1 00:00:00,000 --> 00:00:09,600 Welcome to the Azure Security Podcast, where we discuss topics relating to security, privacy, 2 00:00:09,600 --> 00:00:13,280 reliability and compliance on the Microsoft Cloud Platform. 3 00:00:13,280 --> 00:00:17,720 Hey everybody, welcome to episode 97. 4 00:00:17,720 --> 00:00:20,760 This week it is myself, Michael, with Sarah. 5 00:00:20,760 --> 00:00:22,560 Everyone else is busy. 6 00:00:22,560 --> 00:00:26,200 And our guest this week is Richard Diver, who's here to talk to us about artificial 7 00:00:26,200 --> 00:00:27,600 intelligence security. 8 00:00:27,600 --> 00:00:31,640 Actually, the note that we have here says AI security stuff. 9 00:00:31,640 --> 00:00:35,160 But before we get on to our guest, let's take a little lap around the news. 10 00:00:35,160 --> 00:00:36,160 I'll kick things off. 11 00:00:36,160 --> 00:00:39,200 I actually want to start with a personal news item. 12 00:00:39,200 --> 00:00:42,720 I've actually taken a new position within Microsoft, which I'll start in the new fiscal 13 00:00:42,720 --> 00:00:45,040 year, which is July the 1st. 14 00:00:45,040 --> 00:00:49,280 So right now, as some of you probably know, I work in the Azure Data Platform. 15 00:00:49,280 --> 00:00:54,440 So I work on Azure SQL Database, SQL Server, Cosmos DB, Postgres SQL and MySQL, all from 16 00:00:54,440 --> 00:00:55,440 a security perspective. 17 00:00:55,440 --> 00:01:01,960 I'm now actually moving over to basically that sort of similar work, but for the whole 18 00:01:01,960 --> 00:01:03,200 of Azure. 19 00:01:03,200 --> 00:01:08,280 So I'll be working on engineering stuff and sort of learning from attacks and how we can 20 00:01:08,280 --> 00:01:09,280 change our processes. 21 00:01:09,280 --> 00:01:12,440 So I started that in January the 1st. 22 00:01:12,440 --> 00:01:13,440 Really excited. 23 00:01:13,440 --> 00:01:15,440 I just, you know, I love the security stuff. 24 00:01:15,440 --> 00:01:20,600 And by the way, I couldn't say anything nicer about the Azure Data team. 25 00:01:20,600 --> 00:01:22,920 It's a fantastic team of engineers. 26 00:01:22,920 --> 00:01:23,920 They really know what they're doing. 27 00:01:23,920 --> 00:01:25,600 They've learned so much about database products. 28 00:01:25,600 --> 00:01:28,280 The stuff that I didn't buy, there's still a lot of stuff I don't know. 29 00:01:28,280 --> 00:01:30,640 I don't know about databases. 30 00:01:30,640 --> 00:01:34,000 Just the magnificent database team or engineering team in general. 31 00:01:34,000 --> 00:01:37,000 So it's going to be a bit of a bit of a bittersweet to be honest with you. 32 00:01:37,000 --> 00:01:39,880 But I'm really looking forward to this new endeavor. 33 00:01:39,880 --> 00:01:47,240 So in terms of news, first one is Azure Kailh Studio now has a feature where it can pause 34 00:01:47,240 --> 00:01:49,840 processes inside of virtual machines. 35 00:01:49,840 --> 00:01:55,200 This is really useful when it comes to things like mimicking updating processes rather than 36 00:01:55,200 --> 00:01:56,800 just rebooting the VM. 37 00:01:56,800 --> 00:01:57,800 So it's good to see that. 38 00:01:57,800 --> 00:02:02,480 Just one, you know, again, just one more thing to add a little bit of chaos to your environments. 39 00:02:02,480 --> 00:02:06,600 Azure API management now supports the circuit breaker pattern. 40 00:02:06,600 --> 00:02:10,360 So there's a whole bunch of sort of design patterns that are out there. 41 00:02:10,360 --> 00:02:11,720 And one of them is a circuit breaker. 