Latest / Microsoft Cloud IT Pro Podcast / Episode 434: Reading the Spec Sheet on Microsoft's Agentic SOC
Transcript
- 0:03- Welcome to episode 434
- 0:06- of the Microsoft Cloud IT Pro podcast recorded
- 0:09- live on 08/10/2026.
- 0:11- This is a show about Microsoft three sixty
- 0:14- five and Azure from the perspective of IT
- 0:16- pros and end users, where we discuss a
- 0:18- topic or recent news and how it relates
- 0:20- to you. In this episode, Ben and Scott
- 0:23- provide their perspective on Project Perception,
- 0:26- Microsoft's latest AI security tool that provides you
- 0:30- with a workforce of AI agents. We discuss
- 0:32- what perception really is, what you need to
- 0:35- have to get started, and the agents and
- 0:37- tools that are contained within the product.
- 0:40- Join us as we unpack
- 0:42- Project Perception.
- 0:45- Scott, I was just I mentioned this to
- 0:47- you. I had a statement
- 0:49- that I made two weeks ago when we
- 0:52- recorded on the whole authentication and passkeys, and
- 0:55- we mentioned that you can go discuss with
- 0:57- us on Reddit. And somebody was like, hey,
- 0:59- Ben. You said something incorrect, and they called
- 1:01- me out on Reddit. So, yes, I made
- 1:03- a couple of comments. One, I said I
- 1:05- thought synced pass keys were tied to authenticator,
- 1:07- and I had it backwards in my head.
- 1:09- The device bound our authenticator. You can't do
- 1:11- synced pass keys in authenticator.
- 1:13- So,
- 1:14- yeah, I had it backwards in my head,
- 1:16- and
- 1:17- we were talking about one password. I also
- 1:19- discovered one password does work now for synced
- 1:22- pass keys. I just had a misconfiguration
- 1:24- of my tenant that was preventing me from
- 1:25- doing it. I think I have misconfiguration of
- 1:27- my tenant or my employees. I can you
- 1:30- make configuration
- 1:31- changes in your tenant? Not like that. They
- 1:33- don't let you go just add random
- 1:36- passkey
- 1:38- GUIDs to I don't even know. Is it
- 1:40- GUIDs? Is it GUIDs? Is it yeah. I
- 1:42- say GUID. You say GUID. I interchange them
- 1:44- sometimes.
- 1:45- You've got that whole Pan Panhandle thing going
- 1:48- on. Like, there there's a there there's a
- 1:50- little bit of an accent from the Panhandle.
- 1:51- Yeah. Usually, I say GUID, and then every
- 1:53- once in a while, I get that GUID
- 1:55- that sneaks in there. No. The AA GUID,
- 1:57- they don't let you just add those randomly
- 1:59- to your tenant? Probably not. Do you have
- 2:01- agents
- 2:01- scanning your tenant to look for those types
- 2:03- of things? I'm sure. I do. But, I
- 2:05- mean, what is an agent?
- 2:07- What is an agent?
- 2:09- What is a skill? What is AI?
- 2:13- Yes. You can pontificate.
- 2:16- Agent Smith? Agent
- 2:17- Smith? Oh, oh, no. We're gonna go Skynet.
- 2:19- Agent Smith. We're gonna go Skynet first. Oh,
- 2:20- Agent Smith. No. We can do Agent Smith.
- 2:22- Do you realize how old that movie is?
- 2:25- We talked about this before. Came out '99,
- 2:27- Scott. That that that We're going on twenty
- 2:29- seven years. Mhmm.
- 2:31- Yeah. Yeah. My my
- 2:33- we were watching a TV show the other
- 2:35- day, and
- 2:36- there was a line in there. It was
- 2:38- a it was a teenager talking to we're
- 2:41- watching a movie, but it was a teenager
- 2:43- talking to her parent.
- 2:45- And
- 2:46- her dad was born in 1980, like, from
- 2:49- the nineteen hundreds Like kinda kinda Yeah. And
- 2:51- and and she asked this quipped to her
- 2:53- dad that
- 2:54- you were born closer to 1940
- 2:57- than I was to 1980.
- 2:59- Meaning, like, when we were born, we were
- 3:00- closer to the forties than a kid today,
- 3:03- forty five, forty six years later,
- 3:05- would actually be to 1980.
- 3:07- That that was a really weird one for
- 3:09- me. It's like I'm I'm old in multiple
- 3:11- dimensions.
- 3:12- Yeah. I don't know that I wanna think
- 3:14- about that too hard. Yeah. I just wanna
- 3:16- pretend I'm still 20. Yeah. Well, I I
- 3:18- don't know that I know this could tear
- 3:19- all all my gray hair and everything, but
- 3:21- I got one more for you. Listen. You
- 3:23- hear that? The sound of silence? I don't
- 3:25- hear the sound of the kids. Are back
- 3:27- in school.
- 3:29- I get my house back. I get my
- 3:31- office back. How do you get how does
- 3:33- it feel to be back in silence?
- 3:35- It would be okay, except I got this
- 3:37- puppy behind me. So if he starts screeching,
- 3:39- you you you'll know why. Either he didn't
- 3:41- like the content or he really needed to
- 3:43- go outside. It is a cute new puppy.
- 3:45- Yeah. Alright. I will let you enjoy the
- 3:47- puppies. I'm not gonna enjoy the puppies. I
- 3:49- even had a good segue into agents. We
- 3:51- got distracted by puppies in square islands. Sorry.
- 3:54- I took you there. We'll get back. It's
- 3:55- okay. Human in the loop. Human in the
- 3:57- loop. We're back. Human striked. Okay. That was
- 4:00- quite the perception there, Scott, of
- 4:02- the
- 4:04- wow. Okay. So agents. Human in the loop,
- 4:07- a new product from Microsoft.
- 4:09- A new service? I don't know. At what
- 4:10- point in time, speaking of agents and skills,
- 4:13- what's a product now and what's a service?
- 4:15- I don't know anymore. Who knows? Up is
- 4:17- down. Up is down. Left is right. But,
- 4:19- yes. Prod Project Perception? Project Perception.
- 4:23- What are we perceiving? Project Perception. What are
- 4:26- we perceiving? We are perceiving
- 4:29- tenant security.
- 4:30- Like,
- 4:31- this is an interesting one. It is a
- 4:34- new security
- 4:36- product or service from Microsoft
- 4:38- and right from
- 4:40- the website for Project Perception, a workforce.
- 4:43- So now you aren't maybe this isn't even
- 4:45- a service or a product. It is a
- 4:47- workforce.
- 4:48- It's a workforce of specialized
- 4:50- AI agents,
- 4:51- red, blue, and green,
- 4:53- that reason across your security data tools, workflows
- 4:56- to expose gaps, investigate threats, remediate them, use
- 4:59- up the strategy, agents carry the load. So
- 5:02- this is an agentic
- 5:04- security
- 5:05- based service
- 5:06- from Microsoft
- 5:08- that you can
- 5:10- go in and enable at some point in
- 5:12- time. Right now, it's like a limited public
- 5:15- preview,
- 5:16- so
- 5:17- very selective in who can get access to
- 5:20- it. But all the documentation, the websites, everything
- 5:23- else is out there so that you can
- 5:26- now
- 5:27- enable agents to secure your environment.
- 5:30- Can you though?
- 5:31- Ish.
- 5:32- What is an agent, Scott?
