Cisco AI Defense: Product Update and Discovery

Oh alright so now what I'm going to request you is take a.00:00Few seconds to answer the poll on the right side. The intent is to customize both the content covered, right? And the demo which follows, right? We are going to tailor this session for you.00:15Right? And for that, we need your inputs. So unfortunately though, we may not be able to take a lot of time for these slido, answering the poll, right? So once you get going with that, I guess we will switch.00:30To the households, ok? I'm gonna give another like 15 s if you guys can hop in and answer those.00:47Alright, so I'm just gonna leave it open just another 10 s and then we're gonna kick off and get going with question to tailor the demo. All right? Alright, so.01:04Thank you. I'm gonna lock this in. All right, welcome again to the Cisco AI defense webinar. I'm central, your host and moderator. So I'm gonna go over some quick house rules, right? Once we get to the house rules screen.01:21Right, so now again this is something that we gotta share with you all, right? This is for your quick reading skill test, right? So, so that it is and let's go into the house rules, not much.01:42Real quick, so we are not going to use the raising hand for questions, right? The reason is you're going to use Q and A, which is at the bottom right hand side of your Webex window, right? There are three dots, you click on that and you'll be able to see Q and A.01:59Right? You use that to ask the questions. And the reason is that we have an army of experts, AI security engineers ready to answer your question as the answer questions occur to you. You don't have to hold on to those questions, right? You can focus on the demo.02:19The presentation, right? So like army of people ready to answer live on Q and A, right? So now why would you use the raising hand if at all, right? If you are not able to hear anything that I'm saying, alright, you are running into some technical issues, you raise your hand and.02:36One of the panelists would reach out to you.02:57Directly and potentially help you solve that technical issue. All right? Okay, so with this, right, I mentioned that we gathered some information from the poll to tailor the actual presentation which is coming up. So I'm going to just put in a poll to help tailor.03:00The demo itself, right? And I'm going to leave that on for the next 1015 minutes, but I would request you to answer that as soon as possible, and that is to help us tailor the demo, right? Alright, so now with that.03:20I'm going to hand it over to two of the presenters. The presenter who will be going up 1st is our AI security engineer, Shikhaar and the other engineer who would also, the AI security engineer who would be demonstrating the product is Yanos. 1st up, I'm going to invite.03:37Shikar to take you and walk you through the AI defense intro. Shikar, over to you.03:57Thanks until. Great introduction, so let me quickly start us for today's session. So good morning, good afternoon, good, good evening. Depending on where you are joining us from. So over the next hour, we will focus on a very practical question that many organizations are asking right now.04:05How do we adopt AI quickly but securely?04:24AI is already changing how businesses build applications, how they support customers, how they automate workflows, how they analyze data, and basically how they improve productivity. But as AI adoption accelerates, it also introduces a new class of risks that are traditional security.04:28Tools were not originally designed to address. So today's session is basically structured into three parts. 1st we'll try and spend some time discussing the AI risk landscape and how AI defense helps secure your enterprise AI deployments. Then we will move into a deb.04:47Demo where we will show how the product supports discovery, detection and production across your AI life cycle.05:07Finally, we will leave some time for Q and A So the goal is to leave this session with a clear understanding of why AI security is becoming critical, what makes it different from traditional application security and how Cisco AI defense can help your organization move forward with confidence.05:15So, starting the session basically, when we talk about AI risk in enterprise, it's useful to separate it into two broad categories. The 1st one is 3rd party AI tool usage. This includes your employees using, let's say public or external AI tools such as chart GBT.05:37Gemini, Notion AI or any other SAS based AI tools.05:57These tools can create risk around sensitive data leakage, around intellectual property exposure or let's say compliance and shadow AI usage. Cisco basically addresses this area through capabilities such as Cisco secure access, helping you manage and secure the employees.06:01Device use of 3rd party AI tools.06:212nd category, this is the 1st party AI applications. These are the AI systems that your organization builds or deploys directly. They may use internal data, they can connect to business systems, maybe they can rely on models, agents, your vector databases, API.06:25These are sometimes 3rd party components as well. So this is where Cisco AI reference comes in picture.06:45We basically help organizations secure the development and operation of 1st party AI applications across the business. A simple way to frame this would be secure access helps manage employees use of external AI tools and Cisco AI defense helps secure the AI applications. Your business is building and running.06:52Basically. Now, AI adoption, it creates new risk because AI applications are fundamentally different from traditional applications. Let's say a traditional application is in picture. It usually follows a deterministic logic.07:12If a user clicks on a button or they, let's say submit a form, the application follows a