WEBVTT 00:00:03.679 --> 00:00:06.240 Welcome to the Azure Security Podcast, where 00:00:06.240 --> 00:00:08.759 we discuss topics relating to security, privacy, 00:00:09.039 --> 00:00:11.480 reliability, and compliance on the Microsoft 00:00:11.480 --> 00:00:15.880 Cloud Platform. Hey everyone, welcome to episode 00:00:15.880 --> 00:00:19.100 112. This week it's just myself, Michael. Everyone 00:00:19.100 --> 00:00:22.079 else is busy. And our guest this week is Ryan 00:00:22.079 --> 00:00:24.920 Munch, who's here to talk to us about security 00:00:24.920 --> 00:00:27.920 co -pilots and security co -pilot agents. But 00:00:27.920 --> 00:00:30.140 before we get to our guest, I just want to take 00:00:30.140 --> 00:00:31.300 a little lap around the news. I actually have 00:00:31.300 --> 00:00:34.759 some pretty important items to talk about. The 00:00:34.759 --> 00:00:37.179 first one, this is honestly just from my old 00:00:37.179 --> 00:00:40.200 stomping ground in Azure Data, but this is actually 00:00:40.200 --> 00:00:41.939 really important. I really want to stress this 00:00:41.939 --> 00:00:44.780 one. There's a blog post written by one of my 00:00:44.780 --> 00:00:47.179 old colleagues about why it's time to migrate 00:00:47.179 --> 00:00:51.060 from system .data .sqlclient to microsoft .data 00:00:51.060 --> 00:00:55.100 .sqlclient. It's not just about adding features. 00:00:55.179 --> 00:00:58.719 It's also about the fact that Microsoft .data 00:00:58.719 --> 00:01:02.859 .sqlclient will support more security capabilities 00:01:02.859 --> 00:01:06.079 moving forward. So for example, things like TLS 00:01:06.079 --> 00:01:08.780 being enabled by default, that's all enforced 00:01:08.780 --> 00:01:12.560 in the client. So if you've got managed code 00:01:12.560 --> 00:01:15.799 that is currently running on system .data .sqlclient, 00:01:15.900 --> 00:01:18.879 you need to start moving across to Microsoft 00:01:18.879 --> 00:01:24.590 .data .sqlclient. Next one is if you're using 00:01:24.590 --> 00:01:28.530 Azure DevTest Labs, we now support Generation 00:01:28.530 --> 00:01:32.349 2 VMs using Trusted Launch. So this will give 00:01:32.349 --> 00:01:36.750 you much higher kind of robustness and basically 00:01:36.750 --> 00:01:39.810 just more secure workloads because it will support 00:01:39.810 --> 00:01:43.430 things like secure boot, virtual TPM, virtualization 00:01:43.430 --> 00:01:45.890 -based security or VPS, as well as Microsoft 00:01:45.890 --> 00:01:49.000 Defender for cloud integration. Next one is virtual 00:01:49.000 --> 00:01:52.140 network tap. So we now support a virtual network 00:01:52.140 --> 00:01:55.120 terminal access point or a tap that allows you 00:01:55.120 --> 00:01:58.099 to continuously stream virtual machine network 00:01:58.099 --> 00:02:01.000 traffic to a network packet collector or analysis 00:02:01.000 --> 00:02:04.400 tool. Very, very cool. Very commonly used, but 00:02:04.400 --> 00:02:07.980 now we support it in our VMs. Azure Front Door 00:02:07.980 --> 00:02:10.960 Web Application Firewall, or the WAF, now supports 00:02:10.960 --> 00:02:13.800 Capture. This is actually in preview, but it's 00:02:13.800 --> 00:02:15.879 great that they've now added Capture support, 00:02:16.199 --> 00:02:18.939 which is there to help block malicious automated 00:02:18.939 --> 00:02:24.060 requests. Next item is in public preview. We 00:02:24.060 --> 00:02:27.460 have inbound private endpoint support for standard 00:02:27.460 --> 00:02:31.500 version 2 tier of API management. I mean, I've 00:02:31.500 --> 00:02:34.000 talked about this many times, but there are certain 00:02:34.000 --> 00:02:35.860 things that are happening across Microsoft, things 00:02:35.860 --> 00:02:38.250 like customer managed key support versus system 00:02:38.250 --> 00:02:41.590 managed keys, managed identities versus usernames 00:02:41.590 --> 00:02:44.270 and passwords. Another big one is better support 00:02:44.270 --> 00:02:47.229 for private endpoints. So it's great to see that 00:02:47.229 --> 00:02:50.590 API management is now supporting that for standard 00:02:50.590 --> 00:02:53.689 V2. Another one which is kind of a little bit 00:02:53.689 --> 00:02:55.069 what we're going to talk about for the rest of 00:02:55.069 --> 00:02:57.990 the day is Azure Firewall is now integrated with 00:02:57.990 --> 00:03:01.240 Microsoft Security Copilot. I'm sure that Ryan 00:03:01.240 --> 00:03:02.860 will explain this in a little bit more detail, 00:03:03.139 --> 00:03:04.759 but this is something we're certainly seeing 00:03:04.759 --> 00:03:06.240 across the board. I probably should add to the 00:03:06.240 --> 00:03:09.699 list, managed identities, CMK, and private endpoints. 00:03:10.039 --> 00:03:12.219 We're seeing more support for Copilot across 00:03:12.219 --> 00:03:16.139 various security workloads. Next one is that 00:03:16.139 --> 00:03:20.400 the April 2025 progress report for our Microsoft 00:03:20.400 --> 00:03:24.080 Secure Future Initiative is now available. It's 00:03:24.080 --> 00:03:26.800 not only available as a document, it's also available 00:03:26.800 --> 00:03:29.020 as a recording as well. So you can listen to 00:03:29.020 --> 00:03:31.430 it for those who, who have vision impairments 00:03:31.430 --> 00:03:33.530 or prefer to listen to it, it's available for 00:03:33.530 --> 00:03:36.110 you to listen to as well. And the last item, 00:03:36.169 --> 00:03:39.189 before we finally get to Ryan, there's now versionless 00:03:39.189 --> 00:03:41.569 customer managed key support in Azure Database 00:03:41.569 --> 00:03:44.090 for PostgreSQL. There's a blog post written by 00:03:44.090 --> 00:03:46.210 my old colleague and actually co -author as well, 00:03:46.270 --> 00:03:49.150 Heinrich Gantenbein, that