- 5:34- I think that's one of the biggest things
- 5:36- that I I ran into this. But before
- 5:39- we get, like, too deep into that, maybe
- 5:41- we should talk a little bit about
- 5:44- what Project Perception, like, proposes to be and
- 5:48- what the agents it proposes to be
- 5:51- that are in this limited public preview that
- 5:53- are out there. And then we can maybe
- 5:55- come back a little bit to what is
- 5:57- an agent and
- 5:59- what does, like, supervision
- 6:00- look like
- 6:01- in particularly context of project perception and what
- 6:05- they've released in this first iteration here? Yeah.
- 6:07- And even some of the onboarding of this,
- 6:10- I had to pull together.
- 6:13- Again, limited public preview,
- 6:15- all over the documentation, things are liable to
- 6:17- change,
- 6:19- But I had to pull together a few
- 6:21- different things
- 6:23- to figure out, like, what do you even
- 6:25- need or what does this look like from
- 6:27- an onboarding perspective? Because
- 6:30- you have the home page, and then when
- 6:32- you go down in the documentation and start
- 6:34- looking at it, it has
- 6:37- certain requirements.
- 6:38- So if this is something you're interested in,
- 6:42- just getting started with it and looking through
- 6:45- some of the prerequisites
- 6:46- is this is something that's gonna live in
- 6:48- the Microsoft Defender portal, so you'll need to
- 6:51- have access to Defender, Microsoft Defender XDR Got
- 6:55- security center. Defender. Yes. I can keep coming
- 6:57- up with acronyms. Yeah. Security.microsoft.com.
- 7:00- You do need to have a unified
- 7:03- role based access
- 7:04- control, RBAC? RBAC. This is this was a
- 7:07- new acronym that was popping up all over
- 7:09- the documentation for this one.
- 7:11- Yes. That needs to be enabled across all
- 7:13- Defender workflow or workloads.
- 7:16- You have to be a security admin, security
- 7:18- reader.
- 7:19- And then even as you get into some
- 7:21- of the agents then, it talks about,
- 7:24- I believe, with some of these agents, like
- 7:26- different licensing, you need to have
- 7:28- Defender for Cloud enabled to look at your
- 7:31- Azure
- 7:32- stuff. You need to have
- 7:34- certain prerequisites
- 7:36- for various workloads, so Defender for Office three
- 7:38- sixty five
- 7:40- plan two. If you wanna look at email
- 7:43- and collaboration
- 7:44- stuff, you need to have Defender for Cloud,
- 7:46- Defender for Containers.
- 7:48- If you're wanting to look at your Azure
- 7:49- cloud,
- 7:50- Defender or Entra ID p two, Defender for
- 7:53- Identity, Defender for Cloud apps to look at
- 7:55- identity alerts.
- 7:56- So this is one part that for me,
- 7:58- it's like all through the documentation, it take
- 8:00- talks about the licensing and Urbach and all
- 8:03- of that. But then when you go back
- 8:05- to the actual product page and you scroll
- 8:08- all the way down and you look at
- 8:09- how is project perception priced, then it comes
- 8:12- up with you pay for only what you
- 8:13- use measured
- 8:15- in security computing units.
- 8:17- The mythical the magical
- 8:19- SCU. Yeah. So good old security Copilot
- 8:22- SCUs.
- 8:23- So if this is something you're interested in,
- 8:26- much like some of the other things we've
- 8:28- talked about, this is not I would not
- 8:30- call this a low cost barrier to entry.
- 8:33- You're probably looking at you already have
- 8:36- Microsoft three sixty five e five
- 8:39- or at least a lot of the security
- 8:41- products from e five, and you're going to
- 8:43- be willing to pay for security compute units
- 8:46- to be able to go in and run
- 8:48- these agents. So that is my here's where
- 8:51- you go, how you wanna think about if
- 8:53- this is something you're interested in just from
- 8:55- a pure pricing, licensing,
- 8:57- onboarding perspective in that respect. It's got a
- 8:59- lot of
- 9:00- moving pieces and parts, and I think you
- 9:03- can see probably where it's, like, limited public
- 9:05- preview because it's for those folks out there
- 9:07- who are a little bit more in, like,
- 9:09- token maxi land or I think just kinda
- 9:12- some of the bleeding edge, like, really interested
- 9:14- in the
- 9:14- frontier programs and other things, like, that are
- 9:17- available today. Yeah. So once you turn all
- 9:19- that on, once you have everything licensed and
- 9:21- enabled and permission set, then you can go
- 9:24- start looking at these agents. Well, I mean,
- 9:27- I just get an AI that does it
- 9:28- all for me. Right? Yeah. You just tell
- 9:30- it go do your thing. Go be my
- 9:32- security team, my security workforce. Go be my
- 9:36- security workforce
- 9:38- indeed.
- 9:39- Well,
- 9:40- so so, yeah, you you get all this
- 9:42- stuff lit up. There is a lot of
- 9:43- licensing.
- 9:44- There is a lot of configuration.
- 9:47- I didn't see, like, configured this for me
- 9:49- agent as part of onboarding, so you do
- 9:51- have to go and kinda wrangle that, well,
- 9:54- altogether. But once you do get that going,
- 9:57- you get these
- 9:59- categories of agents, like, I guess, broadly. Right?
- 10:02- They group them into Yep. Red team, blue
- 10:04- team, and
- 10:06- green team. So red team,
- 10:09- traditional
- 10:10- kind of penetration side of things, like, out
- 10:12- there to
- 10:13- identify
- 10:15- what's exploitable
- 10:16- and potentially
- 10:17- exploit it, blue team to
- 10:21- monitor for those threats.
- 10:22- When they see them come through. Like, can
- 10:24- they reconstruct them, get through the path,
- 10:27- close any gaps? Like, if there was something
- 10:29- exploitable, like, hey. Let me come and close
- 10:32- those up real quick. And then green team
- 10:34- to
- 10:35- remediate
- 10:36- and harden
- 10:38- configuration. So
- 10:39- while blue team might be doing, like, active
- 10:42- monitoring, triaging alerts, things like that, green team's
- 10:45- out there kinda
- 10:47- constantly
- 10:48- analyzing your posture,
- 10:50- making sure to give you, like, a set
- 10:51- of
- 10:52- prioritized recommendations,
- 10:54- the things like that that are gonna help
- 10:56- you ultimately, like, harden that posture. So I
- 11:00- think that's, like, all good on the surface.
- 11:02- And you go like, Well, that would be
- 11:04- kind of interesting. Right? Like, so I've got
- 11:05- this thing that can get out there, and
- 11:07- it can try and Yep. Poke holes in
- 11:09- my configuration
- 11:10- and see what can actually get through. I've
- 11:13- got another piece over here that can kinda
- 11:15- monitor things and potentially
- 11:18- operationalize
- 11:19- that monitoring or that observability
- 11:22- and go close those gaps real quick. And
- 11:24- then I've got something else on the back
- 11:25- end that's kinda sitting more in a constant
- 11:27- loop and basically doing continuous improvement and run,
- 11:31- run, run, kinda like CICD
- 11:33- for your tenant configuration is what it sounds
- 11:36- like. I think it's very easy to say,
- 11:38- like, oh, hey. That's what it all reads
- 11:40- as. But, ultimately,
- 11:42- back to that original question of what is
- 11:45- an agent,
- 11:46- that that very much, like,
- 11:48- bleeds back into
- 11:50- what project perception
- 11:53- is here because,
- 11:55- like, that that that was the thing that
- 11:56- struck me. So you go out, you read
- 11:57- about it, you read the blog post, things
- 11:58- like that, you go through the docs, and
- 12:00- you're like, cool. And then you get down
- 12:01- to it, and
- 12:03- the the things that I was saying of,
- 12:05- hey. Let's do the the whole red, blue,
- 12:07- green, and it's a little bit more automated.