defined path. AI applications are way more complex. They are non deterministic, meaning the output output can depend on let's say the prompt or model behavior context which you are giving.07:29Integrations which you have or even the prior conversation history.07:49So what is the risk? The key challenge is that AI applications do not behave like traditional applications. A user can interact with an AI system in almost unlimited ways. The same user can rephrase some requests a lot of times basically. Malicious user can attempt to manif.07:55The system by changing the wording or let's say they try to roll play, they try to break a request into multiple steps or they try to use indirect prompt injection. So all these we'll cover during today's slide, but just a quick overview. This creates a new risk vector.08:16Let me give you some examples. So, let's say you have a customer facing chatbot which maybe designed to answer product related questions, but if it's not properly protected, users may try to manipulate it into making unauthorized comments or revealing internal instructions or let's say gen generating.08:35Inappropriate content or interacting with back end systems in unsafe ways. A well known public example involved a car dealership where the chatbot was reportedly manipulated into agreeing to sell a car for $1. Now this is important listen and it's not for the specific $1.08:54The lesson is that AI applications can be socially and technically manipulated into, let's say, producing outputs that the business never intended to do.09:14Another publicly discussed category of examples involves branded chat experiences. So being prompted to move outside of their intended purpose is what we've seen, such as let's say generating code or responding to ways unrelated to the business, and this is very common and we are.09:25Examples let's say related to food industry where the chatbot is now being prompt engineered to generate code.09:45These examples show why enterprises need security controls that understand AI specific behavior, not just traditional web, application security. So AI risk, these are increasing with all the capabilities which.09:53They're coming up.10:12Let's say a simple AI chatbot. It may answer questions based on limited knowledge based. We now have rag applications. They can access your internal documents, they can access your policies, product information or even customer data which you have. With that we also have.10:14Into AI applications which can go even further, so it can make actions, it can call tools, interact with applications, query or databases, it can create tickets, summarize records, and most importantly, it can trigger workflows. So each step makes AI more useful.10:33But each step also increases the risk. As AI systems gain access to sensitive data and there's more autonomy given to them, they become more attractive targets. The potential business impact also increases because the AI system is NO longer just generating text. It maybe interacting with.10:53Real enterprise systems. This is why securing AI deployments is not only about blocking bad prompts. It's about understanding the full AI life cycle.11:13Including the model, data, tools, agents, and runtime behavior. Now, when we are talking about AI security and the ecosystem, organizations are currently increasingly looking.11:23To industry frameworks and standards to understand AI specific risks. We have oas, we have MITRE, we have NIST, and these frameworks help define and categorize AI security risks such as prompt injection, data extraction, model manipulation, poisoning, supply chain.11:40Risk let's say adversarial behavior. So for the enterprise security team, this is important because AI security needs to be measurable.12:00Repeatable and aligned to the recognized standards. We provide validation and reporting that map to AI security standards so the organization can better understand, communicate and prioritize AI risk. Moving from there, if let's say.12:09The AI risk is not managed. The consequences, they can be significant.12:29There are not only technical issues, they can be, let's say legal or compliance or customer trust brand reputation. So I'll give you few examples. Let's say if an AI application exposes confidential information, that becomes a data sense security and privacy issue. If like.12:35Say the chatbot provides unauthorized or incorrect commitments to a customer that can become a reputational issue or a business issue.12:55If let's say an agent takes an unsafe action through connected tools, that becomes an operational and security issue. So this is why AI security requires cross functional collaboration between security, AI development teams, data teams, legal and business stakeholders.13:03So there are four reasons why A security is so difficult right now. The 1st one, the technology is changing very quickly. Models agent tools, and attack techniques are evolving at a pace that traditional security, they are struggling to match.13:24Basically, your AI application often involve multiple groups, so they can be data science team, they can be your security team, your cloud team, maybe your legal or compliance. So if these teams do not have a shared visibility, AI risk can fall through the cracks.13:43Then let's go to the next one. Manual validation, let's say and they can be very expensive and resource intensive especially when it needs to be done repeatedly every time the model or prompt or the application behavior changes.14:02And the last one, AI security is a specialized field right now. Many organizations have strong application security teams, but AI specific risks such as prompt injection, model back doors, data poisoning or agent misuse, they require new skills. Cisco AI defenses.14:19Trying to reduce this complexity by giving organizations a structured way to discover