Name: From Shadow to Secure: Managing AI Risk with Full Cloud Context
Uploaded: 2025-10-09T16:50:05.364Z
Duration: 24 min 16 s
Description: From Shadow to Secure: Managing AI Risk with Full Cloud Context

Transcript for "From Shadow to Secure: Managing AI Risk with Full Cloud Context": Alright. Welcome, everybody. We'll get started in just a minute or two. We'll have some folks filter in, but stay tuned. We'll get started in just a couple of minutes. In the meantime, as we're get letting people flow in, get started, grab your coffee, grab your after work beer, depending on what time it is for you, if if you have questions during the session, you can ask them in the chat box on your right hand side. We've also got the docs tab right next to chat where you'll find some resources related to with AIS PM. We are recording today's call. I will repeat that a number of times throughout this webinar. So you should get the on demand recording in your inbox by tomorrow morning. If at any point you're interested in learning more about Wiz, hit the meet the meet with a wizard button at the top, and we will connect you with a wizard after the event. So we will get started in just another minute or two. We'll let a couple more people flow in. But if you've got any early questions or wanna check out our docs, go right ahead. Okay. I'll give it about thirty more seconds. One minute. We know it's a quick AIS PM webinar, so, don't wanna don't wanna give it too long, but I wanna make sure everyone has a chance to join. Alright. It looks like we've got a quorum. So we will get started. Hello, everyone, and welcome to today's fifteen minute demo on live demo, in fact, on AISPM with Snegha and Aviel. We're excited to have you join us as well as them join us today, and appreciate you taking the time out of your busy schedules to be here. There are a number of different ways that you can engage with us during the webber webinar. Excuse me. On the right hand side of your screen, you'll see chat docs and q and a. We'll be monitoring the chat for any questions that you might have. Chat and q and a are kind of the same thing, so feel free to drop your questions or your chat in either one. We also have our docs tab where you can find some other was related materials for a s p m a I s p m. Excuse me. Not to be confused with application security posture management, or a s p m. So if you we also have in the docs tab some as I mentioned, some helpful resources. If you wanna meet with a wizard, or learn more free feel free to hit the meet a wizard button on the top right hand corner of your screen. And with that, I am going to introduce our product marketers, Snegha, who will start going through the demo for you. And, again, don't be don't be shy. Ask some questions, and we'll get this moving. Thanks, Laura. Hi, everyone. My name is Snegha, and I'm on the product marketing team here at Wiz. We're really excited to talk to you today about AI security posture management, and I'm joined today by Aviel on the product team to both hear your feedback, questions, and where we're going. So let's get started. Why are we talking about AI security? I know that AI has been the big thing in everything that we talk about for innovation, new services, and new complexity. But there are some data points that I like to call out when we're thinking about AI and security. First, we've seen that over 85% of the organizations, even as Wiz, are using some form of AI. This really shows a adoption rate that has been unparalleled. Second is thinking about the rate of adoption. If we take the example of DeepSeq, the curve of how quickly it was adopted is something that we haven't seen before. This in combination with the kind of risks we're seeing, the complex environment, and just how fast things are changing, just make it really important to make sure that you have what you need to secure your use of AI and use it in a way where you can answer what you need both as a business and for your customers. So when we think about AI visibility, there are multiple layers to this, and we just wanted to quickly go through what that could look like. First is thinking about the infrastructure. Where does the AI service even live? This could be from a CNAV perspective, the services and platforms you're using, what the access looks like, and more. A lot of you may be existing with customers and have been using this with us for a longer time. Second is starting to think about more of the new things that we've started to build on the AI side. Be that be the model that's looking at the data or the tools you're building or agents or otherwise. Being able to look at that posture itself of that area with AISPM. Third is data. One of the things we talk a lot about both with customers and otherwise is AI and data really go hand in hand. And they're two sides of the same coin and being able to connect the dots between having a very strong data security posture and how that connects to your AI tools and services is critical. Last but not the least is the actual application that you're building that connects to different aspects of your cloud, the different components of data and AI to really put together the actual offering here. Now when we think about this as a puzzle and we put it all together, we really believe that we're able to meet you on this journey and grow with you on this journey for building and using AI in a secure manner. We will be talking about how this shows up in your environment and the steps you can take, but the combination