- 12:09- Turns out it's really heavy on human in
- 12:12- the loop. Everything is actually manually triggered,
- 12:16- and and and effectively read only, like, with,
- 12:19- like, one or two caveats, but pretty much
- 12:22- across the board, read only.
- 12:24- And if that's the case
- 12:26- and you're kinda like, you're doing all the
- 12:29- supervision and you're doing all the monitoring and
- 12:31- it is very heavy on human loop, is
- 12:32- it truly agentic, or is it more like
- 12:34- a set of capabilities that you've been provided
- 12:37- maybe as, like, a security engineer within your
- 12:40- organization
- 12:41- to go and kinda help you have additive
- 12:43- tools to understand your environment configuration, things like
- 12:47- that. I can see where it goes in
- 12:48- the future, but it's it's hard to to
- 12:51- today to to kinda
- 12:53- be able to, like, really wrangle it together.
- 12:55- And then I wonder back to the whole
- 12:57- pricing aspect of it. Like, is it worth
- 12:58- it? Like, what are you getting here?
- 13:01- I kinda walked away at the end. And
- 13:02- I know this is a little flippant of
- 13:04- me, but, like, you walk away at the
- 13:05- end of, like, there's a bunch of, like,
- 13:07- scripts and things out there that help you
- 13:09- evaluate your environment already. Like, what's additive here
- 13:12- that you're getting out of it? I have
- 13:13- done some of this type of stuff even
- 13:16- using existing
- 13:17- AI. So I think seeing some of this
- 13:20- and having it built in was a little
- 13:22- intriguing to me. I think as I look
- 13:25- at this,
- 13:26- certain agents are probably
- 13:29- I probably get different value from different agents
- 13:31- in this. And maybe we dive into them
- 13:34- like the red team agent. Like you said,
- 13:36- that's gonna be identifying the weakness.
- 13:38- This one to me is
- 13:40- I think there's some value here,
- 13:42- but I think you also have some limitations
- 13:45- here. This one, when you go look through
- 13:47- it, Microsoft has the recon agent that's on
- 13:50- the red team. There's multiple agents. There there's
- 13:53- Right. There's multiple agents under these different teams.
- 13:56- Yeah. Under the teams categories. Yes. So yeah.
- 13:59- So so as we go through this, it's
- 14:01- like, yeah. Red team has recon agent, blue
- 14:04- team has
- 14:05- four agents,
- 14:07- and green team has one agent under it.
- 14:09- So you could also potentially see, like, additional
- 14:12- capabilities being introduced over time
- 14:15- and having that stuff come through. But yeah.
- 14:18- So so maybe we start with the recon
- 14:20- agent out there
- 14:22- scouting around. These are always interesting. Right? Recon
- 14:25- agent to me because it's simulating attack behavior,
- 14:29- trying to expose vulnerabilities
- 14:31- that
- 14:32- adversaries could exploit. It's looking at things like
- 14:35- mapping attack paths,
- 14:37- looking at choke points in your configuration. It
- 14:40- does have vulnerability
- 14:41- code scanning in there. I think maybe that
- 14:44- one jumps out at me a little bit,
- 14:46- exposes detection gaps, but it's all simulated.
- 14:49- Right? And this is the part where you
- 14:52- start thinking of you can only get so
- 14:55- far, I think, in a simulation. You can
- 14:57- only hit an API so many times before
- 14:59- you actually would have to make a configuration
- 15:01- change, like, to go through the next step.
- 15:04- So while I think it can find
- 15:06- it's almost like it could find surface level
- 15:08- stuff, but it has a limit to the
- 15:11- depth that it could go to. Because if
- 15:13- you have to expose one vulnerability,
- 15:15- how can it know that if it maybe
- 15:17- gets past the first layer, that it can
- 15:19- move laterally?
- 15:20- Or if it exploits the certain code vulnerability,
- 15:24- how does it know if it can keep
- 15:26- taking the next step? Like, this is maybe
- 15:29- that first line. Maybe some of the AI
- 15:32- stuff can kinda say, well, if I can
- 15:33- get past here, then I can go do
- 15:35- this, then this. It's
- 15:38- but I also think there's some value here.
- 15:39- Some of the other things I've done with
- 15:42- some of my other AI tools and looking
- 15:44- at this is taking some of the correlation
- 15:46- too between like what's known out there going
- 15:48- out looking at the web. What are known
- 15:50- CVEs out there? What are known exploits? What
- 15:53- are some of those things that are known
- 15:55- to be exploitable?
- 15:57- Do you see them in the configuration of
- 15:59- my tenant where an exploit from over here
- 16:01- could be used over here in my tenant?
- 16:04- The red team to me, a lot of
- 16:05- the investigatory
- 16:08- probing,
- 16:09- I think there's a level, but I also
- 16:11- think there's a lot of
- 16:13- correlations and maybe looking at things from different
- 16:17- angles
- 16:18- where you can get some value from AI
- 16:20- here. I think it's back to the whole
- 16:21- what's an agent thing. I think it's interesting
- 16:23- that they called it an agent because, like
- 16:25- like, there is I'm not arguing that there's
- 16:27- not value in it, but I think the
- 16:28- presentation's interesting.
- 16:30- So if you go and read the docs
- 16:33- and you start hopping in and you say,
- 16:35- okay. What's the recon agent and what's it
- 16:37- do? It's very clearly positioned
- 16:40- as
- 16:41- a single pane of glass
- 16:43- across
- 16:44- multiple Defender products. Right? So so this is
- 16:47- additive to Defender. So Defender is kinda the
- 16:49- base. It's out there. It's what protects you.
- 16:52- So so that's number one is it's a
- 16:54- little kinda like it is that little canned
- 16:56- reporting kinda thing, right, of saying, hey. Let's
- 16:58- go down the path, analyze your environment configuration.
- 17:01- I think where the AI thing comes in
- 17:02- and where it's a little bit less of
- 17:03- a canned report, but it's still very much
- 17:05- a report at the end of the day,
- 17:07- is the ability to go look at your
- 17:09- specific tenant configuration.
- 17:10- Let's say, the configuration
- 17:13- for
- 17:13- maybe one of your subscriptions in Azure.
- 17:16- It's able to hop through that subscription specifically
- 17:19- and kinda do
- 17:21- more
- 17:23- more pivots
- 17:24- and kinda what if analysis
- 17:25- than a traditional,
- 17:27- alright, go check this binary, if this, then
- 17:29- that,
- 17:30- kind of thing. I think that's where AI
- 17:32- helps out these things. It makes it a
- 17:33- little bit more of a choose your own
- 17:35- adventure.
- 17:36- But, ultimately, getting to
- 17:37- that single
- 17:39- pane of glass
- 17:41- that that's out there, I think even that
- 17:43- single pane of glass is
- 17:45- interesting the way it's kinda presented
- 17:47- back to you. So, like, when you go
- 17:49- and run these things,
- 17:51- you you don't get a traditional dashboard. Like,
- 17:53- it's not like you get, like, a Power
- 17:54- BI report or anything like that. You get
- 17:56- a bunch of markdown files in a kinda
- 17:59- known format. But if you were to run
- 18:01- it against my tenant and then run it
- 18:03- against your tenant, we should get different sets
- 18:05- of we get the same set of output
- 18:07- mark markdown files, but with different guidance in
- 18:09- each one. The markdown files to me makes
- 18:11- sense because
- 18:13- one of the other things, like, if we
- 18:15- wanna go take this the next step, is
- 18:17- you get all this markdown.