42 00:02:11,720 --> 00:02:16,560 And the idea of this is that when a front end becomes completely overloaded, rather 43 00:02:16,560 --> 00:02:22,440 than sort of failing in a horrible mess and, you know, impacting all the stuff behind the 44 00:02:22,440 --> 00:02:25,840 API management, it does so in a graceful way. 45 00:02:25,840 --> 00:02:27,840 And that's the circuit breaker pattern. 46 00:02:27,840 --> 00:02:30,960 So you can now actually enable that patterns, a whole bunch of parameters you can have in 47 00:02:30,960 --> 00:02:31,960 there. 48 00:02:31,960 --> 00:02:34,840 I'll provide a link not just to that news item, but also a link to the circuit breaker 49 00:02:34,840 --> 00:02:38,800 pattern, which is on the Azure architecture center. 50 00:02:38,800 --> 00:02:44,960 The other bit of news is Azure Bastion now has a developer skew. 51 00:02:44,960 --> 00:02:49,560 Right now, if you deploy Bastion, you have to pay for it. 52 00:02:49,560 --> 00:02:50,800 There's no sort of free version. 53 00:02:50,800 --> 00:02:56,560 So what Bastion lets you do is, I know the Bastion guys are going to hate me for saying 54 00:02:56,560 --> 00:02:58,840 this, but it's almost like a jump box. 55 00:02:58,840 --> 00:03:02,520 So you're not connecting directly to your environment. 56 00:03:02,520 --> 00:03:06,400 You're going through the Bastion and then you can put all sorts of policies around the 57 00:03:06,400 --> 00:03:08,080 Bastion. 58 00:03:08,080 --> 00:03:12,280 And that way you're connecting just through that, as opposed to potentially compromising 59 00:03:12,280 --> 00:03:13,680 back ends if there's an issue. 60 00:03:13,680 --> 00:03:17,880 So it's a really magnificent product because it's really seamless the way it works in Azure. 61 00:03:17,880 --> 00:03:19,400 But historically you had to pay for it. 62 00:03:19,400 --> 00:03:22,600 Well, now there's a developer skew, which is free of charge. 63 00:03:22,600 --> 00:03:27,020 I'm not sure about all the limitations and what have you, but it is essentially free 64 00:03:27,020 --> 00:03:28,020 of charge for developers. 65 00:03:28,020 --> 00:03:33,600 And I think it only allows you to connect to one VM, I think, but I could be wrong there. 66 00:03:33,600 --> 00:03:39,200 And the last bit of news that I have is Azure Front Door now has log scrubbing of sensitive 67 00:03:39,200 --> 00:03:41,320 data and that's in general availability. 68 00:03:41,320 --> 00:03:42,720 This is really nice. 69 00:03:42,720 --> 00:03:47,000 So if you've got logs, you can put rules in there to scrub out sensitive data. 70 00:03:47,000 --> 00:03:49,320 So I'll put like asterisks where there may be sensitive data. 71 00:03:49,320 --> 00:03:54,200 So for example, if you're logging, you may decide you don't want to log all of an IP 72 00:03:54,200 --> 00:03:58,460 address or you may not want to log parts of a URI, that kind of stuff. 73 00:03:58,460 --> 00:04:04,520 Then you can have as your front door, just essentially scrub that out of the log so it's 74 00:04:04,520 --> 00:04:05,520 not there. 75 00:04:05,520 --> 00:04:11,660 Again, this kind of stuff is just great to see because it's not uncommon for attackers 76 00:04:11,660 --> 00:04:16,680 to go after log files because they might have sensitive information in there. 77 00:04:16,680 --> 00:04:20,080 So anything you can do to get that data out of there in the first place is always a good 78 00:04:20,080 --> 00:04:21,080 thing. 79 00:04:21,080 --> 00:04:22,080 So that's my news. 80 00:04:22,080 --> 00:04:23,080 Sarah, what you got? 81 00:04:23,080 --> 00:04:26,080 So a couple of bits. 82 00:04:26,080 --> 00:04:31,800 I've got some Kubernetes and AKS and then some copilot for security. 83 00:04:31,800 --> 00:04:35,960 So let's go for the AKS type stuff first. 84 00:04:35,960 --> 00:04:43,200 So firstly, the open source project called Draft that Microsoft keeps an eye on and maintains 85 00:04:43,200 --> 00:04:44,520 has been updated. 