Latest / Microsoft Cloud IT Pro Podcast / Episode 411 – Securing AI with Microsoft Purview: Data Security Posture Management for AI
Transcript
- 0:03Welcome to episode 411
- 0:05of the Microsoft Cloud IT Pro podcast recorded
- 0:08live on 09/24/2025.
- 0:12This is a show about Microsoft three sixty
- 0:14five and Azure from the perspective of IT
- 0:16pros and end users, where we discuss a
- 0:18topic or recent news and how it relates
- 0:21to you. In this episode, we circle back
- 0:23to our conversation around preparing for Microsoft three
- 0:26sixty five Copilot
- 0:28and, frankly, all AI in a conversation about
- 0:31data security posture management. We discuss what's available
- 0:34to you within DSPM for AI and Microsoft
- 0:38Purview, what you need to do in order
- 0:40to get started with it, and how you
- 0:42can leverage this tool within your company to
- 0:44keep your sensitive data in Microsoft three sixty
- 0:47five safe and secure from AI.
- 0:50Let's dive in.
- 0:54Scott, it feels like forever since we've talked.
- 0:57I think this might be the longest we've
- 0:59gone in a while without talking.
- 1:02Well, that's because you went to a conference,
- 1:04and
- 1:05you let Jay and Joy take over the
- 1:07podcast Saturday. That was a mistake. That's fine.
- 1:10Nah. It was all good. Oh. Don't you
- 1:12go back and listen to that one if
- 1:14they haven't? Joy and Jay take over and
- 1:16talk change management, episode
- 1:19four ten Four ten. Of the Microsoft Cloud
- 1:21IT Pro podcast. Yes. That was a good
- 1:23one. We need to get you back to
- 1:24a conference,
- 1:26Scott. I was at a conference last week.
- 1:28I was just at a different conference. Which
- 1:30conference were you at? I was at the
- 1:32storage developer conference
- 1:34out in Santa Clara, California. Oh, how are
- 1:38storage developers these days?
- 1:40They keep on developing things. I had a
- 1:43talk on retrieval augmented generation and object storage.
- 1:46So where does object storage fit into
- 1:50the
- 1:51kinda AI pipeline when you think about an
- 1:54end to end spectrum
- 1:55of AI training all the way through to
- 1:59AI inference
- 2:00and actively, like, serving
- 2:02these workloads and LLMs and things like that
- 2:05are on it. But my talk was a
- 2:07little bit kind of about, like, the three
- 2:09quarters of the way through the AI pipeline
- 2:11with fine tuning and retrieval augmented generation. Got
- 2:15it. Interesting. I was gonna say, because you
- 2:16probably need storage, like, at every step of
- 2:18the pipeline. I mean, AI kinda depends on
- 2:21data a little bit, which kinda depends on
- 2:23storage.
- 2:26Yeah. If you're training a foundational model, you
- 2:28require petabytes and petabytes or exabytes of storage.
- 2:32If you're just doing fine tuning,
- 2:34you might be in the petabyte range, or
- 2:36you could be smaller. You could be down
- 2:37in the terabyte range and taking a very
- 2:40specific dataset to
- 2:43teach a model the new set of skills
- 2:46along the way. But it turns out that
- 2:48object storage platforms
- 2:50are super important
- 2:52when it comes to retrieval augmented generation as
- 2:55well, because you've got this thing where you
- 2:57have to effectively
- 2:59kinda rather than teaching new skills like you
- 3:01would fine tuning, you have to teach new
- 3:03facts. Like, hey. Here's something that is going
- 3:05to augment your knowledge. So if you think
- 3:08about, like, Copilot notebook, for example. So we've
- 3:10talked about those in the past. So you
- 3:11could go into a Copilot notebook, and you
- 3:13can put whatever it is, ten, twenty documents
- 3:15in there. When you put those documents into
- 3:18the notebook,
- 3:19Copilot
- 3:20and the underlying
- 3:22LOM that's serving the response back to you,
- 3:24they're not actually reading that document, like, as
- 3:26a document. Like, they don't open up a
- 3:28Word document or a PDF and read it
- 3:30page by page. They have to go ahead
- 3:33and
- 3:34generate embeddings. So effectively, they take that PDF
- 3:37for that Word document, and they turn it
- 3:39into a bunch of chunks, so and or
- 3:41embeddings. Those embeddings go to a vector database
- 3:44and
- 3:45vectorize,
- 3:46which means then you need a
- 3:48a vector search mechanism that can augment the
- 3:51LLM response. But object storage is kind of
- 3:54important, or if you think about, like, putting
- 3:55a new document into a Copilot notebook, something
- 3:57like that Yep. Because it's gotta be chunked,
- 4:01and it has to be generated, and embeddings
- 4:03need to be generated, and those embeddings need
- 4:05to be pushed over to a vector database.