explains the rationale 00:03:49.150 --> 00:03:52.330 why it was done and how that is supported in 00:03:52.330 --> 00:03:54.409 Azure Database for PostgreSQL. So again, here's 00:03:54.409 --> 00:03:57.370 another example of a product using customer managed 00:03:57.370 --> 00:03:59.909 keys. All right, so that's the news out the way. 00:04:00.569 --> 00:04:03.270 Now let's turn our attention to our guest. As 00:04:03.270 --> 00:04:05.469 I mentioned at the top, our guest this week is 00:04:05.469 --> 00:04:08.090 Ryan Munch. Ryan, we'd like to take a moment 00:04:08.090 --> 00:04:10.909 and introduce yourself to our listeners. Thanks 00:04:10.909 --> 00:04:12.590 for having me on, Michael. It's great to be back. 00:04:13.000 --> 00:04:16.220 I'm part of the Microsoft Security Copilot product 00:04:16.220 --> 00:04:19.100 team. So in the team, I'm focused on a few different 00:04:19.100 --> 00:04:21.740 things. One, growing out our partner ecosystem, 00:04:22.040 --> 00:04:24.600 specifically our MSSPs and how they think about 00:04:24.600 --> 00:04:27.279 building agents. And then a myriad of different 00:04:27.279 --> 00:04:30.500 functionalities, such as logging within Security 00:04:30.500 --> 00:04:34.120 Copilot and agents. All right, so let's get stuck 00:04:34.120 --> 00:04:36.639 into the agenda. Look, so I mean, you hear sort 00:04:36.639 --> 00:04:39.019 of Security Copilot or the term Security Copilot 00:04:39.019 --> 00:04:42.660 kind of thrown around almost with abandon. So 00:04:42.660 --> 00:04:45.300 what is going on with security co -pilot and 00:04:45.300 --> 00:04:47.319 agents, which is what we're here to talk about? 00:04:47.560 --> 00:04:50.360 So what should our listeners know about? What's 00:04:50.360 --> 00:04:52.100 the elevator pitch? If you were to walk out of 00:04:52.100 --> 00:04:54.079 an elevator after talking to somebody, what would 00:04:54.079 --> 00:04:56.000 they know about security co -pilot and agents? 00:05:00.430 --> 00:05:03.129 We have moved fast in the past year and a half. 00:05:03.230 --> 00:05:06.069 We were the first solution out to market. We 00:05:06.069 --> 00:05:08.610 have learned a lot, and we have now evolved into 00:05:08.610 --> 00:05:11.410 what I think you see typical across a lot of 00:05:11.410 --> 00:05:13.550 different Microsoft AI solutions, and that is 00:05:13.550 --> 00:05:18.009 we are now an agent -based AI platform. All right, 00:05:18.069 --> 00:05:20.970 so a really, really dumb question. When would 00:05:20.970 --> 00:05:23.329 people use this? Why would they use agents? There 00:05:23.329 --> 00:05:26.399 are examples of where it's being used. I really 00:05:26.399 --> 00:05:28.819 want to keep it super practical. What can people 00:05:28.819 --> 00:05:30.839 walk away with understanding how it can actually 00:05:30.839 --> 00:05:34.860 help them? Yeah, and we've seen an incredible 00:05:34.860 --> 00:05:37.600 journey of our customers over the past year and 00:05:37.600 --> 00:05:40.279 a half since our private preview and even our 00:05:40.279 --> 00:05:44.339 launch. And it started with customers going in 00:05:44.339 --> 00:05:46.620 and looking at things like scripts and email 00:05:46.620 --> 00:05:49.100 headers and trying to see what AI could do for 00:05:49.100 --> 00:05:51.939 them in a security context. We started off working 00:05:51.939 --> 00:05:55.379 pretty broadly. And where it's evolved and what 00:05:55.379 --> 00:05:58.620 we've seen from our customers is that they are 00:05:58.620 --> 00:06:00.920 still workflow and very much task -oriented. 00:06:01.540 --> 00:06:03.899 And where we can help the most and where agents 00:06:03.899 --> 00:06:07.420 help the most is, we believe, with specific tasks 00:06:07.420 --> 00:06:10.600 that are either mundane, repetitive in nature, 00:06:10.660 --> 00:06:13.300 or take a lot of manual effort. Stuff that really 00:06:13.300 --> 00:06:15.899 pairs well with a human and needs a... analysts 00:06:15.899 --> 00:06:18.300 in the loop, but stuff that you don't necessarily 00:06:18.300 --> 00:06:22.339 want to do. So a good example of this is user 00:06:22.339 --> 00:06:25.160 -submitted phishing email triage. That's one 00:06:25.160 --> 00:06:28.060 good case. Vulnerability analysis with whether 00:06:28.060 --> 00:06:30.220 or not you patched a system or patched different 00:06:30.220 --> 00:06:33.720 applications. That's another practical and pragmatic 00:06:33.720 --> 00:06:37.100 use case. The other one that I really like as 00:06:37.100 --> 00:06:39.120 well, it kind of gets into more of something 00:06:39.120 --> 00:06:41.860 you don't always need, but something that a lot 00:06:41.860 --> 00:06:44.649 of security practitioners maybe aren't necessarily 00:06:44.649 --> 00:06:46.990 as familiar with as those that live it and breathe 00:06:46.990 --> 00:06:49.790 it every day, which is threat intelligence. And 00:06:49.790 --> 00:06:52.509 knowing what a threat actor is, how they operate, 00:06:52.670 --> 00:06:55.709 or how they specifically would attack your organization 00:06:55.709 --> 00:06:58.490 isn't something that everyone is always up to 00:06:58.490 --> 00:07:01.709 date on. So that becomes something that can empower 00:07:01.709 --> 00:07:04.250 people in a way that would otherwise take a lot 00:07:04.250 --> 00:07:06.170 of time. I mean, there are lots of tools that 00:07:06.170 --> 00:07:11.310 people use outside of the Microsoft. Are there 00:07:11.310 --> 00:07:14.029 integrations with other services that people 00:07:14.029 --> 00:07:18.089 may use on a daily basis? Yeah, and this is a 00:07:18.089 --> 00:07:20.509 great question. And candidly, it's one of my 00:07:20.509 --> 00:07:23.649 favorites because I think it represents a philosophical 00:07:23.649 --> 00:07:26.850 shift of how people perceive what Microsoft is 00:07:26.850 --> 00:07:30.449 doing, but also reflects a lot of what we passionately 00:07:30.449 --> 00:07:32.910 believe that security should do and how it should 00:07:32.910 --> 00:07:36.779 operate across any ecosystem. When we look at 00:07:36.779 --> 00:07:38.899 what we're trying to achieve with Security Copilot, 00:07:38.920 --> 00:07:42.519 we recognize that a workflow could exist in any 00:07:42.519 --> 00:07:44.740 state. You could be starting it. You could be 00:07:44.740 --> 00:07:46.740 beginning it. You could need to augment it anywhere 00:07:46.740 --> 00:07:50.139 in between. And a workflow could persist across 00:07:50.139 --> 00:07:52.879 any number of different systems, and it could 00:07:52.879 --> 00:07:56.060 exist exclusively outside of a Microsoft system. 