of the visibility, context, and being able to really see AI as just an extension of everything you've been doing with cloud, but making sure we're tailoring it to what you need is critical. In terms of the steps we take, first, gaining visibility. Being able to look at your environment and asking what is even available. Where is the AI used? This surfaces as an AI bomb and inventory and a true service catalog to connect the dots. Then you start thinking about AI misconfigurations. These are things unique to the security of the different components that you're using. That could be your models itself, guardrails in use, MCP endpoints, and more. Then you start to think about reducing that posture risk. This is where you want to start to understand what are the toxic combinations and concerns that are being created, not just within AI, but everything it touches in your environment using the range of cloud context available to you. Last but not the least, as much as we look at proactive security and secure by design, it's also critical to be able to monitor AI and run time. Being able to detect what may have changed, any anomalies, and more is super important. We are going to see several different updates here across the build side. So from code to cloud to run time, and you'll see this as we go into our demo as well. So step one, you're going to see how you're able to surface things with our AI bomb, the shadow AI, and more having end to end pipeline visibility. Step two, being able to surface those AI misconfigurations. Be that be with SageMaker, Bedrock, and more and asking yourself what has applied, what feels missing, and can you apply that all the way from build to run time. Then being able to connect this risk to an AI pipeline across the board and be able to look at things like, oh, not only is there an exposed endpoint, it has access to a model that has access to training data that may be exposed. Being able to understand why that's a priority and how to respond to it. Last but not the least, being able to detect this real time with threat detection as well as anomalies and behavior and respond as you need to these threats. These all come together and really being able to see this from a visibility, risk, and response, and governance standpoint, and the Wiz platform really unlocks that for you with our AISPM offering. With that, I'm going to hand it to Aviel to help us walk through a demo of this platform. Hey, everybody. Nice to meet you. My name is Aviel. I'm the product manager for AISPM. Second, let me get my screen up here for you. And we're gonna walk through the demo and see those four stages which we've just went through on the deck. So for those of you who are a little bit less familiar with the Wiz platform, I actually just wanna start off down here and highlight our lens. Right? So when you wanna look at AI security, we really recommend that you select this AI security lens. What it's gonna do for you is automatically surface everything relevant, for AI security for you in this dashboard. So let's go through those four steps one by one, and I wanna show you how what we have up here on the screen is gonna provide value for each one of those. So starting with visibility. Let's take a look at what we actually have in this dashboard. So up top here, we have different, frameworks and issues, which are toxic combinations. Actually, I wanna put those on on the side for just one second when we get to them in stage three. And keep on scrolling down our dashboard to look at all of the different data that we are providing at your fingertips from a visibility perspective. So starting up on the top left over here, what we have are widgets which provide, great visibility for all of your AI pass and your self hosted, models. So what is that? This here is going to this widget here is going to surface all of your models, agents, guardrails, knowledge bases, which you have in the different platforms, that the main platforms from the cloud service providers such as AWS Bedrock. I'm gonna scroll down a little bit so you can see the icons. But, Azure, ML, Azure OpenAI, Vertex, etcetera. So all of those are gonna be right here at your fingertips, for you to go over, and you can mark them as approved, unwanted, or that they need more. They need to, the security needs to dig into them a little bit further. Down here, we have all of these self hosted models. Right? So in contrast to the models up here, which are being run within those, CSP platforms, these are models which are being run directly in your environment, which we, provide visibility for you as well. Moving over here to moving over on the right. Over here, we have all the different AI software, which we're finding in your environment. If I just scroll down here, I'm sure you're gonna see a lot of, familiar names and logos. So for example, MCP, model complex protocol, etcetera etcetera. Once again, you can mark all the different softwares, as approved or unwanted, unreviewed so that the security team can keep tabs on them. As I scroll down here, we have this widget as well, for exposed, secrets, which is obviously a problem within itself. But from a visibility standpoint, what we found is many times, we can surface shadow AI actually by detecting these secrets in your environment for AI services which you might not have known or been aware of their actual use. So from a visibility visibility standpoint, this can provide great value as well. Right? So I think that's an, quite a bit in terms of, examples from this