your AI assets, then let's say detect vulnerabilities inside those and finally protecting EA applications at runtime.14:39As discussed, enterprise AI application, they often rely on 3rd party components. This may include your open source models. This can include your 3rd party data sets. It can be MCP servers, tools or let's say repos or libraries. So these models.14:56Tools may contain backdoors or malicious behavior or let's say the data sets which you are using, they maybe poisoned or they may contain sensitive information or let's say tools and servers maybe introduced to vulnerabilities or let's say unsafe access parts. So this is why I see now.15:16On the AI security, it must start before the deployment. It needs to include supply chain risk. It needs to include your model and data set validation and controls that can prevent insecurity assets.15:36From entering your production workflow. Hopefully I was not going through the content too fast. If you feel like you want me to slow down or go a little bit faster, just put it in the chat. I'm trying to look at the chat as well.15:50Here, let me quickly remove the pen and I have a quick demo. So many AI applications, they rely on internal instructions. They are called system prompts. So to define basically how the assistant should behave, this is what we use. If those instructions are except.16:08Post attackers may learn how applications work and may use that information to bypass controls. So you can see on the screen I have a running example where we were able to just write a simple prompt and we got the system prompt back.16:28This is scary. Let me go to another one. So let me run this. Basically many rack applications, they use vector databases and embeddings to retrieve relevant content. We'll just look at the simple rag.16:44The application we built which allows users to upload PDF documents and let's say ask questions about that PDF document.17:04So we uploaded a very safe document which was looking to us like it's, it's nothing it's safe, and still there was a small hidden link which was there in the PDF. So after uploading the document and asking it a simple question, we got provided a link, we clicked on it, and that link is my.17:12Issues looking to exfiltrate data from our systems. So these kind of attacks, these kind of scenarios we are seeing commonly nowadays.17:32Here I have a 3rd example. This is for supply chain vulnerabilities. So AI application, they increasingly depend on external models, data sets, tools, and repositories. If those components are compromised unsafe or poorly controlled, they can introduce disk into the.17:43The entire AI application. So.18:03What we did is we just re uploaded a file to HuggingFace and you could see like the vulnerabilities were not scanned yet. If let's say you try to pull that model in, you can see what happens. We are redirected to that malicious link. So all these are examples.18:05Real world which are happening nowadays like this is what the threat looks like. This is what you need to be protected against. So we've seen here applications can be attacked through prompts, data, models, tools, plugins.18:23Infrastructure or even supply chain components.18:43We provide a framework to reduce these risks through discovery, detection, and protection. With AI defense, the message is simple, as enterprises build and deploy AI, they need security controls purpose built for AI.18:46Traditional controls remain important, but AI introduces new behaviors and new attack paths.19:06Cisco AI defense helps organization answers critical questions such as what assets do we have or let's say what models, what agents, what databases we are currently using or let's say what applications and models we have which are vulnerable.19:13We can ask questions like 3rd party AI components are they safe or not?19:33Can our AI applications be manipulated into, let's say, answering in a way we don't want it to or are we protecting the production AI applications in real time? So Cisco AI defense is designed to give organizations the visibility validation, and runtime protection needed to adopt.19:38AI securely.19:58Here at Cisco, we've organized this into a simple three step framework, the discovery detection and production which I've been repeating. So 1st, you cannot secure what you cannot see. Discovery helps uncover AI assets including your models, your agents, your knowledge bases, which are.20:02Distributed across your network.20:22Second, we have detection. Once the assets are identified, organizations need to test for AI risks. They need to test for vulnerabilities and sustainability to attack. This includes your model and application validation through automated red teaming.20:24And finally, after the risks are understood organizations need to put guardrails and guardrails that protect data and defend against runtime threats. This includes blocking malicious prompts or let's say unsafe responses in real time. All of this is supported through a unified management.20:43With Cisco Security Cloud control.21:03This framework helps move AI security from reactive to proactive. Awesome. If you have any questions, guys, feel free to use the Q and A panel. I'll take a quick water break.21:06Right hope we are good. Now moving on, Cisco AI defense, it provides coverage across AI lifecycle through four main capabilities and let me bring my pointer here. So let's start with the AI cloud visibility.21:25This basically helps identify AI assets, including your models, agents, and connected data sources across your distributed cloud environments so teams can understand usage and gotcha the risk.21:452nd part, which is the AI supply chain risk management. This capability scans model files, repositories, and MCP servers to help identify threats and proactively block malicious or unsafe AI assets before they impact