86 00:04:44,520 --> 00:04:46,720 It's including a new feature called Validate. 87 00:04:46,720 --> 00:04:52,240 So what Validate will do is it allows you to scan your manifest to see if you're following 88 00:04:52,240 --> 00:04:58,280 best practices so you can catch any issues early in the development lifecycle. 89 00:04:58,280 --> 00:05:03,480 And this is all part of our new AKS feature around deployment safeguards. 90 00:05:03,480 --> 00:05:08,260 So if people make silly mistakes, we catch them early, early. 91 00:05:08,260 --> 00:05:16,280 And then also in GA is support for disabling Windows Outbound NAT in AKS because we know 92 00:05:16,280 --> 00:05:20,320 it can cause a bit of problems in AKS pods. 93 00:05:20,320 --> 00:05:26,480 So now if you need to, you can go and turn that off when you're creating new Windows 94 00:05:26,480 --> 00:05:28,560 agent pools. 95 00:05:28,560 --> 00:05:30,800 Next is could it be my new baby? 96 00:05:30,800 --> 00:05:36,720 I don't know because I'm very loyal to Sentinel, but Microsoft Copilot for Security has a couple 97 00:05:36,720 --> 00:05:38,800 of new things in public preview. 98 00:05:38,800 --> 00:05:41,280 It has got Azure Firewall integration. 99 00:05:41,280 --> 00:05:47,240 So that means that you can retrieve IDPS signatures from an Azure Firewall. 100 00:05:47,240 --> 00:05:54,080 You can look at the threat profile of an IDPS signature and do some other things there. 101 00:05:54,080 --> 00:06:00,700 And then also, we also have WAF integration, public preview and copilot for security. 102 00:06:00,700 --> 00:06:05,920 So that means that you can have a look at, you can summarize and get copilot to pull 103 00:06:05,920 --> 00:06:10,360 things from the WAF and you can have a look at it in the prompt. 104 00:06:10,360 --> 00:06:12,160 So definitely exciting. 105 00:06:12,160 --> 00:06:17,680 If you're already on the Copilot for Security bandwagon, you should go and have a look at 106 00:06:17,680 --> 00:06:19,480 that. 107 00:06:19,480 --> 00:06:21,360 And that is my news. 108 00:06:21,360 --> 00:06:24,960 And with that, I get to kind of be Michael this time. 109 00:06:24,960 --> 00:06:29,040 We will turn our attention to our guest this week, who is Richard. 110 00:06:29,040 --> 00:06:34,520 Richard, do you want to tell us who you are and what you do at Microsoft? 111 00:06:34,520 --> 00:06:36,440 Hey, thank you very much. 112 00:06:36,440 --> 00:06:37,440 Yeah, Richard Diver. 113 00:06:37,440 --> 00:06:43,200 I've got 29 years experience now in the tech and security world, starting in the Navy back 114 00:06:43,200 --> 00:06:44,200 in the 90s. 115 00:06:44,200 --> 00:06:49,720 But my current role is the technical story design lead in Microsoft security marketing. 116 00:06:49,720 --> 00:06:52,640 Now, Richard, what does that mean? 117 00:06:52,640 --> 00:06:56,840 For those of us who might not know what technical design marketing is. 118 00:06:56,840 --> 00:07:00,080 I get to work between engineering and marketing. 119 00:07:00,080 --> 00:07:06,040 So I get to work with fun people like yourself and Michael and many other experts, work out 120 00:07:06,040 --> 00:07:10,120 the behind the scenes story and then bring it to the real world. 121 00:07:10,120 --> 00:07:11,800 That's basically what I get to do every day. 122 00:07:11,800 --> 00:07:12,800 I know. 123 00:07:12,800 --> 00:07:16,520 Well, I do know because I actually talk to you a lot, Richard. 124 00:07:16,520 --> 00:07:20,080 So, there's a couple of things. 