dashboard. Mind you, I just wanna briefly go over here. We have our AI inventory. So for each one of these topics, you can really dig into it, looking at all the AI technologies, the deployment. So deployment is, all of the models that have actually been deployed and are accessible. All of your agents, right, your models, pipelines, data, etcetera. There's a lot here. I'm not gonna be able to get into it within the time we have today. But there's really lots of different ways in which you can slice and dice this in order to get full visibility. And finally, just in terms of visibility, what I'll mention is that this is all correlated to the projects. Right? So, for example, if you wanted to see what models are running in your production environment, you could select the, relative projects and then it would all be correlated to that, specific project or environment. Let's move on to misconfiguration, step number two. Right? So as I scroll down here, we provide this widget here, which surfaces all of the different resources, that have critical or high AI misconfigurations. Right? So let's just take one so we can kinda color that in a little bit. What we have here is an AWS bedrock guardrail. We immediately, here see all of the filters for the input and output of that guardrail. We can also see here that it has one high configuration finding. Let's go see what that is. So we see there, there's a couple of them over here. So here, this one is it's not protected for prompt injection. Obviously, that, is, a high severity, or missing PII protection. We we could dig into each one of these and see exactly what's missing. But the point being that we ingest these guardrails, from the platforms in the, CSPs, and then we scan them, for any, findings with or misconfigurations, which you need to be aware of according to severity. Right? So that, I think, is a good example of misconfigurations. Keeping an eye on the time, I think, I'd like to move, into our issues, which are essentially the toxic combinations. We mentioned before in the deck that what we're able to do is utilize the full context of our security graph. And by connecting all of the various dots together for you, we can actually highlight, attack paths, which need your immediate attention or potential attack paths. So let's just take a couple of different examples to highlight that, different types of toxic combinations, which we can surface for you. So just starting, with the top one over here, what we have is a publicly exposed VM with initial access vulnerabilities, which has access to a bucket used for AWS model training. So it's very similar to the one we saw on the deck. But let's just review this because it's quite a bit here. So here we have an e c two instance, right, which, via this application endpoint is externally exposed to the Internet. It has some sort of vulnerability on it. Now if an attacker were to gain access to this resource here, it would be able to assume this role, which has access to this s three bucket which is used, for, AI model for its l m training. Right? So obviously, that is a a high criticality in terms of that toxic combination. Let's go see, another one. However, let's do this one with MCP. Right? So just let it load. So here we have an application endpoint exposing an MCP with high, critical or high vulnerabilities, on a VM with access to sensitive data. Right? So let's go see what that looks like on the security graph. So here we have this, MCP server, which is, running on this resource over here. Both the MCP server and the resource are exposed to the Internet via these three application endpoints. This MCP has two vulnerabilities. Right? So if an attacker buy the MCP, where it'd be able to gain access, to this resource, they would be able to assume this admin role. Right? So that's another type of toxic combination, which we surface for you. And let's go do one more. Let's take something, a little bit more simple, but I actually think that not any less, interesting. I know that's not the one I wanted to do. Hold on. Here. Let's do this one. Right? So the other ones we, were a little bit more complex, but I think this one is also, super interesting and relevant to a lot of our customers. So what we have here is, I I skipped the the title, but fine tune AI model trained on dataset with sensitive data. Right? So what what it looks like on a security graph is that this model down here went through this fine tuning job using this file which had PII and emails in it, ultimately producing this, fine tune model. What that means is that downstream, anything that this model does, we need to regard as having access to sensitive data just given the fact that it was actually trained on the data to begin with. Right? So that that's gonna be another, combination, which we which we surface for you. One last thing, not an issues in terms of, our capabilities of servicing these toxic combinations, which I wanna, highlight is our a our proprietary AI model scanner. So, anytime our workload scanner detects, hosted AI models within your environment, what it immediately does is triggers our our proprietary AI model scanner to scan them for any vulnerabilities. Let's go see what that looks like here. So here we're working at an AWS e c two instance here, which was defined as an AI workload. Right? And the reason it was defined as an AI workload is because it runs these two AI models. So when the workload scanner detects these two models, what, again, it's gonna immediately do is trigger our model