- 18:19- It does talk about, and this is maybe
- 18:21- where some of that value comes in with
- 18:23- project
- 18:24- perception, is it can take the output of
- 18:27- one agent and feed it to the next
- 18:28- agent. Which you kinda need. Right? Like, if
- 18:31- you're gonna go and say, hey. I found
- 18:33- something wrong in your configuration. You kinda wanna
- 18:35- pass that off to the next thing to
- 18:37- have it evaluate. Say, like, okay. Is this
- 18:39- important? Where does it land? And that's where
- 18:41- kinda, like, the blue and green come in.
- 18:43- We'll just say, hey. Let's go what's prioritization
- 18:46- look like here? Like, that's the responsibility of
- 18:48- the green team agent. It's not the responsibility
- 18:50- of any of the blue team agents that
- 18:52- are out there that are kinda doing active
- 18:54- triage or things like that. Yeah. And that's
- 18:55- where that red team it's like, we've gone
- 18:57- in. We've done the recon. We figured out
- 18:59- all these things. Now let's pass it over
- 19:01- to that blue team agent.
- 19:03- And this one, there's the most in here
- 19:05- where you have
- 19:07- four agents in the blue team. You have
- 19:08- the triage agent, you have the attack
- 19:11- investigation
- 19:12- agent,
- 19:13- the threat intelligence agent,
- 19:15- and the detection
- 19:16- authoring agent. And I think this is one
- 19:19- where that red team agent can then take
- 19:20- those marks down to the blue team and
- 19:22- pass it over to the triage agent, where
- 19:24- this agent can now
- 19:26- go
- 19:27- take all of that
- 19:29- information.
- 19:31- This is where
- 19:33- I have mixed feelings on this because
- 19:35- the triage agent it would be interesting to
- 19:38- see how it passes. The triage agent has
- 19:39- defined classifies and triages as alerts in Defender
- 19:42- for different workloads, for different alert types.
- 19:45- It does have some things like autonomous
- 19:48- triage. It has some learning based on feedback,
- 19:51- record classification.
- 19:52- But if your
- 19:54- red team agent
- 19:56- is read only,
- 19:57- it shouldn't, in theory, actually trigger any alerts
- 20:01- in Defender because it's not taking any actions,
- 20:04- I don't think. Maybe it would. Maybe it
- 20:07- would try
- 20:09- like, what is a read only agent, I
- 20:10- guess, Scott? Is it actually gonna try to
- 20:12- log in? Is it gonna try, like, brute
- 20:15- forcing a password a little bit? In my
- 20:18- mind, that's not really read only. It's doing
- 20:20- a little bit more writing.
- 20:22- And to really get things to trigger in
- 20:25- Defender,
- 20:25- you have to do that type of stuff.
- 20:27- It would be nice to see this evolve
- 20:29- a little bit where
- 20:31- maybe the red the recon agent is allowed
- 20:34- to poke at your tenant a little bit
- 20:36- more. Microsoft's running it all, it would be
- 20:38- able to recognize it and flag it accordingly
- 20:42- so that blue team agent could actually
- 20:45- like, you could see a little bit more
- 20:47- of a simulated attack versus just a, let's
- 20:50- go read all the configuration. It's what I
- 20:52- was kind of planning for. It was like,
- 20:53- hey. When do these things get out? What
- 20:55- if we get a little more proactive? But
- 20:58- e even touching things can be a little
- 21:00- touchy sometimes when they were not there. So
- 21:02- if you look at these agents and so
- 21:04- so per agent, there's a configuration,
- 21:08- and that configuration comes with a set of
- 21:10- permissions
- 21:11- that are gonna be layered into the graph
- 21:12- and your tenant and other things that are
- 21:15- out there. So
- 21:16- if you look at, like, the recon agent,
- 21:18- it's a
- 21:19- it's it gets some pretty privileged
- 21:22- role assignments. Like, it does get user read
- 21:24- all. It gets group member read all, application
- 21:27- read all, delegated permissions read all, audit log
- 21:30- read all,
- 21:31- things that are out there. It also gets
- 21:32- a bunch of Azure RBAC assignments.
- 21:35- So you have to kinda go through, add
- 21:37- a subscription scope, and
- 21:40- assign these things down or assign it at
- 21:42- a higher level, like a management group, something
- 21:44- like that. Let it bleed through. But it
- 21:46- gets reader,
- 21:47- specifically Key Vault reader and Log Analytics reader,
- 21:50- as well as
- 21:51- security reader that are out there. So I
- 21:53- think even going through and just kinda poking
- 21:56- at things, maybe it'll set some things off
- 21:58- depending on how your environment's configured and what
- 22:01- it is. But Right. I really look at
- 22:03- like, today, like, the red team
- 22:05- agent would be out there analyzing your environment.
- 22:09- It's gonna, like, tell you
- 22:11- what the broad scope of maybe think about,
- 22:13- like, a funnel. Like, you're always gonna be
- 22:15- funneling it down. So broad set of things
- 22:18- that are out here. The blue team is
- 22:19- gonna be sitting out there and saying, like,
- 22:21- okay. What are my tasks I need to
- 22:23- go? What do I need to triage that's
- 22:26- out there? If I and this is where
- 22:28- maybe, like, even, like, the red agent goes
- 22:30- away a little bit is, like, you have
- 22:31- things like an attack investigation agent. So that
- 22:34- could be super helpful, right, because you've got
- 22:35- an active tenant. And you're out there. You
- 22:37- don't need the red team to come in
- 22:38- Yep. And look at things like that, tell
- 22:39- you about misconfigured
- 22:42- users,
- 22:43- anonymous logins,
- 22:44- all all those kinds of things.
- 22:46- So, yeah, you get that funnel, pass it
- 22:48- to blue team, and then, you get a
- 22:49- little bit more of the the the green
- 22:51- team expertise to say, like, hey. Here's what
- 22:53- we're gonna prioritize
- 22:55- to go and fix
- 22:57- that's out there, which is interesting. Like, the
- 22:59- green team is really focused on prioritization. Like,
- 23:01- there's one agent today. It's the posture prioritization
- 23:04- agent.
- 23:05- Like,
- 23:06- like, the says it on the tin where
- 23:09- it's gonna go and what it's gonna do
- 23:11- for you.
- 23:13- If you are managing file storage today, you've
- 23:16- probably felt the pressure. Refresh cycles are harder
- 23:19- to plan, costs and complexity are increasing, especially
- 23:22- as file workloads scale across regions.
- 23:25- That's why Microsoft customers trust Nasuni as an
- 23:28- Azure certified software.
- 23:30- Nasuni provides a global file system designed for
- 23:33- enterprise scale environments,
- 23:35- keeping file data secure, governed, and accessible across
- 23:38- Azure environments without added infrastructure complexity.
- 23:42- It's the foundation to keep your unstructured data
- 23:45- stable even as your infrastructure evolves. To see
- 23:48- how much you could save while keeping your
- 23:50- initiatives on track, visit nasuni.com/tco.
- 23:56- That's nasuni.com/tco.