125 00:07:20,080 --> 00:07:24,960 Now as Michael had said earlier, we're going to talk about AI security stuff, which I have 126 00:07:24,960 --> 00:07:30,200 to apologize was my note that I wrote on the podcast notes that we have. 127 00:07:30,200 --> 00:07:33,080 But let's talk about that. 128 00:07:33,080 --> 00:07:37,980 So we've had some guests on already talking about different aspects of AI security. 129 00:07:37,980 --> 00:07:43,320 So some earlier episodes, we had Ryan who talked about co-pilot for security. 130 00:07:43,320 --> 00:07:47,920 And we also had AI Red Team folks. 131 00:07:47,920 --> 00:07:50,120 We had Amanda and Pete on. 132 00:07:50,120 --> 00:07:59,360 But I know that you tell stories more kind of, how do I put this, straight up AI security. 133 00:07:59,360 --> 00:08:04,080 So if people are trying to get their heads around all of this, and obviously many people 134 00:08:04,080 --> 00:08:09,680 are trying to use AI, security people are worried because they need to understand it. 135 00:08:09,680 --> 00:08:12,880 Where would you say people would start with that? 136 00:08:12,880 --> 00:08:14,800 Good question. 137 00:08:14,800 --> 00:08:20,160 So AI security is not much different to any other security we've done before. 138 00:08:20,160 --> 00:08:24,080 There's a lot of basics in the data and identity layers. 139 00:08:24,080 --> 00:08:27,720 I think those are the two that most people get worried about is who's got access to it 140 00:08:27,720 --> 00:08:29,640 and what are you going to do with my data? 141 00:08:29,640 --> 00:08:34,360 So we have to go back and review some of the things we've maybe not done properly in the 142 00:08:34,360 --> 00:08:35,360 past. 143 00:08:35,360 --> 00:08:40,920 But the AI specific security problems, and this is where the Red Team comes in very handy, 144 00:08:40,920 --> 00:08:45,120 is that it's a little bit like you can social engineer it. 145 00:08:45,120 --> 00:08:46,720 And so that becomes a whole new problem. 146 00:08:46,720 --> 00:08:49,320 You can't social engineer machine learning. 147 00:08:49,320 --> 00:08:50,920 Machine learning has a start and an end. 148 00:08:50,920 --> 00:08:52,560 You put data in, you get data out. 149 00:08:52,560 --> 00:08:54,800 It's pretty solid. 150 00:08:54,800 --> 00:09:00,300 But generative AI, as we have today, that ends up with its own special unique characteristics 151 00:09:00,300 --> 00:09:08,120 that we have to be aware of, put some guardrails in place, and then monitor and watch and not 152 00:09:08,120 --> 00:09:09,120 let it fall down. 153 00:09:09,120 --> 00:09:10,640 That's basically where it goes. 154 00:09:10,640 --> 00:09:16,560 So you mentioned that people are concerned about identity and data. 155 00:09:16,560 --> 00:09:17,760 There's some top concerns. 156 00:09:17,760 --> 00:09:19,480 So let's start there. 157 00:09:19,480 --> 00:09:26,440 Now I know you did a Build session with my boss last week at Build or last week from 158 00:09:26,440 --> 00:09:28,360 the time we're recording this. 159 00:09:28,360 --> 00:09:30,660 That was all about data security. 160 00:09:30,660 --> 00:09:33,120 So why don't we start on that one? 161 00:09:33,120 --> 00:09:39,800 Tell me problem, what we should do about it, why it's different in the AI space. 162 00:09:39,800 --> 00:09:46,800 Yeah, so we luckily came up with the three golden rules that made life a lot easier for 163 00:09:46,800 --> 00:09:48,120 us to present it. 164 00:09:48,120 --> 00:09:50,220 And we had a lot of fun doing that one. 165 00:09:50,220 --> 00:09:56,820 So the rules basically say that data already has a lot of protections in place. 166 00:09:56,820 --> 00:10:00,800 So whether you're accessing a database or you're connecting to file servers, or maybe 167 00:10:00,800 --> 00:10:06,520 your user is uploading an email, looking at a website, every kind of data type you might 168 00:10:06,520 --> 00:10:09,840 use to interact with AI comes from somewhere. 