the operation.22:01Coming to validation, this helps identify safety and security vulnerabilities across the models at scale using algorithmic red teaming, which we are gonna deep dive into later. And finally, the runtime production, so this protects the production.22:21AI applications and agents with guardrails embedded in the network helping block attacks and harmful responses in real time.22:41Together these capabilities support the full journey from visibility to assessment to enforcement. Now let's quickly introduce these features and we are gonna have a very interactive demo by.22:50He'll be covering all of these in depth on his demo, but I'll just quickly introduce them to you right now.23:10So let's start with the 1st one which is AI cloud visibility. One of the 1st challenge our organizations face is simply knowing where the AI is being used. AI assets may exist across cloud environments, development teams, repos, applications, and business units. What.23:16AI cloud visibility helps you is basically we automatically uncover AI assets across your distributed cloud environments. This can include your models, your agents or connected data sources like knowledge bases.23:36This is critical because unmanaged AI assets can create blind spots for you. If security teams do not know which AI application exists or what data they are connected to or let's say where they are deployed, it becomes very difficult to govern and protect them.23:50AI cloud visibility gives organization the starting point they need, which is inventory and context. The next capability is supply chain risk management. So AI applications often depend on components that come from outside the organization.24:08That may include models, data sets, repos MCP service, tools, and supporting code.24:27Cisco air Defense helps scan models in private repos to identify vulnerabilities such as code execution and suspicious imports. It also helps scan MCP servers to basically have inventory of the tools and detect these tools for poisoning attacks as well.24:35The goal is to prevent insecure models and 3rd party EA assets from entering or impacting your production environment. This is especially important because AI supply chain can be introduced before the application is ever deployed.24:54If a model tool or data set is compromised at the source at the enterprise may inherit that risk unless it has a way to inspect and control these components. Now moving to validation, so AI model and application validation.25:10Helps organizations identify vulnerabilities in the model and applications through automated.25:31Algorithmic red teaming. This is important because AI vulnerabilities are often behavioral. A model may appear safe during normal testing but behave differently when challenge with adversarily prompts, multi tenant interactions, indirect instructions or attempts to extract data.25:37We generate a report that maps to AI security standards. This can also help generate guardrails that address specific model vulnerabilities and better protect the AI application. The value here is not only finding the risks. It is tuning and turning these validation findings.25:56To an actionable protection.26:16Now with Cisco AI defense, we support risk focused value validation. This includes multi turn validation and custom goals. This is important because real world attacks are often not single prompt events. Let's say a user starts with a harmless question, then gradually steers the conversation towards something which is un.26:21Safe. This is called multi ten scenario.26:41We also simulate realistic conversation attacks that evolve over the multiple terms. With customizable attack goals, we allow the validation to be tailored to your business context. A software company may want to test whether the agent can be pushed into an unsafety tool usage.26:44So we enable you to validate against the risk that matter most to your business.27:04Algorithmic AI red teaming, it uses AI to generate and refine the prompts that test the AI system. So you can see different prompt variation that help to basically move towards an unsafe outcome. Attackers, they rarely stop after one field prompt. They rephrase, they try to roll.27:11They break the request into smaller pieces or they try to exploit the ambiguity.27:30We use the algorithmic techniques to test how A application respond under those conditions. This helps you find the vulnerabilities before the attackers do. It also helps reduce the time and effort required for manual testing.27:37The outcome is clearer understanding of where AI application is resilient and where guardrails need to be improved. Now, coming to AI teaming, it can be done manually as well. The point is it can take up to seven to 15 weeks to test just a single model.27:53Depending on the scope. That work may include, let's say, identifying the frameworks.28:13They can be designing and writing test code. They can be setting up environments and infrastructure or creating wrappers and integrations or let's say running the individual tests or comparing different models and this is not a one time activity.28:18Testing should be repeated whenever the model or application is modified during the development. This is where our automated and algorithmic validation becomes very important. Now we've discussed validation, it's very essential, but the protection at runtime.28:34It's equally important. AI runtime protection allows organization to define bi directional guardrails for applications and agents. Bi directional means protection applies to both sides. User prompts going into the AI application or the AI responses which are coming out of the application.28:54This is important because risk can happen both directions. A malicious or unsafe prompt may attempt to manipulate the model or let's say an unsafe response may expose sensitive