00:07:56.639 --> 00:07:59.439 So what this means is that we have to build out 00:07:59.439 --> 00:08:03.560 a solution, a platform that accounts for all 00:08:03.560 --> 00:08:06.430 of that and provides for that possibility. Otherwise, 00:08:06.589 --> 00:08:10.569 as a good steward of a solution provider in the 00:08:10.569 --> 00:08:13.430 security community, if we're providing a solution 00:08:13.430 --> 00:08:16.589 that doesn't allow that expansion, we're creating 00:08:16.589 --> 00:08:18.509 another security gap. We're creating a blind 00:08:18.509 --> 00:08:22.110 spot. So we are creating and working with agents 00:08:22.110 --> 00:08:24.649 from the mindset that this could tie into any 00:08:24.649 --> 00:08:27.730 solution in any workflow state. So this ties 00:08:27.730 --> 00:08:31.230 into the exact philosophy that we're trying to 00:08:31.230 --> 00:08:34.870 part. There's a lot of significant news we've 00:08:34.870 --> 00:08:37.129 announced over the past few months, both at Securin 00:08:37.129 --> 00:08:41.269 and as well as at RSA. But for me, I think ServiceNow 00:08:41.269 --> 00:08:44.269 was a bit of the most significant bits because 00:08:45.019 --> 00:08:47.299 it backs up exactly what we've been saying in 00:08:47.299 --> 00:08:49.620 that we have to work with partners and we have 00:08:49.620 --> 00:08:51.980 to meet users where they are and however they 00:08:51.980 --> 00:08:54.860 work. So with ServiceNow, they released a plugin, 00:08:54.940 --> 00:08:57.139 but more significantly, they also integrated 00:08:57.139 --> 00:08:59.940 Security Co -Pilot into their security operations 00:08:59.940 --> 00:09:02.580 center. And that's now something that their customers 00:09:02.580 --> 00:09:05.120 in a private preview can go out and enable and 00:09:05.120 --> 00:09:07.759 begin to work with. And what this reflects and 00:09:07.759 --> 00:09:10.460 what you'll start to see across more and more 00:09:10.460 --> 00:09:15.230 different platforms and Microsoft partners, is 00:09:15.230 --> 00:09:16.950 you will see security co -pilot live in more 00:09:16.950 --> 00:09:19.309 interfaces. You will see agents that are more 00:09:19.309 --> 00:09:21.929 accessible. And you will see agents that then 00:09:21.929 --> 00:09:24.710 are able to work with you in whatever workflow 00:09:24.710 --> 00:09:27.289 you are tasked with. And so it becomes something 00:09:27.289 --> 00:09:29.830 that's hands -on and right there for you. Now, 00:09:29.830 --> 00:09:33.490 specifically for ServiceNow, what this allows 00:09:33.490 --> 00:09:36.710 is kind of exactly what you'd think. Immediate 00:09:36.710 --> 00:09:40.360 incident triage. and ability to then advance 00:09:40.360 --> 00:09:43.000 it using the power of the security copilot platform. 00:09:43.379 --> 00:09:46.299 So sort of keeping with that topic, I mean, why 00:09:46.299 --> 00:09:49.919 is it better to use agents than what we've been 00:09:49.919 --> 00:09:53.740 using before? And that sounds really silly. Can 00:09:53.740 --> 00:09:55.399 you give us some examples of what we were using 00:09:55.399 --> 00:09:58.320 before? I mean, what sort of things can using 00:09:58.320 --> 00:10:00.580 agents improve over what people are currently 00:10:00.580 --> 00:10:04.940 using? Great question that you raised there, 00:10:04.960 --> 00:10:06.960 Michael, for a few different reasons. I think 00:10:06.960 --> 00:10:09.600 it splits into two facets for me. One is like 00:10:09.600 --> 00:10:12.159 the journey of AI, and two, it's the traditional 00:10:12.159 --> 00:10:15.019 journey of how you solve solutions in a security 00:10:15.019 --> 00:10:17.700 context. So if you think about the AI journey 00:10:17.700 --> 00:10:20.340 and where Security Copilot started, we launched 00:10:20.340 --> 00:10:24.659 a chat GPT -style interface. And we did so building 00:10:24.659 --> 00:10:27.399 on top of what now everyone refers to as a compound 00:10:27.399 --> 00:10:30.080 AI system before people even really knew what 00:10:30.080 --> 00:10:33.620 that was. We built a stateful AI. We had session 00:10:33.620 --> 00:10:37.200 memory. And we worked with multiple LLMs based 00:10:37.200 --> 00:10:39.860 on what we decided was relevant for a particular 00:10:39.860 --> 00:10:43.820 task that needed to be done. At the core of this, 00:10:43.879 --> 00:10:46.279 we built what was called an orchestrator. In 00:10:46.279 --> 00:10:48.860 a lot of ways, it was a precursor to what agents 00:10:48.860 --> 00:10:52.279 are today. And agents effectively are a means 00:10:52.279 --> 00:10:55.419 to decide what system to use against the task 00:10:55.419 --> 00:10:58.879 at hand and to provide the end -to -end solutioning, 00:10:58.879 --> 00:11:01.100 if you will, of whatever you're trying to achieve 00:11:01.100 --> 00:11:04.610 or what the agent is designed to achieve. And 00:11:04.610 --> 00:11:07.830 so with agents, in a lot of ways, it represents 00:11:07.830 --> 00:11:11.330 an evolution of what we started with, with Security 00:11:11.330 --> 00:11:13.950 Co -Pilot and the Orchestrator, to probably what 00:11:13.950 --> 00:11:17.090 is something that is more robust, more evolved, 00:11:17.230 --> 00:11:20.889 and has a greater industry support. Then if you 00:11:20.889 --> 00:11:23.470 pair that with, well, should I build an agent? 