169 00:10:09,840 --> 00:10:15,800 So hopefully it's either a trusted source like SharePoint or Office, or maybe it's coming 170 00:10:15,800 --> 00:10:18,600 from the internet, which means it might be untrusted. 171 00:10:18,600 --> 00:10:23,880 So knowing the kind of data you've got coming in, the idea is you should make sure that 172 00:10:23,880 --> 00:10:29,960 if there are already access controls or maybe API security, or maybe you started doing the 173 00:10:29,960 --> 00:10:36,440 more advanced labeling of content to say exactly what kind of sensitivity label this data has, 174 00:10:36,440 --> 00:10:41,360 you would make sure that when you get to your AI system, you do not want to let that AI 175 00:10:41,360 --> 00:10:45,800 system have unlimited access to the data on its own, because otherwise I will trick that 176 00:10:45,800 --> 00:10:49,040 AI into giving me all the data it has access to. 177 00:10:49,040 --> 00:10:53,560 So the first layer is make sure that your data is protected and that you don't undo 178 00:10:53,560 --> 00:10:58,720 those protections by giving even read-only access through a service account is not a 179 00:10:58,720 --> 00:10:59,720 good idea. 180 00:10:59,720 --> 00:11:05,360 So instead, you should use the user's credentials and get access to the data on behalf of the 181 00:11:05,360 --> 00:11:06,360 user. 182 00:11:06,360 --> 00:11:09,920 There are different ways of doing that, but that's the fundamentals of that first one. 183 00:11:09,920 --> 00:11:12,400 And then the second rule is about identity. 184 00:11:12,400 --> 00:11:17,300 And this one was surprising to a few customers I spoke to about this, but you should think 185 00:11:17,300 --> 00:11:21,360 about multi-factor authentication, what device the user is using. 186 00:11:21,360 --> 00:11:25,880 So when you get to the prompt of an AI system, that should be locked down and controlled 187 00:11:25,880 --> 00:11:27,560 to your trusted users. 188 00:11:27,560 --> 00:11:32,320 Now if you're making a customer-facing or public-facing AI app, you have a different 189 00:11:32,320 --> 00:11:36,880 set of problems, but mostly we're talking to enterprises and businesses that want to 190 00:11:36,880 --> 00:11:42,240 build some kind of chat GPT type functionality, but on their own data. 191 00:11:42,240 --> 00:11:46,080 And for that, they need to limit the scope down to the users that are trying to get to 192 00:11:46,080 --> 00:11:48,400 the data that that AI has access to. 193 00:11:48,400 --> 00:11:53,920 I have spoken to a few customers of building monolithic applications that has all the data 194 00:11:53,920 --> 00:11:57,320 with all the applications and all the users in one place. 195 00:11:57,320 --> 00:12:01,440 And really you're asking for trouble there because you're trying to expect that the LLM 196 00:12:01,440 --> 00:12:06,720 or the large language model is controlling access and I would not use that as a security 197 00:12:06,720 --> 00:12:07,920 layer. 198 00:12:07,920 --> 00:12:12,440 And then the third one is layers of security around the model. 199 00:12:12,440 --> 00:12:15,040 So models today, they're getting better. 200 00:12:15,040 --> 00:12:21,660 There's a, our red team works regularly with our own and other model makers to give feedback 201 00:12:21,660 --> 00:12:25,560 on how to improve model robustness in the first place and they're becoming more secure 202 00:12:25,560 --> 00:12:26,560 by default. 203 00:12:26,560 --> 00:12:32,480 However, I would always recommend that you put layers of controls like content safety, 204 00:12:32,480 --> 00:12:33,800 some kind of filtering. 