information or provide inappropriate output or maybe just violate the business policy.29:14AI defense helps configure guardrails that cover specific model vulnerabilities and fit unique AI application. This also helps protect against rapidly evolving AI threats, including threats to MCP servers. The key, value here is we are protecting at real time, you are provi.29:32Production A application and agents.29:52Now, we've discussed so much AI risk appears at multiple decision points in AI lifecycle. A public customer facing chatbot has a different risk profile than let's say an internal summarization tool. The model maybe proprietary or open source or let's say.29:57Tested or customized. So these options introduce different security and governance considerations. Customizations may involve sensitive business data, fine tuning, let's say implementation of rag, embeddings or prompt engineering. This can introduce risks around poisoning, extraction and.30:16Leakage.30:36Now we've covered the product, but what are your deployment options? So, I'll quickly cover these. We basically support multiple deployment options to meet all the customer and their needs. The 1st one which we have here is the Cisco cloud. So in this model, data is sent to the cloud.30:39And back to the customer's environment.30:59This is best for customers who are looking for a simple flexible deployment with zero infrastructure to manage. The next one, which is VPC. So in this, the model data plane traffic never leaves the Cisco leaves the customer's cloud environment. So this is best for customers looking to balance.31:01Data control and compliance with cloud scalability of theirs.31:22And finally, we have data centers. So in this model, basically the data plane traffic, it never leaves the customer's data center. This is best for customers that want to manage AI workloads themselves rather than relying on hyperscalers.31:27Here I have a capability matrix which helps summarize which capabilities are available across the different deployment models. I'll quickly leave this here for like say 10 s and we can share this as a link for you to bookmark as well in the chat, so keep a look out for that.31:44With that, we are at the fun part, which is the demo. So let me hand over the ball to over to you. Thank you. And that was an excellent overview of the key use cases, discovery detection and protection, and you also threw in the deployment option, a quick overview on that. Excellent.32:04So thank you again for the folks who participated in the poll to help customize the demo. I've been sharing updates and snapshots of that now and then to Yanos, so just running a live demo in itself is.32:22Challenging and Yanos said, hey, you know what, I'm going to even customize it for the participant's preference. I just hope the demo gods are going to be with him. Good luck to you, Yanos over to you. Thanks a lot. Are you just sharing my screen?32:37And Congratulationshikar on the awesome presentation on setting the stage for us. Thank you. Thank you Senthil. Yeah, together we'll see AI defense in action and as you can see from the from the results of slider.32:53You're more interested in applying the to block prompts and responses. Be sure that we'll see that and I will show you how we can do it with some customer applications that are built on my own. And with that let's get started. So, as you can see I'm already already logged in in my security cloud control account, which is.33:11The place where we go to have access to AI defense, if we have this deployment already configured, so the 1st option that she had explained in the last slide so everything is Cisco cloud hosted. So I have it here.33:31I will click on it and I will be straight away into the dashboard. So as long as we manage to log in, this is what we see 1st. Now.33:47Today at the demo, I will show you the three main use cases that explained. So we'll see what is visibility, how we can validate a model together, how we can protect an application, and I will come back to this dashboard at the very end where I hope that you will be able to read everything.34:03Without me explaining and what do I mean by that what why some applications are protected and not and some not? What are the, how can we have the 40 agents and the model that you have and and the data sets, but I will come and close the session with the dashboard that we initially see.34:23So if you remember and she can explain very well, we cannot protect with what we cannot see. So with that being said, we'll see the 1st use the 1st use case, sorry by next and you see we have two options here.34:43In for the sake of the demo today, we'll focus on 2nd one, so the multi cloud defense visibility. But as you can see there is an option called AI inventory, and let me briefly explain what those two are and what are the differences. So if you are a multi cloud defense customer.34:59You have integrated your public cloud accounts there. So AWS GCP, Azure and this is 2nd issue where we will see I have my own AWS account. I have integrated my AWS account with multi cloud defense and I've connected my multi cloud defense with a.35:16By defense and I will show you the tab in a bit. The 1st.35:36Option called AI inventory is the use case where you don't have multi cloud defense running. You don't have you haven't integrated actually your accounts there and you want to integrate it straight with AI defense. That's for the sake of the demo, let's click on 2nd one.35:40So here, a piece of information that comes and it's powered by multi cloud defense, I can see my AI assets, all of the 732 of them. Yes, I explained before I only have an AWS account.35:56What if I had a GCP or an Azure one, this is the place where I would see that info. So 700 of my assets are models, AI models, two of them are agents. I don't have any data insights. Neither