- 4:07Well, that means anytime I upload a new
- 4:09document, if I put a new document or
- 4:11I take a document out of my Copilot
- 4:13notebook, anything like that, you actually need to
- 4:15kinda kick off back end process for data
- 4:18freshness. So having things like at least in
- 4:21our platform on the Azure storage side or
- 4:23the object storage side, we have things like
- 4:25change feed, which can
- 4:27be leveraged by these systems to kinda give
- 4:29them a proactive notification and say, hey. There's
- 4:31a new object here. Go and run whatever
- 4:34you need to run on top of this.
- 4:35For example, maybe you integrate Azure AI search
- 4:38into your workflow, so you pump a bunch
- 4:41of documents into a storage account. AI search
- 4:43can actually respond to the change feed notifications
- 4:47that are coming out, and it can automatically
- 4:49do the
- 4:51chunking, embedding, and store those embeddings in the
- 4:54vector database
- 4:56just based on the virtue of having the
- 4:58ability to notify in, like, a durable ordered
- 5:00list with something like a change feed along
- 5:03the way. It's kinda super interesting because I
- 5:05don't think folks really think about the back
- 5:08end of these things too much. Right? If
- 5:10you're playing with a Copilot notebook and you're
- 5:12putting documents in there, it's all just kinda
- 5:14like magic
- 5:15that all of a sudden it can reason
- 5:17over my document in really near real time.
- 5:21It doesn't it doesn't take too long to
- 5:24generate those embeddings, especially with smaller documents and
- 5:27smaller datasets, which is what most people are
- 5:30using in things like Copilot notebooks today. But
- 5:33there's a lot of moving pieces on the
- 5:35back end to get that all going and
- 5:37get it to where it needs to be.
- 5:39Yeah. So anyway, I had a session on
- 5:41that. That'll be up on YouTube. I'll I
- 5:43can share it in a couple of months
- 5:45once once those come out for public consumption.
- 5:48And then I also spent a couple days
- 5:50at a a plug fest.
- 5:52So SNIA
- 5:54and the broader
- 5:55storage developer conference, they organize effectively these hackathons
- 5:59that bring a bunch of vendors together under
- 6:01NDAs so that we can go and do
- 6:04interop testing with each other and
- 6:07kinda do debugging in real time.
- 6:10So there's a bunch of those plug fests
- 6:12that's needed hosts. They host one for Swordfish.
- 6:14They host one
- 6:15for SMB for, like, vendors to test their
- 6:19SMB implementations
- 6:21that might live outside of Windows or things
- 6:23like that. And then the one that I
- 6:24did was an object storage plug fest. So
- 6:26we have focused on, like, restful API surfaces
- 6:29for object storage vendors of which Microsoft is
- 6:32one
- 6:33with Azure storage, and then you've got s
- 6:36three protocols, which are kind of the elephant
- 6:39to the room as s three is not
- 6:41only an s three, but, you know, Google
- 6:43supports s three as a protocol. And they're
- 6:45a cloud object storage platform, Alibaba,
- 6:48Backblaze, and then a whole ton of on
- 6:50prem vendors, IBM Cloud Object Storage,
- 6:53Oracle,
- 6:54both their on prem offering,
- 6:56OCI as well, and their cloud infrastructure offers
- 6:59an s three protocol. So spent a couple
- 7:01days kinda hacking away at that and
- 7:04then seeing if we can get some of
- 7:05our products
- 7:06that we have that talk to s three.
- 7:08Like, while Azure Storage does not have a
- 7:10native s three service, we definitely have products
- 7:13that talk s three, especially
- 7:15for data management, data movement scenarios.
- 7:17So things like like AZCopy,
- 7:20StorageMover,
- 7:21you pick up one of our SDKs,
- 7:23we absolutely allow you, and we've got some
- 7:26great enabling APIs to help you move your
- 7:28data from S3 over to us. Got it.
- 7:30Very cool. Well, hopefully we'll get you at
- 7:32a, at another conference yet this year, Scott,
- 7:35because we have Ignite coming up
- 7:37in November, November.
- 7:40So hopefully doing podcast stuff there again.
- 7:44I actually have a couple I have one
- 7:45before that too. I'm going down to Orlando
- 7:47too. It was dev intersections,
- 7:49and now they're also doing, like, co locating
- 7:52dev intersections with cybersecurity intersection
- 7:55and the Next Gen AI conference. So I
- 7:58have a couple talks,
- 7:59presentations Oh, nice. At cybersecurity intersections. And then
- 8:03I'm actually doing a couple more, two in
- 8:04December in Dallas at Workplace Ninjas.
- 8:08So Cool. I've got two more coming up.
- 8:10Well, three more coming up, hopefully Ignite, and
- 8:12hopefully we can get you out to Ignite
- 8:14too. Yeah. We'll see. Pump it out on
- 8:16the socials. You can always find Ben on
- 8:18LinkedIn. You can find the podcast on LinkedIn.
- 8:20Can you find me on LinkedIn as well?
- 8:21Yeah. I think it's becoming one of our
- 8:23preferred communication method
- 8:25mechanisms for the podcast here. I would say
- 8:27LinkedIn is absolutely becoming
- 8:30the preferred one.
- 8:31All of the other ones. This is not
- 8:33just a any social media. Like, all of
- 8:35the other ones I have had to turn
- 8:37into looking at them all as just pure
- 8:39entertainment.