205 00:12:33,800 --> 00:12:37,120 You don't want the user to work directly with them with an AI. 206 00:12:37,120 --> 00:12:42,200 You want some layers in the middle that look for potential bad things, filter them out. 207 00:12:42,200 --> 00:12:46,560 And then even when you get the answer back from the AI, you double check it and say, 208 00:12:46,560 --> 00:12:51,800 is the sensitivity of the data you're giving me either malicious or higher level than I 209 00:12:51,800 --> 00:12:54,040 trust this user to have access to. 210 00:12:54,040 --> 00:12:57,280 And that's, you get one last chance before you hand it to the user. 211 00:12:57,280 --> 00:13:02,360 So we've got, if you watch the build talk that we did, there's diagrams in there and 212 00:13:02,360 --> 00:13:04,880 it makes it much easier to explain. 213 00:13:04,880 --> 00:13:09,440 So many, many moons ago, I worked on this little thing called the security development 214 00:13:09,440 --> 00:13:10,440 life cycle, SDL. 215 00:13:10,440 --> 00:13:15,280 I heard through the great via that there's been some changes made to the SDL in light 216 00:13:15,280 --> 00:13:17,280 of AI security. 217 00:13:17,280 --> 00:13:20,280 Yes, there is because the build conference is coming up. 218 00:13:20,280 --> 00:13:24,280 We wanted to celebrate the 20 years of SDL. 219 00:13:24,280 --> 00:13:29,320 And so just before the build conference, we made it the new updated site go live. 220 00:13:29,320 --> 00:13:36,800 If you go and look at the SDL site now, we'll put a link in, but aka.ms slash Microsoft 221 00:13:36,800 --> 00:13:43,880 SDL, you'll see that we've created new 10 new practices and these practices cover end 222 00:13:43,880 --> 00:13:50,240 to end security for both developers and through DevSecOps and into live operations. 223 00:13:50,240 --> 00:13:52,880 And the intent is we will continue to update this. 224 00:13:52,880 --> 00:13:59,240 We have an initial set there today and then we'll continue to add detailed practices within 225 00:13:59,240 --> 00:14:02,360 each of these 10 in the future. 226 00:14:02,360 --> 00:14:06,540 Some of the best ones that we've added into the SDL is things like performing security 227 00:14:06,540 --> 00:14:08,800 design reviews and threat modeling. 228 00:14:08,800 --> 00:14:13,240 The number, I asked people to raise their hand in the session, but also other conversations 229 00:14:13,240 --> 00:14:17,120 I've had when I asked people if they do threat modeling, there's a little bit of an awkward 230 00:14:17,120 --> 00:14:18,600 silence of what is that? 231 00:14:18,600 --> 00:14:21,080 And I don't think we do it very well. 232 00:14:21,080 --> 00:14:25,320 So I'm glad that's now being called out something serious you need to go do. 233 00:14:25,320 --> 00:14:28,320 Another one would be the things like performing security testing. 234 00:14:28,320 --> 00:14:31,600 There's a lot of different security testing you have to do apart from pen testing and 235 00:14:31,600 --> 00:14:32,720 red teaming. 236 00:14:32,720 --> 00:14:39,240 And so in the software world and also physical world testing, and then as we get to the end 237 00:14:39,240 --> 00:14:41,280 of that, there's also training. 238 00:14:41,280 --> 00:14:45,760 And so as we all need to do is everybody needs to learn about security, which is why we bought 239 00:14:45,760 --> 00:14:47,880 more security sessions to build this year. 240 00:14:47,880 --> 00:14:52,680 We need everybody to take part in the security conversation going forwards. 241 00:14:52,680 --> 00:14:56,280 From a practical perspective, it's interesting you should bring that up. 242 00:14:56,280 --> 00:15:01,840 I can tell you right now, looking at the threat models that I review in Azure data, over the 243 00:15:01,840 --> 00:15:10,200 last, when I say nine, 12 months, there has been more of a focus on asking machine learning, 244 00:15:10,200 --> 00:15:17,480 large language models, AI protection of data of the models themselves, basically asking 245 00:15:17,480 --> 00:15:19,920 those questions during the threat model reviews. 