- 24:01- I like the attack investigation agent. That if
- 24:04- I was gonna just from the title without
- 24:07- having played with these, look at these, I
- 24:09- can see a lot of value with the
- 24:11- attack investigation
- 24:12- agent.
- 24:13- I would like it to see it get
- 24:15- to a better price point. You need a
- 24:17- lot for this one. Right? Can you come
- 24:18- back to what do you need to turn
- 24:20- on to get all this stuff going? This
- 24:22- one's like XDR,
- 24:24- EU fives
- 24:25- for all the categories of attacks, identity attacks,
- 24:29- and
- 24:30- kind of cloud and compute side of things.
- 24:33- Kinda looking at a lot. Right? The cloud
- 24:35- side, you got Defender for cloud, Defender for
- 24:37- containers.
- 24:38- You got the identity side with Entra
- 24:41- where you need Defender for cloud apps, Defender
- 24:43- for identity, ID p twos. This one actually
- 24:46- calls out you need the e five license,
- 24:48- the M365E5.
- 24:50- But I have a lot of clients, and
- 24:51- I say the value of this one, I
- 24:53- have not a lot. I have a few
- 24:55- clients.
- 24:56- They're relatively small. Right? They don't have, like,
- 24:59- a dedicated
- 25:00- security team. They don't have a dedicated SOC,
- 25:03- but
- 25:04- they're big enough that they are maybe running
- 25:06- e fives, they're running Defender for endpoint,
- 25:11- they're Intune enrolled,
- 25:12- and something gets triggered. Some type of
- 25:16- not necessarily it's a known attack, but something
- 25:18- in Defender is triggered as a potential attack.
- 25:21- Defender is great. It is not the most
- 25:23- intuitive interface
- 25:25- for a IT generalist
- 25:28- to dig through, to figure out if they
- 25:30- need to be worried about a particular
- 25:32- alert or incident or something that is raised
- 25:34- in Defender. If Microsoft could figure out a
- 25:37- way to really
- 25:38- lower the cost of this, so some of
- 25:40- these SMB
- 25:42- type of
- 25:43- companies
- 25:44- could leverage this to help them with an
- 25:46- investigation of certain alerts and incidents raised in
- 25:49- their tenant. I think there's a lot of
- 25:51- value there in helping
- 25:53- what I I will say more IT generalists
- 25:56- understand
- 25:58- what they should be worried about or concerned
- 26:00- about when different alerts are raised in Defender.
- 26:02- And I think that's where this one fits
- 26:05- in the blue team is when you actually
- 26:08- have an attack, not taking inputs from the
- 26:10- red team, but actually looking at attacks
- 26:13- and trying to
- 26:15- help people identify
- 26:17- what actually is going on there, helping them
- 26:20- understand the attack path, the replaying the attack,
- 26:24- giving them a summary,
- 26:26- looking through what was the result of the
- 26:29- incident,
- 26:30- what were the different MITRE attack techniques that
- 26:34- we're seeing, like, what are all the different
- 26:36- components of the attack
- 26:39- to help someone understand that, like I think
- 26:41- that one's there. There's the threat intelligence agent.
- 26:44- Threat
- 26:45- intelligence
- 26:46- agent. I struggle with this one.
- 26:49- Yeah. Why why is that? How is this
- 26:51- different than a secure score? That would be
- 26:53- it in a summary. Threat
- 26:55- Intelligence is
- 26:56- looking at reports
- 26:58- or no. Maybe it wasn't this one. Maybe
- 26:59- it was the green one. No. It's more
- 27:01- the green one. Threat intel
- 27:03- I but I think this is still somewhat
- 27:06- of a
- 27:07- secure score A little bit. Ish. Very very
- 27:10- much so. So
- 27:11- it's taking
- 27:13- kinda public signals
- 27:15- and public reports and then identifying them out
- 27:18- against in environments
- 27:20- that are out there. So or I guess
- 27:22- your environment. This is where it comes down
- 27:24- to,
- 27:25- like, is it a report at the end
- 27:26- of the day? Like, yes. It's a report.
- 27:28- You get some markdown,
- 27:30- comes out of it, things like that. What
- 27:32- do I go and do with that? Then
- 27:34- it's up to you to to figure it
- 27:36- out. It's interesting what the different reports
- 27:39- give back.
- 27:41- So, like, if you look at, like, the
- 27:43- attack
- 27:44- investigation agent, for example, like, one of the
- 27:47- primary things that it's intended to do is
- 27:50- give you
- 27:52- a little bit of guidance. So it has
- 27:53- a a classification in the report, like an
- 27:55- incident verdict. So I'm gonna give you a
- 27:57- classification that says, is this a true positive,
- 28:00- a false positive, or benign with transparent reasoning?
- 28:04- So that can be a little helpful. Right?
- 28:05- Like, they drive you a little bit further
- 28:07- down, like,
- 28:08- something that maybe, like, a human doesn't need
- 28:10- to take a a hard decision on, but
- 28:13- can be human in the loop, things like
- 28:15- that. And that that's a little bit different
- 28:17- than, like, the threat intelligence agent, which is
- 28:19- still gonna give you your executive summary, all
- 28:22- all that kind of stuff. But
- 28:24- it's,
- 28:25- kinda just give you a a a profile
- 28:28- of a potential threat actor
- 28:30- and things that are out there. Interestingly, it
- 28:32- does do a little bit of kinda like
- 28:34- the attack investigation agent or maybe even like
- 28:38- that Rihan agent is doing. It's also kinda
- 28:41- looking for, like, indicators in your environment. So
- 28:44- if you're thinking about, like, identity logins, things
- 28:46- like that, like, that's looking at IP addresses,
- 28:49- domain names,
- 28:50- e email addresses.
- 28:52- And then
- 28:53- it's trying to leverage, because, again, Defender all
- 28:55- under the hood. Right? You've got this common
- 28:57- core platform in there. Yep. Pull in, like,
- 28:59- enrichment data from Defender.
- 29:02- So if there's,
- 29:04- any kind of known attack vectors, things like
- 29:06- that out there, it'll be able to classify
- 29:08- it against your environment. But to your point,
- 29:10- like, couldn't I just go and maybe look
- 29:12- at Secure Score or another dashboard? And, like,
- 29:13- that that stuff should be kinda screaming at
- 29:15- me to say go fix it. Maybe this
- 29:17- gives you a little bit more. The other
- 29:18- one that I feel like is kinda similar
- 29:20- to this is you already have and maybe
- 29:22- this is a superset. You already have vulnerability,
- 29:25- the vulnerability assessment, vulnerability management as well that's
- 29:28- in Defender that's already looking at your environment
- 29:31- and telling you where you're vulnerable, where you
- 29:34- have
- 29:35- users, machine software,
- 29:37- settings that are creating vulnerabilities
- 29:39- in your
- 29:41- environment. Again, maybe that's just one of the
- 29:43- items that feeds into this, but some of
- 29:46- these I feel like it's maybe a little
- 29:48- bit of a
- 29:49- stretch to say we need this agent doing
- 29:52- it because a lot of that functionality
- 29:55- is
- 29:55- already there to a certain extent. But you
- 29:57- kinda, like, want it to go just a
- 29:59- little bit more. Right? Give me the report,
- 30:01- take the output. Like, let me do something.
- 30:03- So so so, like, threat intelligence, for example.