- 8:41There is nothing of the state of the
- 8:42world. Yeah. Not much of value that I
- 8:44have found on many of the other social
- 8:46platforms. So LinkedIn is absolutely becoming
- 8:49that preferred social media for
- 8:52the podcast,
- 8:53although we are still on all the other
- 8:54ones. But speaking of AI and storage and
- 8:58data, we should probably dive into
- 9:01this is like the second part of episode
- 9:04four zero nine
- 9:06when we started talking about preparing for Copilot
- 9:09in Microsoft March,
- 9:11getting ready for it. And we didn't quite
- 9:14get to kind of the last
- 9:16point that I covered in the presentation I
- 9:19gave
- 9:20wherever I was for that one, Branson, Missouri
- 9:23around
- 9:24kind of diving into one of the, I
- 9:26would say, very quickly growing product in purview
- 9:29or a
- 9:31solution technically in Purview or doing a lot
- 9:34of the management, controlling,
- 9:36monitoring,
- 9:37copilot,
- 9:38and really other AI usages is the data
- 9:42security
- 9:44posture management or d
- 9:47DSPM for AI
- 9:49that exists in Purview. We should kind of
- 9:52talk through this one. It was a little
- 9:54bit of an interesting new area for me.
- 9:56Yeah. So if you haven't looked at it,
- 9:59I definitely recommend if you're doing anything with
- 10:01Copilot or really anything at all and you
- 10:04care about your dating, keeping it safe from
- 10:06AI, because there's aspects of this that don't
- 10:08apply to just Copilot. You should go check
- 10:10this out. So it's in Purview. And once
- 10:13you go into
- 10:15Purview,
- 10:16you'll just start on, like, a general
- 10:18overview,
- 10:19typical to a lot of the different admin
- 10:21centers that gives you some things to get
- 10:24started with. So there's kind of four
- 10:27big aspects
- 10:28to this whole data security posture management for
- 10:31AI that you need to enable that are
- 10:34required to really get going in terms of
- 10:37monitoring
- 10:38your data, because this is all about discovering
- 10:41what are people doing, securing AI activity.
- 10:44So in order to
- 10:46secure it, discover it, know what people are
- 10:48doing, you you kinda have to be able
- 10:50to look at the data. So there's Turns
- 10:51out. Yeah. So there's a whole just first,
- 10:54just activate Microsoft Purview audit. This is really
- 10:58just that audit log that we've had for
- 11:00years. That's both security center compliance center. It
- 11:03used to be the unified auto log. Now
- 11:05it's just audit, but then there's also in,
- 11:08I don't even remember when this came out.
- 11:10I don't remember ever talking about it, seeing
- 11:12it in the news. There's a Purview
- 11:14browser extension
- 11:15that you can deploy now, and it really
- 11:18doesn't
- 11:20I would say it doesn't do much. There's
- 11:22nothing really for end users to interact with.
- 11:24It's one of those that you can install
- 11:25and just kinda keep it hidden in the
- 11:27background.
- 11:28There's no reason to show it in your
- 11:29task bar, but it is really just watching
- 11:33browser activity. Install it there so it can
- 11:35see what are you doing within your edge
- 11:37browser. I think there's I think the extension
- 11:39also exists for Chrome. And if I remember
- 11:41right, the edge one may actually just be
- 11:44in the Chrome store. I have it for
- 11:46Firefox too. So edge Chrome, Firefox
- 11:49to really just help watch signals of what
- 11:52are people doing in the browser with data.
- 11:55Again, this kind of ties into, are people
- 11:57going to other AI sites, going to chat
- 11:59GPT, going to claw, going to some of
- 12:02those.
- 12:03So that's the next step. Just get that
- 12:05browser extension deployed.
- 12:07There's guides on, you can go do this
- 12:09with Intune,
- 12:11whatever solutions you kind of use to manage
- 12:14end user devices, leverage those to deploy Purview.
- 12:17And then after that, or along with that,
- 12:19is also onboarding devices to Purview. So this
- 12:23is one too that I've actually seen some
- 12:25people
- 12:25miss Scott is that
- 12:27just because you've onboarded the device into
- 12:31Intune
- 12:32or into other,
- 12:34enrolled it and enter ID joined it to
- 12:36enter ID. There's a whole nother aspect of
- 12:39actually onboarding devices
- 12:41into Purview as well. So you can have
- 12:43a device in Intune, have a device in
- 12:45Entra, and not have it
- 12:47onboarded into Purview. Yeah. I wonder
- 12:50can we take a step back? So Yeah.
- 12:52I'm I'm just kinda, like, watching your screen
- 12:54as it's up. So when you go look
- 12:56at Microsoft documentation and you say, hey,
- 13:00these two folks are talking about data security
- 13:02posture managements, and maybe you just go Google
- 13:04for that or Bing or DuckDuckGo or you
- 13:06search for TPSM,
- 13:08you're going to end up
- 13:11in one section
- 13:12of the Purview
- 13:14documentation.
- 13:16So you're gonna end up in a section
- 13:17around
- 13:18data security solutions.