00:11:23.509 --> 00:11:26.789 And should I work with agents in a security context? 00:11:27.370 --> 00:11:30.889 And I would say, and this is a cop -out, I get 00:11:30.889 --> 00:11:34.529 it, I would say it depends. It depends on where 00:11:34.529 --> 00:11:38.129 you can insert the value of either non -deterministic 00:11:38.129 --> 00:11:42.169 solutions, meaning agents can work really well 00:11:42.169 --> 00:11:44.649 in a way where it needs to consider different 00:11:44.649 --> 00:11:47.190 inputs to provide the ultimate goal and ultimate 00:11:47.190 --> 00:11:50.029 output. So if you are looking across, let's just 00:11:50.029 --> 00:11:53.210 say, a phishing email triage scenario where you 00:11:53.210 --> 00:11:54.870 maybe need to look at a handful of different 00:11:54.870 --> 00:11:56.830 indicators, the indicator types are different, 00:11:56.870 --> 00:11:59.450 and you need to consider, we'll call it N number 00:11:59.450 --> 00:12:01.990 of different sources based on the type of indicator. 00:12:02.720 --> 00:12:05.320 Well, if you were to build one standard agent 00:12:05.320 --> 00:12:07.519 and you were to deploy that against 50 different 00:12:07.519 --> 00:12:10.460 customers, or let's say you were a security provider 00:12:10.460 --> 00:12:14.799 and you're working with all your different customer 00:12:14.799 --> 00:12:17.080 environments, that could mean that you're working 00:12:17.080 --> 00:12:19.000 with any number of different threat intelligence 00:12:19.000 --> 00:12:21.580 sources, any number of different vulnerability 00:12:21.580 --> 00:12:26.000 scanners, any number of different endpoint systems 00:12:26.000 --> 00:12:29.289 that would need to perform the analysis of. does 00:12:29.289 --> 00:12:32.029 this phishing email impact my environment? And 00:12:32.029 --> 00:12:35.429 if so, how? And what an agent can do in that 00:12:35.429 --> 00:12:37.850 case is provide the means in which it can decide 00:12:37.850 --> 00:12:41.570 what is available in this deployment, what should 00:12:41.570 --> 00:12:44.269 I use to solve and triage this phishing email, 00:12:44.429 --> 00:12:46.610 and then provide an output for human review. 00:12:47.370 --> 00:12:51.990 Now, if you contrast that to what you have currently 00:12:51.990 --> 00:12:54.980 with email triage, You're going to have maybe 00:12:54.980 --> 00:12:56.879 some type of script or a SOAR that's going to 00:12:56.879 --> 00:12:59.039 go through, pull out those indicators, and enrich 00:12:59.039 --> 00:13:01.659 them. But you're going to build that very specifically 00:13:01.659 --> 00:13:04.379 for maybe one customer, or maybe you provide 00:13:04.379 --> 00:13:06.840 enough branching logic where you cover all those 00:13:06.840 --> 00:13:10.019 possibilities. The point is that it becomes something 00:13:10.019 --> 00:13:14.179 that's static in nature and also is difficult 00:13:14.179 --> 00:13:18.179 or more challenging to maintain over time. The 00:13:18.179 --> 00:13:20.700 other part too is when I look at SOARs and when 00:13:20.700 --> 00:13:23.200 I think about tips, they generally fall into 00:13:23.200 --> 00:13:26.519 the category for me of automate it and forget 00:13:26.519 --> 00:13:28.580 it. You're scripting it and you have an input 00:13:28.580 --> 00:13:31.580 and output. Agents aren't really built like that. 00:13:31.639 --> 00:13:34.080 They are built more from the consideration of 00:13:34.080 --> 00:13:37.799 based on how we have gone to solve this particular 00:13:37.799 --> 00:13:40.840 task, we want your input to validate the solution. 00:13:41.120 --> 00:13:44.049 And that becomes... That becomes a means through 00:13:44.049 --> 00:13:46.750 which it takes some of the mundane analysis, 00:13:46.909 --> 00:13:50.029 like of a user -submitted phish, automates the 00:13:50.029 --> 00:13:53.009 process, and allows for human input. The final 00:13:53.009 --> 00:13:54.970 thing that I'd tack on to that scenario, I think 00:13:54.970 --> 00:13:58.610 that is also different from a traditional automation 00:13:58.610 --> 00:14:01.740 versus what agents can do. is at any point along 00:14:01.740 --> 00:14:04.159 the way they can invoke LLMs in a meaningful 00:14:04.159 --> 00:14:07.000 way to solve a solution. So the other reason 00:14:07.000 --> 00:14:09.980 why an email is a really good articulation of 00:14:09.980 --> 00:14:13.799 this emerging and evolution of agents in a security 00:14:13.799 --> 00:14:17.019 context is that LLMs can look at the context 00:14:17.019 --> 00:14:20.299 of an email and decide the semantics of the language 00:14:20.299 --> 00:14:23.940 used. Is this business venture from Nigeria something 00:14:23.940 --> 00:14:26.500 I should invest in? Can I really trust this prince? 00:14:27.289 --> 00:14:28.850 Those are the things that it's really good at 00:14:28.850 --> 00:14:31.429 pointing out and saying, you know what? This 00:14:31.429 --> 00:14:33.909 is a phishing email, even though I don't have 00:14:33.909 --> 00:14:37.190 any indicators that match. Based on the other 00:14:37.190 --> 00:14:39.309 attributes that don't match any known indicators, 00:14:39.669 --> 00:14:42.350 we can classify this email as a phishing email. 