246 00:15:19,920 --> 00:15:25,400 Whereas prior to the sort of wave of artificial intelligence, those questions honestly were 247 00:15:25,400 --> 00:15:29,600 not really asked, except in environments that were just pure, true AI. 248 00:15:29,600 --> 00:15:33,680 But yeah, so now there's a much bigger focus. 249 00:15:33,680 --> 00:15:38,980 That comment and those topics are brought up during basically every threat model review, 250 00:15:38,980 --> 00:15:39,980 as they should. 251 00:15:39,980 --> 00:15:44,680 Another work I got started in when I started learning AI last year, and I started to think, 252 00:15:44,680 --> 00:15:49,880 what are we going to say when all these different AI people at Microsoft are doing AI and they're 253 00:15:49,880 --> 00:15:51,160 doing security? 254 00:15:51,160 --> 00:15:54,240 And threat modeling became one of the quickest things we had to talk about because we had 255 00:15:54,240 --> 00:16:00,680 50 plus teams all rapidly building co-pilots and only one team having to check them all 256 00:16:00,680 --> 00:16:01,680 before they went production. 257 00:16:01,680 --> 00:16:05,840 And so that team needed some kind of standardization in how do you talk about AI? 258 00:16:05,840 --> 00:16:06,840 What are the components? 259 00:16:06,840 --> 00:16:14,400 And so building a, even building a tech stack, the fundamental tech stack for AI is the platform, 260 00:16:14,400 --> 00:16:16,060 the application, and then the usage. 261 00:16:16,060 --> 00:16:20,780 If you can go those three layers, then it's a good starting point to bucket things into 262 00:16:20,780 --> 00:16:22,640 at least three big areas. 263 00:16:22,640 --> 00:16:26,440 And then you need to make sure that you think of the threat from the user's prompt in the 264 00:16:26,440 --> 00:16:30,760 usage area through the application and then to the model. 265 00:16:30,760 --> 00:16:36,400 And it's always surprising how simple that is, but really effective as well. 266 00:16:36,400 --> 00:16:40,600 Well, I'm realizing I'm going completely off script here. 267 00:16:40,600 --> 00:16:43,600 By the way, for anyone who wants to know, we don't really have a script in this true 268 00:16:43,600 --> 00:16:44,600 sense of the word. 269 00:16:44,600 --> 00:16:47,920 We have just a couple of bullet points to sort of discuss. 270 00:16:47,920 --> 00:16:49,680 This is not in the list. 271 00:16:49,680 --> 00:16:51,480 So if you look at co-pilots, right? 272 00:16:51,480 --> 00:16:59,080 So co-pilots is the big Microsoft brand around using AI to bring it to the masses essentially. 273 00:16:59,080 --> 00:17:00,080 That's my perspective. 274 00:17:00,080 --> 00:17:08,280 I look at, we were designing the AI, the co-pilot for Cosmos DB and where you could basically 275 00:17:08,280 --> 00:17:11,560 give it a database and it could ask you questions like, how do I do this, that, and the other 276 00:17:11,560 --> 00:17:13,720 build up the SQL statement for you. 277 00:17:13,720 --> 00:17:19,320 And it was interesting, the safeguards that we had to put in place on the understanding 278 00:17:19,320 --> 00:17:27,200 that there was always the possibility that what question you asked could end up producing 279 00:17:27,200 --> 00:17:30,120 a query that was wrong. 280 00:17:30,120 --> 00:17:36,080 And we need to make sure that we had defenses in place to mitigate a potentially incorrect 281 00:17:36,080 --> 00:17:39,600 query coming back from the large language model. 