- 13:21But as you've got the portal up, and
- 13:23if you kinda keep, like, Google being DuckDuckGo
- 13:26search engine of choice,
- 13:28you end up on
- 13:30a different page, which is DPSM
- 13:32for AI,
- 13:34and then starts talking about the specific considerations
- 13:38for that, onboarding to the additional set of
- 13:41capabilities,
- 13:42the co pilots,
- 13:44and things that come with that. So can
- 13:46you help me, like, disambiguate the two? DPSM,
- 13:49DPSM for AI? AI. So DSPM,
- 13:53this is one I haven't done a lot
- 13:55with yet. I actually click over to data.
- 13:57I don't know that I have
- 14:01full access into it. I do have some
- 14:02of it in mind. Where I would say
- 14:05DSPM
- 14:06is just really overall
- 14:08data security. Right? You're gonna go in and
- 14:10look at and if you go into the
- 14:12portal and look at it, it's risky user
- 14:14investigation.
- 14:15Are users doing performing risky behaviors with their
- 14:19data? Or
- 14:20where is that sensitive data? So this ties
- 14:24a little bit more into what I would
- 14:25say of information protection or even
- 14:29a little bit maybe it's not even information
- 14:31protection, it's a little bit higher. So you
- 14:34just have, like, what is my overall
- 14:37posture when it comes to keeping my data
- 14:39safe? It could include some of the information
- 14:41protection. It could include insider risk management.
- 14:45It could include, like, one of the other
- 14:47related solutions that shows up in Purview is
- 14:49data security investigations.
- 14:50It's broader, I would say, is how you
- 14:52look at data security posture management. The DSPM
- 14:55for AI really
- 14:57drills into
- 14:59then taking that data security posture management. How
- 15:02am I securing my data? Now let's look
- 15:05at it just through the lens of how
- 15:07am I securing it when it comes to
- 15:09interacting with AI. So that's things like your
- 15:12co pilots, maybe your agents that are out
- 15:15there, like your declarative agents, knowledge agents, and
- 15:18Yep. The co pilot portal or teams.
- 15:21It's basically like a superset of capabilities or
- 15:24a double click to go and monitor
- 15:27AI workloads specifically?
- 15:29Yeah. That's exactly it and guidance around it.
- 15:32Like, even if you go into the data
- 15:33security posture management portal in Purview, not the
- 15:37AI one, Really, it just talks about, like,
- 15:40there's some stuff here around risky user investigation,
- 15:43sensitive data, but it's really just an overview.
- 15:46You've got some reports around users performing
- 15:49risk related activities
- 15:51on unprotected sensitive assets. So I'm going into
- 15:54SharePoint files where
- 15:56I have
- 15:57credit card numbers, I have PII,
- 15:59I have
- 16:00IP addresses, I have connection strings,
- 16:03and there's people interacting with that sensitive data
- 16:06in some way. It doesn't necessarily have to
- 16:09be AI.
- 16:10It could be creating a sharing link to
- 16:12it. It could be emailing it to external
- 16:14users, some of that information protection type stuff.
- 16:17And it just kind of gives you an
- 16:19overview and some reports there, a few recommendations,
- 16:22but there's, I would say, DSPM, there's really
- 16:24not much there. It's more of a reporting
- 16:26and a little bit more monitoring. It's not
- 16:28it doesn't have many features in it, quite
- 16:30honestly. This is where, like when you mentioned
- 16:32the monitoring and some of those things, so
- 16:35this is the other areas, like, I think
- 16:37it's hard. If you don't find specifically the
- 16:39DPSM for AI section in the docs and
- 16:42you just end up on DSPM,
- 16:44you're gonna get, like, two different sets of
- 16:46guidance for onboarding Yep. Or, like, how far
- 16:49you
- 16:50should go and what you should turn on.
- 16:51So some of those things that you mentioned,
- 16:53like, if you're
- 16:54trying to specifically monitor
- 16:57AI usage within your organization,
- 17:00and maybe that includes things like third party
- 17:03gen AI sites, that could be OpenAI,
- 17:05Claude, Perplexity,
- 17:07things like that, Google, Gemini, things like that
- 17:09that are out there, then you do have
- 17:11to make sure that your devices are onboarded
- 17:14to Purview. So you gotta go through, like,
- 17:16that whole flow. You might need DLP
- 17:18in place. That browser extension becomes super important
- 17:22Yep. To put out there and make sure
- 17:25that's all good. The auditing piece, like, yeah,
- 17:28you gotta go turn all of that stuff
- 17:31on. And then there's always that pesky, like,
- 17:33little licensing consideration.
- 17:35So it looks like for some reason for,
- 17:37like, Copilot in fabric or the security Copilot,
- 17:41then you need the enterprise version of Microsoft
- 17:44Purview data governance, which then lights up a
- 17:46new set of APIs and a new set
- 17:49of capabilities
- 17:50behind it. So there's definitely, like, a lot
- 17:53of reading with this one, I think, just
- 17:54to figure out how it composes.