00:14:42.529 --> 00:14:45.730 And then in that moment, what you have that what 00:14:45.730 --> 00:14:47.990 you wouldn't have before in a traditional automation 00:14:47.990 --> 00:14:52.509 scenario is validation of a phishing email with 00:14:52.509 --> 00:14:56.820 net new indicators that you can then keep. store 00:14:56.820 --> 00:14:59.120 for any other future analysis. And that becomes 00:14:59.120 --> 00:15:01.019 something that's meaningful and impactful in 00:15:01.019 --> 00:15:04.159 a security context. You mentioned before you 00:15:04.159 --> 00:15:07.320 used the word compound. Does that mean you use 00:15:07.320 --> 00:15:09.580 multiple LLMs? Is that the origin of the term? 00:15:09.639 --> 00:15:11.620 Is that what's inferred by the word compound 00:15:11.620 --> 00:15:13.860 or is it something different than that? Yeah, 00:15:13.919 --> 00:15:17.039 great question. So there's a paper that was put 00:15:17.039 --> 00:15:20.679 out by the Berkeley AI Research Center or BEAR. 00:15:21.129 --> 00:15:23.149 And I probably just killed the acronym, but if 00:15:23.149 --> 00:15:25.169 you Bing that, you'll get right there. So they 00:15:25.169 --> 00:15:28.190 have this great paper that details a compound 00:15:28.190 --> 00:15:30.870 AI system if you want to read the full academic 00:15:30.870 --> 00:15:33.629 breakdown. But if you distill it down simply 00:15:33.629 --> 00:15:36.710 to what it represents, when AI systems were first 00:15:36.710 --> 00:15:38.970 built, they were monolithic in nature. Let's 00:15:38.970 --> 00:15:43.230 train an LLM, let's set that loose on a particular 00:15:43.230 --> 00:15:47.279 problem, and let's see what it can do. What they 00:15:47.279 --> 00:15:49.759 found out is that in the edge cases, especially 00:15:49.759 --> 00:15:52.679 in security or in things like healthcare or other 00:15:52.679 --> 00:15:54.440 industries that have a lot of specialization 00:15:54.440 --> 00:15:58.700 and a lot of unique language or context that's 00:15:58.700 --> 00:16:02.889 used, LLMs don't really work. And they don't 00:16:02.889 --> 00:16:05.590 work because they miss that information. So compound 00:16:05.590 --> 00:16:08.090 AI systems started to address that problem by 00:16:08.090 --> 00:16:11.330 integrating or starting to establish more traditional 00:16:11.330 --> 00:16:14.250 paradigms of system architecture around the LLM. 00:16:14.629 --> 00:16:18.470 So you have to find a way to either take an LLM 00:16:18.470 --> 00:16:21.009 and source it with data. So that's what's called 00:16:21.009 --> 00:16:23.710 retrieval augmented generation, where a data 00:16:23.710 --> 00:16:26.529 store can be used to then infuse the result or 00:16:26.529 --> 00:16:28.889 produce a better result for the LLM, where the 00:16:28.889 --> 00:16:31.610 LLM can not only... rationalize the information 00:16:31.610 --> 00:16:34.629 being retrieved, match it against the context, 00:16:34.830 --> 00:16:37.070 and then provide an output. And so that's what, 00:16:37.129 --> 00:16:39.169 for Security Copilot, that's why we created the 00:16:39.169 --> 00:16:41.730 session. And for everyone's own edification, 00:16:42.090 --> 00:16:44.549 that's also why Security Copilot does things 00:16:44.549 --> 00:16:46.950 like includes all of Microsoft's threat intelligence 00:16:46.950 --> 00:16:49.710 as part of the core product, because we believe 00:16:49.710 --> 00:16:51.690 that that should be part of what's included in 00:16:51.690 --> 00:16:54.549 any solution. The other thing that it refers 00:16:54.549 --> 00:16:57.799 to is that... In a compound AI system, you're 00:16:57.799 --> 00:17:00.120 right. You can use multiple LLMs, and you can 00:17:00.120 --> 00:17:03.019 be specialized in how you invoke an LLM based 00:17:03.019 --> 00:17:06.019 on the need, the context, or otherwise. It gives 00:17:06.019 --> 00:17:08.380 some flexibilities in terms of the architecture 00:17:08.380 --> 00:17:12.680 to use a smaller LLM in size and training data 00:17:12.680 --> 00:17:17.099 or specialization, ultimately. Or in some cases 00:17:17.099 --> 00:17:19.940 where you maybe need to process more data. have 00:17:19.940 --> 00:17:22.700 a greater token limit, which can be computationally 00:17:22.700 --> 00:17:25.039 more expensive and take longer to produce a result, 00:17:25.200 --> 00:17:28.019 you can do so. Or if you just need to do a simple 00:17:28.019 --> 00:17:31.079 task and you can use a base model and do that 00:17:31.079 --> 00:17:33.480 faster and more efficiently. So it gives greater 00:17:33.480 --> 00:17:36.119 flexibility for what you'd expect. And that is 00:17:36.119 --> 00:17:39.240 addressing the task at hand in the most relevant 00:17:39.240 --> 00:17:42.599 way and efficient way possible. Yeah, I'd seen 00:17:42.599 --> 00:17:44.579 the term compound thrown around and I just wasn't 00:17:44.579 --> 00:17:49.039 sure of the context. So, I mean, at Microsoft, 00:17:49.220 --> 00:17:50.720 look, I'm going to be honest, I think at Microsoft 00:17:50.720 --> 00:17:53.400 we have a lot of AI solutions right now. And 00:17:53.400 --> 00:17:55.660 honestly, for me, it seems a little confusing 00:17:55.660 --> 00:18:01.259 as to what does what and so on. So why are agents 00:18:01.259 --> 00:18:05.480 different than, say, using Copilot Studio? I 00:18:05.480 --> 00:18:08.319 mean, do I use Copilot Studio to create these? 