- 30:05- Another one of the outputs that it has,
- 30:07- some of the artifacts. So you get this
- 30:09- markdown report that comes out. It also gives
- 30:11- you a bunch of kinda ready to go
- 30:15- Kusto queries, KQL queries for your Defender XDR
- 30:18- telemetry.
- 30:19- So that that's cool. Right? Like, not everybody
- 30:21- is a KQL expert, let alone understanding, like,
- 30:24- the schemas as they hop across maybe, like,
- 30:27- a set of tables and log analytics related
- 30:29- to Azure over to your security environment. Like,
- 30:31- that that stuff's,
- 30:32- like, can be weird to poke at. And
- 30:35- unless you, like, really know the schemas or
- 30:36- have internalized them, like, I could see that
- 30:38- being super helpful, but, like, why can't you
- 30:41- just go run those queries for me? Like,
- 30:43- so great. So you gave me a bunch
- 30:44- of ready to use Kusto queries that are
- 30:46- scoped to my environment that are gonna help
- 30:48- me search for indicators
- 30:50- of threats in my environment. Why don't I
- 30:51- just go run like, keep running it. Right?
- 30:53- Keep running it in a loop,
- 30:55- come back, and then kinda give me the
- 30:57- outputs, and let's go from there. So you
- 30:59- can see, like, the potential. Like, I think
- 31:01- this is where they would go with it,
- 31:03- but who knows? It's expensive to run it
- 31:05- in a loop, Scott.
- 31:06- That's gonna cost me both arms and legs
- 31:09- instead of just one arm and a leg.
- 31:11- Yeah.
- 31:13- But I do agree. Like, it would be
- 31:15- good to see it a little bit more
- 31:17- proactive,
- 31:19- automated
- 31:20- type of stuff. The other one are you
- 31:22- ready to move on to the detection authoring
- 31:24- agent? This is another one that intrigues me.
- 31:27- This one, I would say I can also
- 31:29- see a little more value here
- 31:31- of
- 31:32- this one is
- 31:33- to help you build
- 31:36- exactly what that sounds, detections.
- 31:38- What type of logic do you need to
- 31:40- have? What type of rules do you need
- 31:42- to have in place?
- 31:44- How should you be looking at your data
- 31:46- to help create
- 31:49- detections in your environment?
- 31:50- And
- 31:51- this one, to your point earlier about some
- 31:53- of the KQL and the schemas and generating
- 31:55- that, a lot of detection alerts right now
- 31:58- are based off
- 32:00- of KQL queries that you write or you
- 32:03- get from content packs and Sentinel where you
- 32:05- have all your data going in there, but
- 32:07- you need to write your
- 32:10- KQL queries and build your I completely am
- 32:13- blanking on the name of what they are
- 32:15- right now in Sentinel, but essentially your automated
- 32:17- triggers. It sees this data. It runs this
- 32:20- workbook every five minutes, fifteen minutes, hour,
- 32:22- but it still requires you to author all
- 32:24- of that.
- 32:25- This one
- 32:27- is
- 32:28- designed really to help people create
- 32:31- KQL based
- 32:33- detections,
- 32:34- map
- 32:35- detection coverage to MITRE attacks,
- 32:38- tune their detection methods. Especially your alerting logic.
- 32:42- Right. It's helping you build your alerting logic
- 32:44- in a way that maybe you don't need
- 32:46- to go in and understand all of that
- 32:48- logic to the
- 32:50- level of depth that you need to now.
- 32:52- I look at this kinda like how I
- 32:54- use AI to write PowerShell scripts. I can
- 32:56- go write a PowerShell script.
- 32:58- Sometimes it's easier just to let AI figure
- 33:00- out the logic and the tweaks and the
- 33:02- little nuances of it. I think this is
- 33:04- the same way. You can still go write
- 33:05- KQL. You can go write all of this.
- 33:08- This just helps with some of the nuances
- 33:10- and helps you refine
- 33:12- how you detect
- 33:14- or how you detect more alerts,
- 33:16- potential compromises in your environment, which I think
- 33:18- can be a good thing.
- 33:20- It can be very nuanced to write these
- 33:23- things
- 33:24- where you don't give so many false positives
- 33:26- that you just start ignoring them. I think
- 33:27- you're still a little and that's where it's
- 33:27- human in the loop. Right? Like, you're
- 33:31- depending on those
- 33:32- and so the agents, because it's a funnel,
- 33:33- they're meant to they are meant to feed
- 33:35- into each other. So but you you're sitting
- 33:37- there coming out of that threat intelligence agent,
- 33:40- which had those classifications,
- 33:41- like false
- 33:43- positive, true or or, yeah, false positive, true
- 33:45- positive, benign, and then you're taking it and
- 33:48- kind of passing it through passing parts of
- 33:51- it through, I guess, because you're kinda left
- 33:52- with, like,
- 33:53- the threat intelligence agent seems like it's giving
- 33:55- you a KQL to go and look and
- 33:58- hunt a little bit for yourself.
- 33:59- And then something like the detection authoring agent
- 34:02- is coming in. It's taking that
- 34:05- it's taking that report, and then it's still
- 34:08- generating you yet another set of
- 34:10- alerting
- 34:11- and detection logic that then you would go
- 34:13- and
- 34:14- deploy in your environment if you so choose.
- 34:16- Because, again, there's gonna be more, like, rationalization
- 34:19- here for you to say, hey.
- 34:21- Is this thing that the agent's telling me
- 34:22- something that I actually wanna go do? And
- 34:24- then that would make it very secure. Right?
- 34:26- You're sitting out there looking at a number
- 34:28- of, like, the what what I have to
- 34:29- do now? Like, a is still at 800
- 34:31- or I don't even know. I haven't paid
- 34:33- attention to, like, what the boundary is. But,
- 34:35- like, you're looking at a number, right, that
- 34:36- really doesn't have any meaning, and then you
- 34:38- gotta take a step back and figure out.
- 34:41- So, hopefully, some of these reports
- 34:43- come with them come with some of that
- 34:44- stuff. I would have loved to have seen,
- 34:46- like, sample reports in the documentation or something
- 34:48- like that, but they didn't have any of
- 34:50- that. They just give you the schemas, the
- 34:51- reports, or the high levels. What's gonna come
- 34:54- out of it? And then the final one,
- 34:55- green team. Ready to move on to green
- 34:57- team? The posture prioritization
- 35:00- agent. This is the one I was thinking
- 35:01- of that I'm like, this is just the
- 35:03- agentic agentic version
- 35:05- of the Secure Score. This helps teams and
- 35:08- stakeholders
- 35:10- understand what to fix first and why. In
- 35:12- my mind, that was the whole point of
- 35:13- the Secure Score. Like, let me go look
- 35:15- through the Secure Score and find where I'm
- 35:17- the most vulnerable or where I can get
- 35:19- the most bang for my buck in terms
- 35:20- of fixing and use the Secure Score to
- 35:23- prioritize it. Maybe this takes it a little
- 35:25- step further. It's looking at the inputs of
- 35:28- where are
- 35:30- attacks being seen,
- 35:33- what is
- 35:34- some of the
- 35:36- attack investigation
- 35:37- or the threat intelligence,
- 35:39- the detection,
- 35:40- authoring that you're writing.
- 35:42- Is this taking inputs from the blue team
- 35:44- to kinda reprioritize?