- 17:56And then what are you gonna go turn
- 17:58on? And then like you said, because it's
- 18:00primarily a reporting thing. And to a certain
- 18:02degree, like, remediation. Right? You're gonna use these
- 18:05reports to drive new behaviors in your organization,
- 18:07things like that. But you kinda have to
- 18:09decide
- 18:10what you wanna report on out there. Like,
- 18:11if you wanna report on usage of cloud
- 18:13in your organization, then you've gotta go down,
- 18:15like, this, like, wild path of kinda turning
- 18:17everything on. Yes. And you'll find and we
- 18:20can talk about some of this too. There
- 18:22is even some functionality in here that is
- 18:26pay as you go, where you end up
- 18:28having to go in and connect your Purview
- 18:30instance to an Azure subscription
- 18:33or some of the pay as you go
- 18:35functionality as well. I did see that. So
- 18:37there's this little note, Riley, here, like a
- 18:39little bullet that says, hey. For AI apps
- 18:42other than Copilot and Facilitator, you have to
- 18:45set up pay as you go billing for
- 18:47your organization.
- 18:48And they do guide you into
- 18:51kinda, like, where in the UI
- 18:53to turn it on and things. But it's
- 18:54a little weird because Purview has multiple billing
- 18:58models. It's got the per user billing model,
- 19:00and then you're blending the PAYGo model on
- 19:02top of it. It's not like one or
- 19:05the other.
- 19:06You've gotta get out there and kinda figure
- 19:08out what he uses
- 19:11what. Yeah.
- 19:13Yeah. And you might be there already. Like,
- 19:14that's the other thing I think that I
- 19:16noticed that I saw was
- 19:19remembering, like, my days of Purview. Like, we
- 19:21we already had customers who were doing things
- 19:23like monitoring AWS usage,
- 19:26And we were going out and deploying the
- 19:28proxies
- 19:29and the monitoring solutions and things like that.
- 19:31Those are all out there. Like, if so
- 19:33if you already have, say, like, AWS monitoring
- 19:35in place, you're probably a PayGo customer already,
- 19:37and you've probably already configured the PayGo PayGo
- 19:40billing. It was a little surprising to me
- 19:42to see, like, Azure services fall into the
- 19:45PayGo thing as well. I can understand why,
- 19:47like, Amazon or Dropbox, Google, things like that,
- 19:50but there's stuff like Microsoft Fabric that requires
- 19:53PayGo,
- 19:54monitoring
- 19:54Azure Data Lake Storage or Azure SQL
- 19:57data sources, that all requires PAYGo. It was
- 20:00kind of a a interesting list to look
- 20:03at. I'll make sure to put a look
- 20:05at the the billing docs and the billing
- 20:07considerations
- 20:08in the show notes. Yeah.
- 20:13Do you feel overwhelmed by trying to manage
- 20:15your Office three sixty five environment? Are you
- 20:18facing unexpected issues that disrupt your company's productivity?
- 20:21Intelligink is here to help. Much like you
- 20:23take your car to the mechanic that has
- 20:25specialized knowledge on how to best keep your
- 20:27car running, Intelligink helps you with your Microsoft
- 20:30cloud environment because that's their expertise.
- 20:33Intelligent keeps up with the latest updates in
- 20:35the Microsoft cloud to help keep your business
- 20:37running smoothly and ahead of the curve. Whether
- 20:40you are a small organization with just a
- 20:42few users up to an organization of several
- 20:44thousand employees,
- 20:46They want to partner with you to implement
- 20:47and administer your Microsoft cloud technology.
- 20:51Visit them at inteligync.com/podcast.
- 20:55That's intelligink.com/podcast
- 21:02for more information or to schedule a thirty
- 21:04minute call to get started with them today.
- 21:06Remember, Intelligink focuses on the Microsoft cloud so
- 21:10you can focus on your business.
- 21:14The one thing I would say when I
- 21:16started doing this, they do make it fairly
- 21:18straightforward
- 21:19to get going. So what I did, like,
- 21:21I went through that overview,
- 21:23turned on those those four required things, the
- 21:25auditing, the onboarding devices,
- 21:28extending insights
- 21:30was one of them, and then the browser
- 21:33extension. But even some of these, like when
- 21:35you click on, okay, we need to discover
- 21:37sensitive information,
- 21:39a lot of
- 21:40what DSPM is
- 21:42policy driven off of other areas
- 21:46within Purview. So when you click to start
- 21:48extending your insights for data discovery, what it
- 21:52gives you is
- 21:54what the three policies are. And I already
- 21:56have them turned on in mind, but if
- 21:58they're not turned on, it's like this is
- 22:00the PayGo policy or this is
- 22:02free depending on those interactions
- 22:05in just a one click button to deploy
- 22:07it. And then once you deploy it, you
- 22:09can always go in and look at policy
- 22:11details.