00:18:08.500 --> 00:18:10.740 I mean, I don't understand the relationship between 00:18:10.740 --> 00:18:15.289 security agents or security Copilot agents. and 00:18:15.289 --> 00:18:17.809 all the other sort of AI offerings that we have 00:18:17.809 --> 00:18:21.190 at Microsoft. Can you spend just a few minutes 00:18:21.190 --> 00:18:23.750 and just sort of compare and contrast? Yeah, 00:18:23.769 --> 00:18:26.170 good news, Michael. You're not the only one that 00:18:26.170 --> 00:18:29.150 has this same confusion. I was just at RSA, and 00:18:29.150 --> 00:18:31.730 I was in a room full of CISOs from Spain, and 00:18:31.730 --> 00:18:35.569 they all had similar questions in terms of what 00:18:35.569 --> 00:18:37.829 is this agent thing from security co -pilot, 00:18:37.869 --> 00:18:39.549 and how is it different than all these other 00:18:39.549 --> 00:18:42.009 co -pilot solutions? And I get it. Microsoft 00:18:42.009 --> 00:18:45.299 at this point probably has a... a full air show 00:18:45.299 --> 00:18:49.420 squadron full of co -pilots. But we have a general 00:18:49.420 --> 00:18:53.359 philosophy across Microsoft to learn, to innovate, 00:18:53.519 --> 00:18:56.019 and report back across no matter what engineering 00:18:56.019 --> 00:18:58.720 group we are to see what we can do to advance 00:18:58.720 --> 00:19:01.680 AI solutions. And so what this means is that 00:19:01.680 --> 00:19:04.619 in Microsoft, we have different solutions that 00:19:04.619 --> 00:19:07.599 are... built for a specific purpose. So for example, 00:19:07.680 --> 00:19:10.819 if you wanted to build an AI service yourself 00:19:10.819 --> 00:19:13.299 or an AI solution yourself, you would work with 00:19:13.299 --> 00:19:15.480 one of our AI services, whether it's Azure AI 00:19:15.480 --> 00:19:18.720 Foundry, which deals in AI primitives, or Azure 00:19:18.720 --> 00:19:20.819 Open AI Service if you just want to work maybe 00:19:20.819 --> 00:19:23.779 with some off -the -shelf models. That's there 00:19:23.779 --> 00:19:25.960 if you have the developer mindset and are looking 00:19:25.960 --> 00:19:28.500 to build something on your own. If you look to 00:19:28.500 --> 00:19:30.710 Copilot Studio, It's a way in which you have, 00:19:30.730 --> 00:19:32.809 like, this sea of agents and all these different 00:19:32.809 --> 00:19:35.769 business logic solutions that you can use. But 00:19:35.769 --> 00:19:38.549 it's generally designed to operate in a business 00:19:38.549 --> 00:19:42.230 context for office -related responsibilities. 00:19:43.250 --> 00:19:45.710 Now, you can do things, and this is one of the 00:19:45.710 --> 00:19:48.109 interesting things I like to point out. Like, 00:19:48.170 --> 00:19:52.390 you can, in Copilot Studio, you can call back 00:19:52.390 --> 00:19:55.470 to Security Copilot and invoke a prompt and retrieve 00:19:55.470 --> 00:19:58.779 that prompt response. And for me, that's significant 00:19:58.779 --> 00:20:01.579 because it underscores the point. The reason 00:20:01.579 --> 00:20:04.859 why Security Copilot is its own platform is that 00:20:04.859 --> 00:20:07.559 we have to build agents in a security context 00:20:07.559 --> 00:20:11.299 that can operate across an enterprise in a way 00:20:11.299 --> 00:20:15.779 that demands a high level of workload capability, 00:20:16.259 --> 00:20:20.680 reliability, but also security. And that means 00:20:20.680 --> 00:20:22.380 we have to build it a little bit differently 00:20:22.380 --> 00:20:25.960 to handle all those specific use cases. Copilot 00:20:25.960 --> 00:20:28.039 Studio, if you're going out to summarize a document 00:20:28.039 --> 00:20:30.900 or summarize a Teams chat, you're usually operating 00:20:30.900 --> 00:20:34.460 in a user context that's like one or a few. But 00:20:34.460 --> 00:20:37.819 if you're looking at, for example, find me all 00:20:37.819 --> 00:20:40.259 these user -submitted phishing emails and triage 00:20:40.259 --> 00:20:42.220 them, well, you can be talking about anywhere 00:20:42.220 --> 00:20:44.880 from one email to potentially a few hundred or 00:20:44.880 --> 00:20:47.200 thousand. And that becomes a different way in 00:20:47.200 --> 00:20:50.130 which we have to think about architecture. It's 00:20:50.130 --> 00:20:51.589 interesting you should bring up that you're talking 00:20:51.589 --> 00:20:54.809 to people at RSA. I see that at RSA we announced 00:20:54.809 --> 00:20:59.029 a book of news that talked about partner agents 00:20:59.029 --> 00:21:01.430 and so on. Could you sort of explain that in 00:21:01.430 --> 00:21:04.019 a bit more detail, like what was announced? Yeah, 00:21:04.059 --> 00:21:06.519 at RSA, we announced a few things. We followed 00:21:06.519 --> 00:21:10.079 up on Microsoft Secure and highlighted the partner 00:21:10.079 --> 00:21:12.539 agents across Microsoft that we've built. So, 00:21:12.579 --> 00:21:15.339 for example, there's a conditional access agent 00:21:15.339 --> 00:21:17.500 that exists for Microsoft Intra where you can 00:21:17.500 --> 00:21:19.980 activate in the Intra portal. There's a threat 00:21:19.980 --> 00:21:22.720 intel briefing agent, a vulnerability remediation 00:21:22.720 --> 00:21:25.440 agent from Intune. And then I talked about a 00:21:25.440 --> 00:21:27.440 little bit of the phishing use case. Well, sure 00:21:27.440 --> 00:21:30.420 enough, the Defender team saw the value, too, 00:21:30.440 --> 00:21:32.670 and they built out their agent. to do exactly 00:21:32.670 --> 00:21:36.069 that. And there's also an alert triage agent 00:21:36.069 --> 00:21:39.670 for Microsoft Purview. Now, on the other end, 00:21:39.710 --> 00:21:42.309 we have a ton of stuff that was released by partners. 