- 35:46- Because the one thing I would say about
- 35:47- Secure Score is
- 35:50- it was very much, is this configuration on,
- 35:52- and how many users, how many machines, etcetera,
- 35:54- is it applied to? You just get a
- 35:56- number. The prioritization
- 35:58- of Secure Score
- 36:00- wasn't
- 36:02- necessarily based
- 36:04- on what I would say real world data,
- 36:06- real world analytics, and what's going on in
- 36:08- your tenant other than a settings offer on
- 36:11- across x number of people. So maybe this
- 36:14- helps with a little bit more real time
- 36:16- prioritization,
- 36:18- taking in a few more signals other than
- 36:20- just is a configuration on or off. But
- 36:23- this is going to not go in and
- 36:25- make fixes for you, going back to your
- 36:27- human in the loop, the read only, it's
- 36:29- just going to tell you here's the order
- 36:31- you should go apply all these settings or
- 36:33- fix all these things in your environment. The
- 36:35- other thing we should maybe talk about here
- 36:37- as we're kinda getting towards the end is
- 36:39- orchestrating these things or pushing them together. It's
- 36:42- great. So you you get all these agents
- 36:44- that are out there.
- 36:45- Like, what does it look like
- 36:48- to run them? I think the the
- 36:51- the most interesting one to me was probably
- 36:52- the playbook things or the playbook play. Go
- 36:55- go and kick these things off. So if
- 36:58- you did have something in the Defender portal
- 37:01- pop up, so you get an incident to
- 37:02- where you get an alert that pops up,
- 37:04- you can go run these playbooks, and then
- 37:06- the playbooks act as orchestrators
- 37:09- for the underlying agent definitions that are out
- 37:11- there. So they can kinda
- 37:13- go start the flywheel for you, but then
- 37:15- you'll still be human in the loop for
- 37:17- kinda those decisions along the way that we
- 37:19- talked about. Yep. And these are very much
- 37:21- this is something I would love to be
- 37:23- able to start playing with. Maybe
- 37:25- once it becomes generally available, I'll just spin
- 37:28- it up for a few days to play
- 37:29- with it. But Playbooks very much came from
- 37:31- Security Copilot. Like, people already have
- 37:34- Security Playbooks in there running a series Yeah.
- 37:36- I mean, you can see the evolution here.
- 37:39- Like, this is just I feel like this
- 37:41- is even further integrating Security Copilot into where
- 37:44- people work. Now you don't have to go
- 37:46- over there, you can go trader these playbooks,
- 37:48- tie in these agents that, I mean, realistically,
- 37:51- maybe are a part of Security Copilot or
- 37:54- this is the next
- 37:56- iteration of Security Copilot
- 37:58- within your environment. But, yeah, the playbooks look
- 38:01- like they're gonna be that main
- 38:04- that main service that you're gonna go use
- 38:07- or that main functionality
- 38:09- that you're gonna use to go trigger all
- 38:11- these prompts and these agents from various areas
- 38:15- within Defender or within your environment. Defender then
- 38:17- it's interesting. It becomes part of that single
- 38:19- pane of glass as well, which it was
- 38:21- meant to be anyway. So
- 38:23- you you can kinda come at it a
- 38:24- couple different ways. Exactly.
- 38:27- Anything else? Sessions? I mean, sessions are just
- 38:30- monitoring your agent work. Session, just like any
- 38:32- session in AI. But it's interesting. I'm curious
- 38:35- to see where this goes. I am curious
- 38:37- to get my hands on it and
- 38:39- use it sparingly
- 38:41- so I don't run out of money. But
- 38:43- and to see to your point where agents
- 38:45- are going, this is still very much a
- 38:48- lot of read only.
- 38:49- As
- 38:50- these get better and better, is there a
- 38:53- point we get to
- 38:55- where an alert gets detected
- 38:57- in one of the, I guess, the blue
- 38:59- team agents would be detecting them and then
- 39:02- triggering off a green team agent where it
- 39:04- actually can go in and
- 39:06- make the resolution, make the fix to whatever
- 39:09- the issue is? We were talking about it
- 39:11- earlier. I feel like I'd still be a
- 39:12- little nervous of that. There's a whole weird
- 39:14- thing here. Right? So so do you trust
- 39:16- these identities? So so, like, that recon agent
- 39:18- is out there, but it's it's got a
- 39:20- bunch of those graph permissions. Like, okay. Hey.
- 39:23- And it's got Azure subscription scope rules in
- 39:25- it. So, sure, they're read only, but it
- 39:28- does expose your configuration.
- 39:29- Like,
- 39:30- what what happens if that configuration,
- 39:33- like, report doesn't get published to the right
- 39:35- place?
- 39:36- There's a big question about who supervises these
- 39:39- things.
- 39:39- So if you're gonna be, like, human in
- 39:42- the loop for these things
- 39:44- and you're gonna be supervising it, I think
- 39:46- you'd want those folks to have kinda the
- 39:49- same or greater permissions than the agent. Right?
- 39:51- You wouldn't want the agent to be able
- 39:53- to run away past that? Yeah. Because you
- 39:54- wanna be able to go in and validate
- 39:56- some of the data coming back, I think.
- 39:58- I would. But what does that look like
- 39:59- in a world of,
- 40:01- like, junior to senior level, like,
- 40:04- folks in your security operations center. Like, if
- 40:07- you're sitting out there, typically, like, junior folks
- 40:09- are gonna be, like, very scoped down in
- 40:11- their permission. Like, now you can't have them
- 40:13- fully monitor the agent because, one, they could
- 40:15- be exposed to configuration they're not supposed to,
- 40:17- and, two, they don't have the same rights
- 40:20- as as they need to anyway.
- 40:23- And then I think it will be interesting
- 40:25- as more, like, activity based things come in.
- 40:27- So where you can maybe do things like,
- 40:30- oh, say, like, disable a user account based
- 40:32- on, on, like, anomalous
- 40:34- logins or things like that today.
- 40:36- Where do you start to kinda peel back
- 40:38- permissions
- 40:39- or become more permissive, I guess, in in
- 40:41- what you put out there? And when do
- 40:43- you start to kinda tear down approval gates?
- 40:45- Because if you go through the documentation,
- 40:48- one of the things that'll pop up a
- 40:49- couple times is
- 40:51- the suggestion,
- 40:53- not like the discrete guidance to go do
- 40:55- this, but it is definitely a suggestion
- 40:57- to just reduce, like, approval gates and some
- 41:00- of the friction that's there for those agents.
- 41:03- So, like, hey. It's proven. Like, I I
- 41:05- had to do this thing. It feels like
- 41:06- it's deterministic, and it's out there. Is it
- 41:08- gonna be truly deterministic?
- 41:10- Right. Who knows? Maybe it is today, but
- 41:12- is it when, like, the next underlying, like,
- 41:13- foundation model comes out or things like that?
- 41:16- Oh, that that's the other one to think
- 41:18- about is I didn't see any place where
- 41:22- it's explicitly
- 41:24- called out
- 41:25- what
- 41:26- models
- 41:27- these agents use under the hood. It's kinda
- 41:30- hand wavy. It talks about, like,
- 41:31- OpenAI
- 41:33- Foundry models, Azure OpenAI Foundry models,
- 41:36- but which ones and what's out there. Because
- 41:40- very much, like, sometimes, like, the model does
- 41:42- have a big play
- 41:43- in
- 41:44- what type of output you're going to get
- 41:48- from the same set of inputs. Right? Like,
- 41:50- you'd be as structured as you wanna be
- 41:51- in that agent configuration
- 41:53- and and flow of it. But throwing it
- 41:56- against, like, GPT five five versus
- 41:59- the new five six, like, Sol Luna Terra
- 42:02- stuff, like, gives you very different results. So
- 42:04- my understanding of this
- 42:06- based on the blog post,
- 42:09- and I didn't see it in the documentation,
- 42:11- is that it is very much a
- 42:14- multimodel
- 42:15- where it might be using some OpenAI,
- 42:17- it might be using some Claude.