- 22:12So
- 22:12for extending that data discovery, like, again, the
- 22:15first part of this is what are people
- 22:17doing is like, are people putting sensitive information
- 22:21in AI prompts in Edge or in your
- 22:23browser? And
- 22:25this policy
- 22:26is actually a data loss prevention policy
- 22:30that sits in the collection policies
- 22:33in DLP. So when you click it, it's
- 22:36just going out and creating this. And when
- 22:38you would go into data loss prevention and
- 22:40your collection policies,
- 22:42you'll see all those DSPM policies that you've
- 22:45created, and you can go in and look
- 22:47at what those policies are doing. You can
- 22:49tweak them, turn them on or off.
- 22:52So while it is like a one click,
- 22:54just deploy it, I would also say go
- 22:57in and look at some of these policies
- 22:59because it might also
- 23:01highlight some additional functionality that's available
- 23:04within these different platforms that is I would
- 23:06say they've almost kind of snuck in there.
- 23:08Like these collections I mean, you're paying for
- 23:10it. Right. It's it's out there. And these
- 23:12collection policies, some of the newer ones that
- 23:15are out here now are like network and
- 23:16browser based. So this one was sensitive information
- 23:19shared in the browser, but now it's also
- 23:22detecting sensitive information
- 23:24shared via AI over
- 23:26the network.
- 23:27So not only just what am I doing
- 23:29in the browser, but watching my
- 23:31browsers, my applications,
- 23:33API calls, add ins,
- 23:36watching my different network traffic
- 23:39from my devices
- 23:40for sensitive information going to other
- 23:44AI sites. You brought up licensing. It does
- 23:46require the
- 23:48A little bit. Global secure access, the secure
- 23:51edge
- 23:52licensing as well. So you will start running
- 23:54into,
- 23:55Hey, now I want to start capturing just
- 23:58network data. Guess what? You need some solution
- 24:00to actually capture that network data. So you
- 24:03have some policies there just around data collection,
- 24:06but then this is where you also have
- 24:08a bunch of policies if you go into
- 24:10recommendations
- 24:11after you start doing data collection,
- 24:13and you can go in and start creating
- 24:16different policies
- 24:17around
- 24:18then detecting that risk information.
- 24:21If you're using ChatGPT
- 24:23Enterprise,
- 24:24there's built in policies here for
- 24:27discovering and governing interactions with ChatGPT
- 24:30Enterprise,
- 24:31using Copilot and agents to improve your data
- 24:34security,
- 24:34and
- 24:36a lot of different
- 24:38functionality
- 24:39around creating these different policies to prevent,
- 24:42monitor, alert
- 24:44on where's my sensitive information going when it
- 24:47relates, again, specifically to AI in this particular
- 24:51solution. Yeah. Yeah. We should probably spend the
- 24:54last couple of minutes here and kinda talk
- 24:55about
- 24:56once you've got the insights, what do you
- 24:58do with them? Yeah. So say you're out
- 25:00there. You've got,
- 25:02like you said, your DLP in place.
- 25:05You've got this solution that's doing this monitoring
- 25:08for you. And then what are you gonna
- 25:10do with the information?
- 25:12And how do you leverage those rest of
- 25:14the tools for remediation?
- 25:16You know, I I think about, like, that
- 25:18proxy that gets deployed, like you said, like,
- 25:20the global secure access thing. Well, you can
- 25:22block stuff through that. Yep. You wanna take
- 25:25this report and maybe go turn those kinds
- 25:27of things on. And I think the biggest
- 25:29one here when I start talking to clients
- 25:32and working on it is once you have
- 25:33the detection to your point, what are you
- 25:34gonna do about it, is going in and
- 25:36building out those data loss prevention policies.
- 25:39One of them that
- 25:41kind of comes out of this, you don't
- 25:43need the SPM to do it, but it
- 25:45walks you through getting everything in place for
- 25:47it, is just protecting sensitive information
- 25:50in your own environment from Copilot. And this
- 25:53is an example I give to clients is
- 25:56up until
- 25:57deploying this with DLP,
- 25:59the only option to, like, prevent
- 26:01data from going into
- 26:03Copilot was to disable the index,
- 26:07which doesn't impact just Copilot, but also index
- 26:10impacts your search index, or it's to remove
- 26:12permissions to it. There may be things, Scott,
- 26:14like,
- 26:15I need to get to all of the
- 26:16data in my tenant. I need to work
- 26:18with that. I need to work with, I
- 26:19have w nines out there. If I have
- 26:21bank statements out there, like,
- 26:24this is my data. There is going to
- 26:26be sensitive information out there that I'm gonna
- 26:27need access to, and that I may wanna
- 26:29search, but I may not want Copilot
- 26:32necessarily processing that returning it. So, by going
- 26:35in and starting to apply sensitivity labels to
- 26:37it and then using this protect sensitive information
- 26:40from Copilot,
- 26:41I still have
- 26:54Yeah. I Yeah. I think over time, that
- 26:56granularity is gonna continue to change, like because
- 26:59you're gonna have subsets of people who might
- 27:01need to work on things, and they might
- 27:02need Gen AI and Copilot for them. And
- 27:04then you might have people who don't.