I have a deployment.36:12I'm just to explore one of those so by scrolling down a little bit, let's just around the model. Let's check this one. Okay, what do I know about this model? It's hosted here. The provider is this one, the platform in AWS, and I can see some information.36:28About the model, the model itself.36:48Here I haven't used this one, that's why we don't have more, and I haven't validated this one. That's why I don't have the results here, but we'll do that in 3 min time. And if you have seen this other tab called external visibility, this is what we call shadow AI, currently I don't have any.36:50Traffic flowing here.37:10But I would have, if I had any machine, any VM, Windows Linux you name it, reaching out to an AI FQDN, fully qualified domain name. So if I have some users reaching out to, I don't know, whatever, here is where I would have seen.37:12That traffic but I would I wouldn't see user session details.37:32I would see instant session details, so I would know what was the VM that reached out to that to that domain, to that AI domain. If I wanted to have user session details, if I, if I go to dashboard once again, you see this widget here, I should have integrated my secure access account and through that integration.37:36I would have that info as well. For the sake of the demo, I will not focus on that integration, it can take a session on its own literally. And that includes the visibility part. So right now we do know what we have to protect.37:56And what we have in our environment use case number two is the validation. So we are validating or let's say we are trying to attack models, applications, endpoints to see where they are weak and not.38:13Briefly explaining this stuff and how this works, we need to add a target, of course, who do we want to test, right? Then a profile stating the way that we want to test it, and I'll show you profile and a target, and then if we want to add a custom goal that.38:31Explained before. I for the sake of the time I've already created target profiles and custom goals. Today we will validate together a model from AWS called Amazon micro. So I have already configured this. I got the model ID.38:51Just by navigating to my account and finding that specific model that I want to, to validate and I will show you very very quick if I go to Amazon, for those ones who are familiar with I'm sure you know what I'm doing here. It's, it's the number micro, right? Yep. So if you just.39:09Click on this one you can see the model ID here. I have configured it.39:29The next step should be adding and creating a profile on the way that I want to configure it. I called one called extensive validation and it's extensive because of sorry, it's tensive because it's using all of the content categories. We have, I believe six in total, Yep.39:33And I also added the custom goal and I want to see if the model will reply, let me go to the custom goal to show you if the model will reply to a question like this one. By having those three, I can go and run validation. So.39:51Demo, webinar validation. I will find my targets, I will set the profile, and I will I will select the attack volume. I have two options here standard or a quick scan.40:07A standard is a set of validations of around 2000 if I'm not mistaken, happening one after the other while the quiz scan has fewer validation and it's faster, but it's not as extensive as the standard one. I will not touch any other settings and just to give you an idea here, we do have it.40:25And Cisco focusing only on creating this kind of validation set.40:45I will click on the this will typically take 2 h stops, but of course I have run some before and by seeing that it means that it's successful and it's running, so we will see that here. I can see two different. The one is focused on the custom goal that I specified.40:50While the other one is based on that pool of validation tests that we have and let's check that one that I ran some time ago earlier today and actually let me just change it to dark.41:08Just because it's a bit cooler. Ok. Let's go to the validation. Now I can see that this model, you see out of 2000 attacks, guys, we're talking about 2000 attacks on an AI model.41:24Happening in 2 h. Yeah, I will leave that setting. Now one out of four more or less was blocked only. And by navigating to the techniques and the techniques and the content category I will not go.41:41Through all of them just for us to see one, just to help you understand what we are doing here. Let's go to talking exploitation, probably this and this. Okay, I want to show you one that is an alert. Okay.41:57So by adding one of those validations, we can see of course the techniques, sub techniques and the procedure being used. We can see that every technique that every validation set that we've used is mapped on OS and by clicking here I can see the prompt that AI defense sent to that model.42:14And the response that we shouldn't have gotten, and this is why this is an alert. Now we'll we'll not go into the all of this response here. But just to give you an idea what is happening in the background when we're talking about validating.42:34A model an application, an endpoint. For the sake of the demo I'm validating a standard model provided by AWS. So, we know what we have, we know the vulnerabilities of what we have. What's the next step? What do we do?42:51We need to secure them and we would like to secure them on runtime because we're talking about AI applications that our users are using a lot by for doing that, the procedure in AI defense is a following. We need to connect.43:08The application in AI defense. And we can do that in two ways. And if I go to applications I can show you. Either an API connection or a gateway connection. What are the differences? And I will start with the gateway, which is easier. A