00:21:42.750 --> 00:21:45.750 We have initial five that came out from both 00:21:45.750 --> 00:21:48.930 across what we call our partner developers as 00:21:48.930 --> 00:21:52.720 well as our MSSPs. We had seven agents announced 00:21:52.720 --> 00:21:55.160 there, and we had a few different plugins as 00:21:55.160 --> 00:21:57.720 well. In addition to ServiceNow, we had Splunk 00:21:57.720 --> 00:22:01.619 and HP and Census all provide some plugins. And 00:22:01.619 --> 00:22:04.640 I'm going to quickly differentiate or call out 00:22:04.640 --> 00:22:07.000 plugins are what we used initially with Security 00:22:07.000 --> 00:22:11.039 Copilot to connect to systems. Plugins are now 00:22:11.039 --> 00:22:13.579 something that we can use in agents and use in 00:22:13.579 --> 00:22:17.200 an agent runtime. So it's all complementary architecture 00:22:17.200 --> 00:22:20.559 and complementary solutions. But the significance 00:22:20.559 --> 00:22:23.099 of this news, in addition to what I talked about 00:22:23.099 --> 00:22:27.220 earlier of ServiceNow, is that it shows the depth 00:22:27.220 --> 00:22:30.319 of the ecosystem that's emerging and our desire 00:22:30.319 --> 00:22:32.920 to continue to expand across solutions that are 00:22:32.920 --> 00:22:35.119 most impactful and relevant for our customers 00:22:35.119 --> 00:22:37.559 to solve real -world problems that they deal 00:22:37.559 --> 00:22:39.880 with every day. Okay, let's just bring it back 00:22:39.880 --> 00:22:43.539 to sort of my neck of the woods, which is development. 00:22:45.189 --> 00:22:48.990 So if I want to use one of these security copilot 00:22:48.990 --> 00:22:53.230 agents in my code, what do I need to do? Is this 00:22:53.230 --> 00:22:55.309 something that developers would do? Or is there 00:22:55.309 --> 00:22:57.630 like some drag and drop interface that a non 00:22:57.630 --> 00:23:02.109 -developer can do? Like the current sort of coding 00:23:02.109 --> 00:23:04.049 du jour, you know, vibe coding? I mean, is that 00:23:04.049 --> 00:23:06.670 a possibility? Or is this really something from 00:23:06.670 --> 00:23:08.490 a development, like from a producing something? 00:23:08.690 --> 00:23:10.349 Is this something that a developer would have 00:23:10.349 --> 00:23:12.690 to do writing, say, code in, say, C Sharp or 00:23:12.690 --> 00:23:15.059 Python or something like that? Well, I think 00:23:15.059 --> 00:23:16.799 vibe coding is generally something you could 00:23:16.799 --> 00:23:19.019 do at any time in any interface. I think you 00:23:19.019 --> 00:23:22.359 just need a beer and a lot of will to do whatever. 00:23:23.640 --> 00:23:26.059 Now, for what we're going to do with Security 00:23:26.059 --> 00:23:29.299 Copilot, this is where I got to be careful here 00:23:29.299 --> 00:23:31.839 not to get ahead of some of our news that's coming 00:23:31.839 --> 00:23:35.180 up at Build and other events. But what I'd stress 00:23:35.180 --> 00:23:39.480 is that in our recognition that we are heading 00:23:39.480 --> 00:23:42.660 down the direction of being a... agentic platform 00:23:42.660 --> 00:23:46.400 for a security context, we need to provide different 00:23:46.400 --> 00:23:49.599 ways for the Michaels of the world to go out 00:23:49.599 --> 00:23:52.640 and develop agents effectively. So we've had 00:23:52.640 --> 00:23:55.579 conversations against low code and no code and 00:23:55.579 --> 00:24:00.819 we'll call it, I guess, hardcore coding experiences 00:24:00.819 --> 00:24:03.200 and how you would go out and build agents and 00:24:03.200 --> 00:24:06.420 build agents effectively. Today, what you'd see 00:24:06.420 --> 00:24:08.799 is like at minimum, you could still work with 00:24:08.799 --> 00:24:12.140 a manifest file like you do with plugins. But 00:24:12.140 --> 00:24:14.619 we want to make some of that stuff more accessible 00:24:14.619 --> 00:24:18.160 in ways that people would then extend the security 00:24:18.160 --> 00:24:20.980 copilot and build and bring agent functionality 00:24:20.980 --> 00:24:24.180 more readily in ways that empower them to work 00:24:24.180 --> 00:24:27.190 with their preferred tools. All right, so Ryan, 00:24:27.430 --> 00:24:28.970 I think it's probably time to start bringing 00:24:28.970 --> 00:24:32.309 this episode to an end. One thing we ask all 00:24:32.309 --> 00:24:36.109 our guests is what does a typical day in Ryan's 00:24:36.109 --> 00:24:39.930 day look like? So day in the life of me, it's 00:24:39.930 --> 00:24:42.190 just as chaotic as what you'd think. Working 00:24:42.190 --> 00:24:45.609 in a product team that works on AI, launching 00:24:45.609 --> 00:24:49.329 agents and expanding in the way that we are means 00:24:49.329 --> 00:24:50.990 I'm pretty busy with a lot of different things. 00:24:52.220 --> 00:24:54.599 First and foremost, there's a good amount of 00:24:54.599 --> 00:24:56.940 time that I spend each day trying to stay up 00:24:56.940 --> 00:25:00.559 with what is going on with AI, how I can stay 00:25:00.559 --> 00:25:04.519 smart, remain smart, and anchor in both current 00:25:04.519 --> 00:25:06.819 conversations as well as future ones and future 00:25:06.819 --> 00:25:10.220 designs and how we continue to evolve our AI 00:25:10.220 --> 00:25:13.019 solution and continue to meet the needs of our 00:25:13.019 --> 00:25:15.619 partners and customers. The other thing I spend 00:25:15.619 --> 00:25:18.299 a lot of time on in my day is talking to our 00:25:18.299 --> 00:25:21.259 partners. what they want to build, how they want 00:25:21.259 --> 00:25:23.259 to build it, how they can best service their 00:25:23.259 --> 00:25:25.519 customers so that they can build agents and think 00:25:25.519 --> 00:25:28.740 about how they want to deliver agents to their 00:25:28.740 --> 00:25:31.599 customers to empower them in practical security 00:25:31.599 --> 00:25:35.259 workflows. With agents, it's been something that 00:25:35.259 --> 00:25:38.180 is net new to them, and it represents an exciting 00:25:38.180 --> 00:25:41.220 opportunity. And so for me, as well as them, 00:25:41.359 --> 00:25:43.900 we get to talk about a lot of just general AI 00:25:43.900 --> 00:25:46.859 philosophy, and I love that. And then the final 00:25:46.859 --> 00:25:48.660 bit I think that probably the non -work stuff 00:25:48.660 --> 00:25:51.759 is a good way for me to relax and get out of 00:25:51.759 --> 00:25:54.039 the AI headspace for a bit or maybe think