- 27:07So you're back to kind of just the
- 27:09regular,
- 27:10hey, let me make sure this stuff is
- 27:11secure. But it's also funny, I think, how
- 27:15things do
- 27:17leak in Copilot
- 27:19just because organizations
- 27:21that largely haven't kept up, I think, with
- 27:24having a big corpus of data that lives
- 27:25in SharePoint. And they never really figured out
- 27:28the whole security trimmed search thing, let alone
- 27:31the security trimmed AI search thing and how
- 27:33that comes together. Yeah. And then I would
- 27:36say some of the other ones is like
- 27:38going back to the other AI sites.
- 27:41Once you're starting to capture it and depending
- 27:43on your licenses and what you have turned
- 27:45on, it's not just, yeah, I don't want
- 27:47Copilot to process this, but you know what?
- 27:49I can't,
- 27:51I can't, don't want, however you wanna phrase
- 27:53it, my users going in
- 27:56and using sensitive information
- 27:58in other sites. So some of these DLP
- 28:01policies then that you can apply to devices
- 28:03is auditing that sensitive data that is
- 28:07shared to
- 28:09other sites. And I'm pulling up what these
- 28:11are, like uploaded over the network or one
- 28:15of these policies. It's not this one that
- 28:17I just pulled up is
- 28:19looking specifically for people adding it to
- 28:22BARD or to ChatGPT.
- 28:24Because now that you are auditing it, you
- 28:26can see people are pasting
- 28:28PII or pasting social security numbers
- 28:31into these other third party Gen AI sites.
- 28:34And frankly, we just can't have that and
- 28:37need to prevent that from even happening or
- 28:41prevent it from even occurring. Yeah. I think
- 28:42it's one of those rabbit holes. Yeah. You
- 28:44start going down. I mean, DLP policies in
- 28:46general
- 28:47are their own rabbit hole. I do like
- 28:49having kind of the enabler of the policies
- 28:52are canned and kind of pre baked, and
- 28:54then you can tweak them from there. But
- 28:56there's a whole bunch, I think, to go
- 28:58figure out. And then products like these are
- 29:01so easy to just get caught in the
- 29:03the charts and the reports. But I think
- 29:06figuring out how you
- 29:08actually
- 29:09get it wired up in your environment
- 29:11and get it to where it needs to
- 29:12be, that's another story and another thing that
- 29:15needs to be taken care of. Yeah. Then
- 29:17the other thing I would say is just
- 29:18the activities
- 29:19explore in the last few minutes. If you,
- 29:21you can see, Scott, even some of the
- 29:23stuff that I've done is
- 29:25AI interactions
- 29:26start showing up now in your activity explorer
- 29:28alongside
- 29:30emails,
- 29:31alongside
- 29:32working with files in SharePoint, and it allows
- 29:35you to start even monitoring and alerting on
- 29:38what are users doing. So it will you
- 29:40wanna get super big brother, Scott.
- 29:43Like, your Microsoft three sixty five tenants, your
- 29:44employers, they can spy on you all they
- 29:46want to, and the ability is here. It
- 29:49starts pulling in. I can go in and
- 29:50look at any prompt that I've done in
- 29:52Copilot, what that response was,
- 29:55any additional responses to it, going and looking
- 29:58through, again, just that audit log, having that
- 30:01audit history of AI interaction, which, I mean,
- 30:04I think this does become more and more
- 30:06important as people are using AI for their
- 30:08jobs, for generating
- 30:10documents, for generating
- 30:12emails.
- 30:13Let's face it, we all do it. There
- 30:15is going to start being instances
- 30:18where there may be cases that arise that
- 30:20you need to go figure out
- 30:22where this certain employee misbehaved
- 30:25with
- 30:26AI. And having that audit log, that Activity
- 30:28Explorer,
- 30:29well, yes, it does feel big brotherish by
- 30:32an employer. From a legal perspective, compliance perspective,
- 30:36I think this is absolutely
- 30:38necessary as well. So, yeah, all of that
- 30:40gets built in there. And then there's some
- 30:42data risk assessments. If you want to go
- 30:44run some assessments
- 30:45around
- 30:46the amount of sensitive data,
- 30:48links
- 30:49that have gotten shared,
- 30:51how many links have been created
- 30:53in the
- 30:54last
- 30:55100 or the last 30 or so. There's
- 30:58also some of that. It's similar to what's
- 31:00like in the SharePoint,
- 31:01the data access governance reports in SharePoint. Some
- 31:04of these data risk assessments around
- 31:07what types of links and sensitive data have
- 31:10been shared
- 31:11across my environment so that I can go
- 31:14in and
- 31:15respond,
- 31:17maybe have user training because
- 31:19it's highlighting some bad behaviors of employees.
- 31:22But being able to run some of that
- 31:23reporting and assessment around where are people sharing
- 31:26data and how much sensitive data is out
- 31:28there? Where are they sharing data more than
- 31:30you think? How much is out there also
- 31:32more than you think? A 100%.