gateway think of AI defense acting.43:24As a proxy between your prompt and responses between your.43:44Users and your application endpoints. So whatever is going from the user to the end, to the application, whatever is coming from the application to the user or prompts and responses are going through AI defense and AI defense will allow or block it based on the policy that we will specify.43:49API is the connection where we will request AI defense, we will send to AI defense and prompt all the response or both and get a very quick. AI defense will tell us if this is malicious or not, and then it's up to the developer.44:08To decide what to do with that information. Without any further delay I will show you exactly what do I mean. I have connected this application, it's here, I have specified a policy which I will show you right now, this is a policy.44:25For the sake of the time, I will not go through creating a policy, ok? It's, it's it's it's easy but it's it needs some time to do that. No more than 5 min. But for the sake of the demo I have created that one. I've connected everything, it's up and working and we'll see that in action. So.44:44This is the application to help you understand what I'm doing. Let me just refresh and make sure that everything is working. I'm using a model in the background. We don't care what that model is. And I've connected that application to AI defense by specifying this.45:01So I'm blocking prompt injection and prompt unblocking email addresses both directions, every PII, PHI, PCI, cat information identity information.45:22And some of the our safety gapters. But let's see that in action, right? Let's see what do we do. I have connected this application using the gateway connection and I've already created an attack which should be blocked.45:36We don't care what their targets, it's upon the and it's here. So I can see that AI and how can I see that in AI defense? There is a tab called locks, sorry, events and events locks.45:52If I go to event lock, this is it. You see this time, correct time, the correct application. This is what the user tried to send to the application and it's blocked because it violates.46:08Some rules which are mapped in those two standards and I can see some more information, even the model that my application is using in the background. This is the if I get as an admin of the product.46:25And one more thing as I explained before, 2nd way is connecting an application through an API. What is it? I have already created an application, we don't care exactly what it does.46:39But I can tell you that it's sending a malicious prompt and it's called API test. So i'll try to run it. And you can see let me just one you can see, I know that there's NO but I will show you a better version later. The only important thing we can see from the JSON format of the reply.46:54So to help me understand, I've sent something to.47:15An AI defense replied with a vertic if it's malicious or not and why. The only thing to focus here is this one. It's not safe. Now, for us to visualize it a little bit better, I created another, the same actually script, but I've modified.47:19The JSON format, the JSON reply. So it's the same reply, this thing here, but a little bit easier to be high. I can see that whatever you try to send me should be blocked because it's not safe because it violates.47:35This classification categories and these rules with this event ID. And again, the same way, if I go to logs, let me just refresh it. I should if everything works, worked well.47:51I should see the last block being this one coming from this beautiful. Okay, this is it. This is what I try to send and it violates these rules and which map maps the to the standards.48:07Okay here we were violating a lot that's on purpose, and before I conclude the demo and gives the ball back to let me show you some of the integrations that behind the AAI defense. One of them being Splunk.48:23So if you're familiar with Splunk, you do know that logging in Splunk is pretty cool and quick, so we can even build custom dashboards in Splunk to ingest this information that we see. For the sake of the time I've created one I will just show it. I will not explain what what we can see there. But I've created my own dashboard in.48:42I've created I've integrated my Splunk account with AI defense.49:02And I can see all that info plus some things that I added on my own. This is a custom dashboard. And if I go to my last events, I have created an index, everything is going to that index. And if I explore the last one, which is the API.49:06Application with his name. I can even see the message that I've tried to I've tried to send. So we can see everything from Splunk, we can see everything from AI defense. There are way more integrations for us to explore here, GitHub.49:26Secure access MCD AWS for the validation, some of them are already on my account. And with that I go back to the dashboard. There are more capabilities to discuss and show you guys, but for the sake of the time and since this our 1st session, we try to keep it simple.49:41E.g., the MCP stuff, what you see here is scans or the policy studio, the this can literally be one or two sessions after that. So hopefully now you can understand better what the dashboard show us. It shows us.50:00The events that were blocked or not, applications that have policies assigned or not protected and unprotected if you have agents, models data and so on, and if you have the secure access integration working.50:14With that, I hope I gave you a good understanding of the product. It's a product that is constantly updating as the AAA security landscape, and I will stop sharing and pass the ball back to Shikar. Thank you very much.50:30Thank you. And I loved your demo like where you had the JSON I always struggle to understand what is it sending me and