- 42:19- It also looks like it is using and
- 42:22- we didn't even talk about this, the MAI.
- 42:25- I don't know how you pronounce Microsoft's models
- 42:27- now. May, my, the CyberOne
- 42:30- Flash, which is Microsoft's
- 42:32- new security model
- 42:34- that they claim it scored better above Mythos.
- 42:37- I think all of these companies run their
- 42:39- own benchmarks
- 42:40- on their own models. Your mileage is gonna
- 42:42- vary with benchmarks and what are out there.
- 42:45- Yeah. On
- 42:46- how it is, but it does say it
- 42:48- oh, here. It does. Project perception will take
- 42:50- advantage of that Cyber One Flash model from
- 42:52- Microsoft that they created. For which parts? For
- 42:55- what? For which parts? It doesn't say I
- 42:57- don't know. I don't know. I don't know.
- 42:59- I don't know. I don't know. I don't
- 43:00- know. I don't know. I don't know. I
- 43:00- don't know. I like my kids. Ask me
- 43:02- impossible questions. Where all your SCU cost is
- 43:05- coming from. It's all hidden away under the
- 43:07- hood. That's
- 43:08- my understanding is I don't know that they're
- 43:11- publishing the model because I mean, do do
- 43:13- they even know which one gets passed off
- 43:15- to which model, or is this one of
- 43:17- those
- 43:18- where there's, like, the initial
- 43:20- foundational
- 43:22- model that initially prompts it, and then the
- 43:24- agents go out to find the right model
- 43:27- based on the tasks they wanna do? There
- 43:29- is some blurriness
- 43:31- there in which models it's using where. Other
- 43:34- than that, it is gonna use Microsoft's model,
- 43:37- its
- 43:37- aspects of it. How that propagates out to
- 43:40- other ones
- 43:41- is absolutely
- 43:43- vague.
- 43:44- I think it's gonna take some time to
- 43:46- figure it all out. Yes. Yeah. No single
- 43:48- model will be optimal for every security task.
- 43:50- Cyber defense requires the right model for the
- 43:52- right problem at the right time.
- 43:54- Project perception, the right model is determined by
- 43:57- the combination of quality, reliability, latency, and cost
- 44:00- rather than relying on a single model, blah
- 44:02- blah blah blah blah. So it is using
- 44:04- a slew of models that it picks based
- 44:07- on some rules Microsoft has in place. Yeah.
- 44:10- Got it. Interesting. But we will have that,
- 44:12- the blog post.
- 44:13- So there's the blog post out there. There's
- 44:15- the Microsoft
- 44:16- web web page out there that we'll have
- 44:17- in the show notes, and then
- 44:19- all the learned documentation
- 44:21- is already available. So you can go read
- 44:23- all about this,
- 44:24- see everything, you just can't get access to
- 44:28- it yet. They do have a button for
- 44:30- contact sales
- 44:31- where you would normally
- 44:33- go to, like, buy a preview or get
- 44:35- started with it. Anytime a website tells me
- 44:37- to contact sales, I think of this isn't
- 44:39- gonna be cheap.
- 44:41- That's just
- 44:42- Yeah. My pre programming.
- 44:45- So I'm excited to get my hands on
- 44:47- it and try it out and see what's
- 44:48- in there at some point in time. I
- 44:49- don't know when that will be. We'll have
- 44:51- a follow-up episode when we can actually play
- 44:52- with it. Follow-up.
- 44:54- Ben Ben plays with agents in the clap.
- 44:57- Yeah. No. I think it'll it'll be interesting
- 45:00- to see how it forms
- 45:01- over time and, well, how, like, how the
- 45:04- manifestations
- 45:05- of those agents
- 45:06- change. Like I said, they're very they're very
- 45:09- high level today,
- 45:11- and it's gonna be it's gonna be quite
- 45:13- the journey to see them go from being
- 45:15- able to affect, like, tenant configuration changes, things
- 45:18- like that. It's like a hard no and
- 45:19- read only to
- 45:21- more broad yeses and what can be out
- 45:23- there and be done with it. With that,
- 45:25- I have a meeting coming up. I would
- 45:26- love to hear other people's thoughts on this
- 45:28- if somebody listening actually
- 45:30- is in the public preview and has access
- 45:33- to it and you're allowed to talk about
- 45:34- it. Feel free to reach out. Like I
- 45:36- said, this past week, we had some discussion
- 45:38- on Reddit about some of these things on
- 45:40- the MS Cloud IT Pro
- 45:42- subreddit. Feel free to discuss on LinkedIn as
- 45:45- well. Would love to
- 45:47- chat with people there and get people's thoughts.
- 45:50- Also, Scott, if there is anybody that's gonna
- 45:52- be at TechCon three sixty five in Seattle
- 45:55- the end of August,
- 45:56- it'll be, like, two weeks after this episode
- 45:59- comes out. I will be out there doing
- 46:01- a few sessions and would love to connect
- 46:02- with anybody that will be out in
- 46:05- Seattle at the end of August for that
- 46:07- conference. Got any codes or anything for them?
- 46:09- I do have a code. I just I
- 46:10- should know what it is. Oh, I shoulda
- 46:13- looked it up. It is
- 46:14- it can be in the show notes. I
- 46:16- think it's I'm not going to even
- 46:20- make one up. Kinda put you on the
- 46:21- spot, so we'll put it in the show
- 46:23- notes. There's two Bens you did. We'll put
- 46:25- it in the show notes. There's two Bens,
- 46:27- and I can't remember if I'm Ben or
- 46:28- Ben too. I mean, you are well, you
- 46:30- are Ben as well. That's okay. I am
- 46:32- Ben as well, and I was trying to
- 46:33- find one of my posts here that
- 46:36- oh, here it is. I know where I
- 46:37- can find it. Oh, super quick. How quick
- 46:39- can I go? I found a post that
- 46:42- has my code.
- 46:44- Ben200BEN,
- 46:46- and then the number 200. We'll save you
- 46:48- $200. In February. And TekCon in Seattle? TekCon
- 46:51- in Seattle. Yep. August
- 46:53- '28.
- 46:54- Alright. Well, thanks for
- 46:57- sitting on. It's always good to walk through
- 46:59- a spec sheet with you here, Ben. Yes.
- 47:00- For sure. Thank you. Thank you for your
- 47:02- perspective on perception. Perception. Okay. Your perception of
- 47:05- perception. We're we're gonna get all we're gonna
- 47:07- get all twisted around.
- 47:09- Alright. It's Monday. Go enjoy your meetings. Thanks,
- 47:12- Ben. Alright. Thanks, Scott.
- 47:15- If you enjoyed the podcast, go leave us
- 47:17- a five star rating in iTunes. It helps
- 47:19- to get the word out so more IT
- 47:21- pros can learn about Office three sixty five
- 47:23- in and Azure.
- 47:24- If you have any questions you want us
- 47:26- to address on the show, or feedback about
- 47:28- the show, feel free to reach out via
- 47:31- our website, Twitter, or Facebook.
- 47:33- Thanks again for listening, and have a great
- 47:35- day.