- 31:35The other interesting thing, Aaron, I wonder if
- 31:36you've run into this just, like, with any
- 31:38of your customers,
- 31:40is
- 31:41some of the Gen AI platforms don't actually
- 31:43run their own things. So, like, if you're
- 31:46a perplexity customer coming through a perplexity API,
- 31:49then you're actually using an underlying model from,
- 31:51like, an anthropic or OpenAI or something on
- 31:54the back end. So there there's all these
- 31:56weird little change. I I would also encourage
- 31:59folks, like, if you are looking at this,
- 32:00like and some of the stuff we talked
- 32:02about, like, hey, this thing can monitor AI
- 32:05sites. That's, like, a fixed list of sites
- 32:08that come out. So so they'd is very
- 32:11comprehensive.
- 32:12Like, I I will say, like, if you
- 32:14go through, like, the list of sites that
- 32:16the it does cover, it's a lot. Like,
- 32:18I hadn't even heard of, like, 90% of
- 32:21them easy. But, you know, if there's something
- 32:23out there where you're like, oh, this is
- 32:25a new Gen AI tool
- 32:27or AI products that my organization uses, it
- 32:30might not be on the list. You gotta
- 32:31go figure that stuff out too. Yes. Microsoft
- 32:33does work to keep it updated, but, yeah,
- 32:36I mean, you or I could go stand
- 32:37up a site on Azure, give it some
- 32:39random domain name tied into
- 32:41some LLM on the back end, and all
- 32:43of a sudden, there's a new Gen AI
- 32:44site out there hosted by the podcast.
- 32:47Is that what you wanna do? No. That
- 32:50is not what I wanna do. But we
- 32:51could. Enough, it does monitor.
- 32:54So it does monitor for azure.microsoft.com
- 32:58and specifically the OpenAI service product page. So,
- 33:02like, the page that would tell you about,
- 33:04not the learn docs and, how to use
- 33:07it, but more like the marketing page with
- 33:09the pricing and things like that on there.
- 33:11So that's on the list, but, azure.com is
- 33:14not. So like you said, you can go,
- 33:17potentially spin these things up because you can
- 33:19absolutely get into the portal. But it had
- 33:21a lot of, like, I was kind of
- 33:22impressed. I had like Devon, a lot of
- 33:24the smaller stuff was baked into there as
- 33:27well. Along with that, like, just like the,
- 33:29I think the more like US centric
- 33:32AI or European offerings, but also
- 33:35stuff out of Asia Pacific and
- 33:37China and things like that. So it had
- 33:40broad coverage, but, you know, you gotta check
- 33:42the list to make sure your thing is
- 33:44there. Yep. Yeah. They've done a good job
- 33:46with it. So I think that's kind of
- 33:48the last aspect to it. And that's what
- 33:50I talked about in that presentation is
- 33:52not only do you need to go in
- 33:54and
- 33:55be, let's say, preventative
- 33:57about where this data is going with some
- 33:59of the cleaning up security, cleaning up permissions,
- 34:02starting to get labels on your content, knowing
- 34:04what content is out there, But then leveraging
- 34:06something like DSPM where you can do some
- 34:09of that deal those DLP policies, a lot
- 34:12of this is just baked in to help
- 34:14you kinda get there, I would say, quickly.
- 34:16But then also that kinda ongoing monitoring, like,
- 34:19going in and looking at the activity explorer
- 34:21in here and some of the
- 34:22reports in here to just see,
- 34:25like, I've had 540
- 34:27activities in the last thirty days with different
- 34:29interactions with Copilot and agents.
- 34:31Going and looking at my total visits to
- 34:33other AI apps, you can see I do
- 34:35tend
- 34:36to favor Copilot, but like Notion and Perplexity
- 34:39and Suno and some of those other AI
- 34:42tools that I have used have been showing
- 34:45up in here. So just being aware of
- 34:47where are your users
- 34:50potentially exposing your data to places they shouldn't
- 34:53be. Yeah. I I think it's, it's an
- 34:55important one.
- 34:57Absolutely. But with that, are we supposed to
- 34:59go back to work now? Is that what's
- 35:00supposed to happen? We are we are supposed
- 35:03to go back to work. Oh.
- 35:05Alright. Go back to work. Before you go
- 35:06back to work. This far. And you're still
- 35:08listening.
- 35:09Feel free to reach out. We have a
- 35:11contact us page on the website. Let us
- 35:13know what you'd like to hear about, what
- 35:14topics are interesting to you. Give us a
- 35:16ping on LinkedIn.
- 35:17You can find Ben there. You can find
- 35:19myself there. You can find the podcast there.
- 35:21And let us know
- 35:23what topics,
- 35:24people, interviews, anything like that would be interesting
- 35:28to you. Absolutely. We'd love to hear from
- 35:30you. So
- 35:31thanks. Alright. Thanks, Ben. As always, I appreciate
- 35:34it. Definitely. And we'll talk to you later.
- 35:38If you enjoyed the podcast, go leave us
- 35:40a five star rating in iTunes. It helps
- 35:42to get the word out so more IT
- 35:44pros can learn about Office three sixty five
- 35:46and Azure.
- 35:47If you have any questions you want us
- 35:49to address on the show, or feedback about
- 35:51the show, feel free to reach out via
- 35:53our website, Twitter, or Facebook.
- 35:56Thanks again for listening, and have a great
- 35:58day.