you beautified it and it was so simple. So I'm gonna reach out to you and grab that application for me as well. But yeah.50:49So go ahead. Thanks. And yeah, we are at the end of the session, so keep your questions coming we are still monitoring the Q and A panel. Now let's discuss the Cisco advantage.51:06Basically, that comes from three areas. The 1st is platform advantage. Cisco air Defense is designed for fast low friction deployments that does not require modifying the application and it can enforce policies across within the cloud and data centers.51:23That, 2nd, which is AI model and application validation. Cisco AI defense uses algorithmic AI red teaming to automate assessment of safety and security vulnerabilities. It supports AI readiness, helps guardrail and enforce policy.51:42And it can integrate into your CICD workflows through automations for continuous testing.52:02With that, the 3rd one, which is proprietary model and data. Cisco's AI security capabilities are purposely built for AI security. The key takeaway is Cisco AI defense combines AI specific security expertise with cisco's broader security and networking strength.52:08Awesome, so before we close the presentation portion, I want to bring us back to the central message. Enterprise AI, it's moving very quickly. Organizations want the productivity, automation, innovation, and the benefits which come along AI, but they need.52:32To manage the risk. Those risks can include data leakage.52:49Prompt injection, unsafe responses, and many more which we discussed throughout this session. So what Cisco AI defense does, it helps your organizations secure enterprise AI deployments through three step approach which we discussed. Discover the AI assets, detect the vulnerabilities, and then protect the app.52:53Applications at the runtime. If your organization is building or deploying AI applications, Cisco AI defense is designed to help you securely and confidently do that. So with that over to you.53:13Right, thanks a lot for a nice custom presentation and thanks to you for customizing even the demo, right? As I said, just having a live demo in itself is hard and you customize the demo based on participant input and luckily it went well, right?53:29Oh God with you today, right? So now with this, right, we have about 4 min, right? So please keep sending your questions in the QA and Q and A panel, right? Given that we have a couple of minutes, we're gonna do two things. One, I'm just gonna ask this one poll question on what you thought about.53:49Up this session and I'm gonna kick it off, right? I request you all to try, take a few seconds to take care of that. So now, couple of questions were referred to our army of experts who are answering your questions in the Q and A panel to answer live, right?54:09So I'm gonna just pick up one of two of the questions based on the time we have and have Shikar and Yanos give us some inputs on that, ok? All right, so I'm gonna let you guys decide who wants to answer the question. Hikar and Yanos question which was referred to me is.54:29Why do we need AI specific security? Don't existing security tools already cover this? The existing security tools being I guess the traditional security tools. All right, I'll let you guys go ahead with this.54:48Anyway, yeah, I can take that one, I think and maybe you can pitch in as well. So what my understanding is existing security tools remain very important, but A application introduced new risks. That traditional tools were never designed to fully address.55:04Systems, they are non deterministic, they can be manipulated through prompts, they can be manipulated using multi turn conversations, indirect instructions, and all the things which we discussed, model behavior, data retrieval, tools, agents. So this is why we need.55:25A tool, a security tool for AI specifically, and that is where Cisco AI defense comes in. So it is purposely built for AI security. It helps discover the assets, validate the models and also like enforce runtime guardrails to block these malicious prompts and unsafe responses. So.55:43Hopefully I covered that. Maybe anyone wants to add. Nope, thank you so we have just a couple of minutes. I'm hoping to squeeze in just one more question that was referred to me before we close this out, right? So 2nd question, again, I'll let you choose who wants to go with this.56:04What types of AI applications does Cisco AI Defense help protect? I can, I can take this one. AI defense is a solution is a product built for businesses developing and deploying a 3rd party AI application.56:22Questions.56:41So we are we are talking about enterprises using as Chicago showed in the in the presentation before using any models 3rd party provider or their own customizing them also fine tuning them and then deploying that as their app AI application to be used either by their internal employees or for their own.56:42Customers and partners.57:02So to summarize the question, yeah, it's for businesses deploying their own 1st party AI apps. Right, thanks a lot so we are almost there like you know at time, so I'm gonna thank you all for the feedback that you guys have provided by.57:04Polling. Thanks for joining us today.57:24Thanks a lot to Shikar and Yanus for doing an excellent job with the presentation and the demo and the army of the expert panelists who actually help with the Q and A session. Thanks again for joining us. If you have any questions, you know, please contact Cisco's AI defense account team and look forward for future webinars where we will dive deep.57:26Deeper and add more information to help you all understand air difference and put it to good news. Have a great day. Bye for now.57:46Please answer the survey once, you know, the session ends, you know, just a few questions will help us tailor future sessions for you. Thank you. Bye.57:54