
Dejan Kosutic:Okay, but if everything is on premise, right? Or let's say in the environment that you control, doesn't this really slow down operations, slow down innovation because you're spending more resources on managing this kind of locally hosted infrastructures?
Yagub Rahimov:It depends on what you are doing. Not everything needs to be on prem. But many things can be fully on prem. Like the things that we've just discussed about it, it's a containerized environment and runs purely on CPU and RAM. If you are a team of, I don't know, 5,000 people and you have an Intel Xeon environment in your own operation, or even 50,000, doesn't matter.
Yagub Rahimov:It will add maybe additional $3,000 computational power for your own thing. Or if it is gonna be an investment that you don't have it, you're gonna be buying it fresh, and you're putting in an air gapped environment, it's going to be like 20 five-thirty ks one time investment. And you are the owner of it right now.
Dejan Kosutic:No, no, I understand from a financial point of view, but from the human resource point of view that you actually have to invest into, let's say, engineers that are going to operate this kind of production environment, this is an additional cost and actually additional time that you have to calculate in, right?
Yagub Rahimov:But why do you need to have again, are talking about agentic operations and if you basically automate everything that AI is designed to do so, you're actually reducing that number. Today, whomever we are working with ourselves, they don't... I mean, the highest number of dedicated engineers that are working with us that we have right now is two. And that company is a nuclear company. So before, with us or without us, they probably had a lot more than that.
Yagub Rahimov:It's all about having that visibility being complex, not complicated. So you've got to have a complex, comprehensive visibility coverage in your own activities. The incidents needs to be flagged accordingly. It all comes down to false positives and false negatives and reduce noise. And now there is another part.
Yagub Rahimov:If you're a full on prem operation, the big challenge is that cybersecurity is not a static operation. So it needs to be updated. There is a constant updates on it. A company, whomever it is that's running, it needs to build a kind of package upgrades that can happen within the company's own available time. And it should not take days to upgrade.
Yagub Rahimov:It should be very fast, kind of like replace this package with that one and it's good to go. That's doable, mean we've done it and I'm pretty sure other companies can do that too.
Dejan Kosutic:Okay, now, how do you actually audit these kind of systems that you are building? So basically security slash AI governance systems. How do you basically audit and make sure that everything is working properly?
Yagub Rahimov:So overall, if you're talking about the... Are you talking about the auditing the company or auditing the accuracy of the models? There are two types of audit that we can address in here.
Dejan Kosutic:Okay. It's more about auditing if the level of security is adequate for particular AI systems. So how do you actually audit that part with all these controls that were mentioning?
Yagub Rahimov:So there are a number of nonprofits and associations, etcetera, that are constantly always going through it and trying to benchmark things and so on to validate it. Like as an example, let's think and pick the smallest one that everyone knows about it, privacy control. Everyone can really talk about it. If you go on LLM Arena today, you can actually find out a number of challenges in there and whomever you are working with, have that vendor basically conduct that LLM arena test challenge that you're having and let them show you their own benchmark. We are happy to show our own benchmark on independent datasets such as NVIDIA or whomever it is that is providing.
Yagub Rahimov:So there are benchmarks in there and you need to be able to do that. Another very important thing that I think like from this will be a little promotional, we have an insurance company that rejects many of the privacy liability insurance applications with them if the company has agentic AI operations. But if you are that organization that has been rejected by them and you come on and adopt Polygraph, this insurance company will be giving you the with no question asked privacy liability insurance for your agent corporations as well. And it's because now we have already gone through the benchmarking and auditing etcetera with them too. You're nailing something very important, Dejan.
Yagub Rahimov:I'm sure you've been to Black Hat or any other cybersecurity event in Dubai right now. JISEC is happening as an example. You go there, everyone is like the god of the same thing, everyone is talking the same thing. But I want you to, if you're listening to this, like if they are talking about browser extension, ask them what do you do when I go on incognito mode? Simple question.
Yagub Rahimov:What do you do when I go on incognito mode? They're gonna come and say that, yeah, there is a kernel this and that and that and that and that. What do you do when when I go in incognito? That works up to 80%, you can fix it. What about about the 20%?
Yagub Rahimov:Right? The secondary part is that it must be via the browser extension or it must be a separate dedicated browser. What happens to my other applications that I should be using that your browser does not apply. So I really personally think that security needs to be inclusive in a way. Like you cannot be putting like a box in there saying that this, we are living in an AI era.
Yagub Rahimov:The innovation is out of this world. I don't know what tomorrow really brings and they don't know it neither. If we put your own innovators into a box, they're not going to be able to innovate. My job is to follow the innovation that's happening in the industry today every day and be able to find the failures in this and fix a patch on it, fix a solution on it and deploy it with my client environments. And my client's job is to innovate their own thing and increase their bottom lines and top lines.
Dejan Kosutic:Okay, good. Great. So let's wrap up the discussion for today. So what would be your top recommendations for companies if they want to secure their AI systems?
Yagub Rahimov:First, talk to your own employees if you don't even wanna talk to people. Give them a simple questionnaire like what are the top AI tools that you are utilizing? And then ask on that to elaborate. What do you use it for? What are the main problems that you are utilizing?
Yagub Rahimov:So you'll see the bigger kind of like a comprehensive repeat action that is in there. Identify one of the problems in the thing that is repeat and try to fix the security solution for that one. Learn out of it, move to the next one. This will be the easiest one that you can actually sandbox it ASAP. Within ten days you're good to go and you have a methodology to handle.
Yagub Rahimov:And then you'll see from there on other tools that are emerging. Invest more time for your visibility environment and do not trust and run on your network level security, network level listening and so on because that does itself or those themselves create additional challenge points. Do not, I beg you, please do not make your employees sign that if they make a mistake, they're gonna be fired and so on. It doesn't work. They're not gonna come and tell you when they made the mistake.
Yagub Rahimov:It will be a little too late.
Dejan Kosutic:Yeah. Thanks for these insights, Yagub. It's been a pleasure talking to you.
Yagub Rahimov:Thank you, appreciate you.
Dejan Kosutic:Okay, thanks again and thanks everyone for listening or watching this podcast and see you again in two weeks time in our new episode of Secure and Simple podcast. Thanks for making it this far in today's episode of Secure and Simple Podcast. Here's some useful info for consultants and other professionals who do cybersecurity governance and compliance for a living. On Advisera website you can check out various tools that can help your business. For example, Conformio software enables you to streamline and scale ISO 27,001 implementation and maintenance for your clients.
Dejan Kosutic:White label documentation toolkits for NIS2, DORA, ISO 27,001 and other ISO standards enable you to create all the required documents for your clients. Accredited Lead auditor and Lead implementer courses for various standards and frameworks enable you to show your expertise to potential clients. And a learning management system called Company Training Academy with numerous videos for NIS2, Dora, ISO 27,001 and other frameworks enable you to organize training and awareness programs for your clients workforce. Check out the links in the description below for more information. If you like this podcast, please give it a thumbs up, it helps us with better ranking and I would also appreciate if you share it with your colleagues.
Dejan Kosutic:That's it for today, stay safe!
Creators and Guests