a little 00:25:54.039 --> 00:25:56.740 bit more clearly is I usually will end my day 00:25:56.740 --> 00:26:01.480 with a swim at an outdoor pool for a good 3 ,000 00:26:01.480 --> 00:26:04.400 yards to wrap it up. And once I do that, then 00:26:04.400 --> 00:26:06.740 I'm usually ready for a bit more AI things before 00:26:06.740 --> 00:26:08.880 I call it a day. It's funny you should mention 00:26:08.880 --> 00:26:11.059 sort of staying on top of things and constantly 00:26:11.059 --> 00:26:13.900 learning. I think that's so critically important 00:26:13.900 --> 00:26:16.769 in this. not just today, but also in the area 00:26:16.769 --> 00:26:19.910 of security and AI. And one thing that sort of 00:26:19.910 --> 00:26:21.849 really hit me the other day, just how true this 00:26:21.849 --> 00:26:26.230 is, I was in a call with a corporate VP, actually 00:26:26.230 --> 00:26:28.710 a bunch of corporate VPs, and we were explaining 00:26:28.710 --> 00:26:31.269 a vulnerability class. And this particular VP 00:26:31.269 --> 00:26:33.569 in question didn't understand all the nuances 00:26:33.569 --> 00:26:36.029 of it. It would have been very easy as a corporate 00:26:36.029 --> 00:26:38.890 VP for that person to say, hey, I know about 00:26:38.890 --> 00:26:42.259 this, but he didn't. and he ended up setting 00:26:42.259 --> 00:26:46.240 up a meeting called Make John Smart. And it was 00:26:46.240 --> 00:26:48.299 basically a 30 -minute meeting with a bunch of 00:26:48.299 --> 00:26:50.279 us to go through this vulnerability class so 00:26:50.279 --> 00:26:53.859 that he fully understood the vulnerability and 00:26:53.859 --> 00:26:56.980 all the implications of it. That, to me, is a 00:26:56.980 --> 00:26:59.099 really good example of someone who's incredibly 00:26:59.099 --> 00:27:01.759 senior in the company recognizing they don't 00:27:01.759 --> 00:27:04.319 know something, but also recognizing they kind 00:27:04.319 --> 00:27:06.579 of need to know it, too, and doing whatever it 00:27:06.579 --> 00:27:08.559 takes to make sure that they learn that thing. 00:27:10.059 --> 00:27:12.180 My hat's off to people who want to continually 00:27:12.180 --> 00:27:14.660 learn. I think that's just absolutely just so, 00:27:14.799 --> 00:27:18.059 so, so critical, especially now with AI. So one 00:27:18.059 --> 00:27:21.319 thing we always ask our guests is if you had 00:27:21.319 --> 00:27:23.519 just one final thought to leave our listeners 00:27:23.519 --> 00:27:26.259 with, what would it be? I think it's going to 00:27:26.259 --> 00:27:29.420 be the approach for how everyone... rationalizes 00:27:29.420 --> 00:27:33.559 agents and AI. Having walked the show floor of 00:27:33.559 --> 00:27:36.599 RSA just over the past week, everyone is talking 00:27:36.599 --> 00:27:39.099 about it. It's literally the talk of security 00:27:39.099 --> 00:27:42.700 and the talk of, well, our industry, the broader 00:27:42.700 --> 00:27:45.920 tech industry. So when you rationalize agents, 00:27:46.140 --> 00:27:47.680 you should think about how you're going to adopt 00:27:47.680 --> 00:27:49.640 them and what are your criteria to adopt them. 00:27:50.180 --> 00:27:52.759 And at its core of that, I would say think about 00:27:52.759 --> 00:27:55.880 agent transparency, how you can see what they 00:27:55.880 --> 00:27:58.230 decide. how you can see how they handle data, 00:27:58.390 --> 00:28:01.509 what outputs of data they provide, and probably 00:28:01.509 --> 00:28:03.490 most importantly, what they're doing to show 00:28:03.490 --> 00:28:06.630 you it's logging an activity. Through all of 00:28:06.630 --> 00:28:09.009 that, and I believe fundamentally, you should 00:28:09.009 --> 00:28:12.609 not implicitly trust a new system without getting 00:28:12.609 --> 00:28:15.490 a good handle and understanding of it. When it 00:28:15.490 --> 00:28:18.789 comes to AI, establishing a trust framework is 00:28:18.789 --> 00:28:21.630 even more important. And so as you think about 00:28:21.630 --> 00:28:23.869 adopting agents, make sure you have that trust 00:28:23.869 --> 00:28:27.019 framework in mind. Work to achieve transparency 00:28:27.019 --> 00:28:30.119 and through transparency, learn to understand 00:28:30.119 --> 00:28:32.619 how agents operate and how you can trust them 00:28:32.619 --> 00:28:35.000 or work to trust them inside of your environment. 00:28:36.039 --> 00:28:39.700 Yeah, and on that topic, make sure if you're 00:28:39.700 --> 00:28:41.039 listening, people listening to this and you're 00:28:41.039 --> 00:28:44.099 not sure kind of what Ryan's alluding to, go 00:28:44.099 --> 00:28:46.039 and listen to our previous episode on that very 00:28:46.039 --> 00:28:49.880 topic about trusting agents, you know, because 00:28:49.880 --> 00:28:52.180 the things that you do need to be aware of. All 00:28:52.180 --> 00:28:54.829 right. So let's bring this episode to an end. 00:28:54.910 --> 00:28:56.829 Orion, thank you so much for joining us this 00:28:56.829 --> 00:28:59.430 week. I learned a ton. I learned a ton on all 00:28:59.430 --> 00:29:01.509 the episodes, but this one very, very much so. 00:29:02.170 --> 00:29:05.170 I think that the intersection of security and 00:29:05.170 --> 00:29:11.210 agents is just mind -blowing. I think the possibilities, 00:29:11.529 --> 00:29:14.210 especially when we're dealing with AI -based 00:29:14.210 --> 00:29:17.170 attacks as well, we're putting up some sort of 00:29:17.170 --> 00:29:19.529 defensive mechanisms, so that's huge as well. 00:29:19.710 --> 00:29:22.529 So again, thank you for joining us. And to all 00:29:22.529 --> 00:29:23.950 our listeners out there, we hope you found this 00:29:23.950 --> 00:29:26.309 episode of use. Stay safe, and we'll see you 00:29:26.309 --> 00:29:26.750 next time.