scanner. Now in this case, obviously, for demo sake, we have, we have it finding this vulnerability. In this case, it's a CW five zero two, which is essentially a malicious pickle. So when, this will be, deserialized, this, malicious coin miner, XM regular known coin miner would actually, would actually run. Right? So this is, highlighting the capability of our proprietary model scanner to scan these posted models within your environment and surface any vulnerabilities. But this is actually, I think, a great, point to transition to, our runtime capabilities as well. And now, correlating back to those four stages with its toxic combinations and general posture was stage number three moving out to stage number four in runtime. Let's let's imagine a scenario that this finding, couldn't have been surfaced, with our AI model scanner just because, this is a zero day, vulnerability and there was no way that the scanner could actually, detect it ahead of time. We could still detect it in this case, with the Wiz sensor. So what this would look like is that in real time, the sensor would detect this DNS query to a known crypto mining domain. Right? So again, here we're seeing this XM rig, pop up here to the DNS request. And then we will be able to detect the threat in real time, and take it down immediately. Right? So that's just one example of how we can provide, runtime security, for, your AI workloads as well. Finally, just to briefly touch upon the different frameworks we have up here. I don't have time to dig into all of them, obviously, just to select one, which is, pretty popular, the OAuth top 10 LLM. So for each one of these rules, you can open them up, look at the different controls, and examine how you're, doing in terms of your security posture. This is another really helpful way to, wrap your head around how, how you're doing in terms of AI security within your environment. We're coming up on time. And I think with that, I'm going to end this this, brief demo here. And, we can absolutely open up, with some questions. Awesome. Well, thank you, Snegha and Aviel, for the super quick walk through. As Aviel noted, we're gonna move into the q and a portion. So if you have questions, put them in the chat or the q and a in the top right module. I am also in a few in a minute or two, going to launch a quick short survey for our audience. It's just five questions. And as a thank you, we'll send you a $25 Starbucks or depending on the country you're in, coffee related gift card. But first off, let's get into it. The the first question that we've got a couple times is, is AI SPM a separately licensed model? Say, how will you take that one? Yes. So I think I sent a note on this one, but AI security is part of cloud security for us. So our cloud SKUs come with this functionality. So just like our other billable workloads, just based on how much we're scanning in your environment and you're able to transparently see that in the environment, you will just be able to use it. So if you're a cloud customer, you already have AI security. Awesome. What tool or network protocol is required to take inventory of AI models in a corporate network? So without getting into the nitty gritty here, in general, anything that are, anything that the workload scanner is able to scan within your environment, will automatically trigger our AI model scanner. In terms of the AI models, which our model scanner, can detect and then scan for vulnerabilities, that's all in the documentation, which you can review there. We can also surface that link for you, after the call. Thank you. And how are you providing the real time protection or prompt protection? So that's a great question. Our strategy regard regarding, inline prompt injection is to be, I would you could kinda say the the hub in connecting in various folks. So as of today, we're not planning on rolling out our own proprietary, AI firewall to detect that, in line. We are doing monitoring out of band, which I'm not gonna get into now, just, to keep things short. We're already working on ingesting those types of alerts via the wind platform, and to be able to partner with, leading vendors in the industry so that you can, ingest them into the Wiz portal and have it all under a single pane of glass. Awesome. And we will be sending out a recording of the demo and the whole session. You should have it in your inbox. I'm gonna say in by tomorrow morning, my time, and I'm in Seattle. So do the math depending on where you are in the world, but you will definitely get the recording and the slides will be in there as well. One more question we've got. Can Wiz Sensor play a role in real time guardrails for the prompts that are coming in and out from LLM? Bonus points for MCP traffic. Currently, no. We are looking into it, and I think we can share a little bit more about our road map offline exactly regarding how the sensor is gonna be, playing a role here. Awesome. Well, I just launched the survey. Thank you very much, Aviel and Snegha, and thank you for all of you all for joining us today for this quick demo. Again, you will get this recording plus the slides in your inbox in, let's say, the next twenty four hours so we don't have to do the math. And there is a survey for you, five short easy questions, $25 coffee gift card in it for you as well. And again, you can still meet a wizard. If you have any other questions, you can always follow-up with us via email. But thank you all for your time today, and have a wonderful rest of your day. Thanks, everyone.