282 00:17:39,600 --> 00:17:42,240 There's all these sort of things you have to think about, right? 283 00:17:42,240 --> 00:17:47,840 I mean, could you imagine, you know, you say, how do I query for ABC? 284 00:17:47,840 --> 00:17:53,080 And it gives you the prompt back or SQL statement that drops the table instead. 285 00:17:53,080 --> 00:17:54,080 Right. 286 00:17:54,080 --> 00:17:59,000 So, you know, we do things like warning people like, hey, you know, you're connecting as 287 00:17:59,000 --> 00:18:00,580 an admin. 288 00:18:00,580 --> 00:18:03,240 You really shouldn't be running these queries as an admin. 289 00:18:03,240 --> 00:18:06,040 First of all, it's just a bad idea anyway. 290 00:18:06,040 --> 00:18:11,840 But you know, if something goes wrong, then it can go wrong because you're an admin, you're 291 00:18:11,840 --> 00:18:14,120 violating the principle of least privilege. 292 00:18:14,120 --> 00:18:16,440 So you really probably shouldn't do this. 293 00:18:16,440 --> 00:18:22,480 And there's stuff in there that's like, you know, there's prompts to kind of just educate 294 00:18:22,480 --> 00:18:24,140 people on that. 295 00:18:24,140 --> 00:18:29,080 But there's also some real good stuff that we did under the covers to help mitigate and 296 00:18:29,080 --> 00:18:34,680 detect a potentially, you know, a hallucination of a query. 297 00:18:34,680 --> 00:18:36,840 Do you see that kind of stuff happening as well? 298 00:18:36,840 --> 00:18:39,720 Yeah, that makes a lot of sense. 299 00:18:39,720 --> 00:18:42,960 From a zero trust perspective, you should trust that the users probably put something 300 00:18:42,960 --> 00:18:48,420 in wrong and then trust that the, or don't trust that the LLM interpreted what they wanted 301 00:18:48,420 --> 00:18:50,640 to do correctly. 302 00:18:50,640 --> 00:18:54,560 And LLMs or large language models, even small language models, you know, they're very broad 303 00:18:54,560 --> 00:18:56,860 in all the things they can do. 304 00:18:56,860 --> 00:18:59,160 That doesn't mean they're very good at specific tasks. 305 00:18:59,160 --> 00:19:04,220 And so there was a really good session, a build between Mark Rusinovich and Scott Hanselman, 306 00:19:04,220 --> 00:19:08,600 where they were doing a very basic demo where they were trying to clean up the desktop icons. 307 00:19:08,600 --> 00:19:10,360 It's a really good storyline. 308 00:19:10,360 --> 00:19:14,800 But in doing so, what they got the AI to do or the generative AI to do was actually call 309 00:19:14,800 --> 00:19:17,160 features and functions that knew what they were doing. 310 00:19:17,160 --> 00:19:21,160 So basic tasks like count how many text files are on my desktop. 311 00:19:21,160 --> 00:19:23,680 The LLM, terrible at that. 312 00:19:23,680 --> 00:19:25,920 It didn't, it got the number wrong every time. 313 00:19:25,920 --> 00:19:29,600 But if it called a function and it called the correct function, it would get the correct 314 00:19:29,600 --> 00:19:30,600 answer back. 315 00:19:30,600 --> 00:19:36,400 And so I think what we're seeing is AI using AI is where we're going in the future. 316 00:19:36,400 --> 00:19:41,880 And that's good and bad because it all comes down to the developer putting these Lego pieces 317 00:19:41,880 --> 00:19:43,000 together. 318 00:19:43,000 --> 00:19:47,360 But if one AI gets it wrong and another AI gets it wrong, how many chances you got of 319 00:19:47,360 --> 00:19:48,720 getting it right. 320 00:19:48,720 --> 00:19:55,120 So the final part of that is we're looking at using something called an AI watchdog. 321 00:19:55,120 --> 00:20:02,400 And just purely from a security perspective of jailbreaks and prompt injection, you can't 322 00:20:02,400 --> 00:20:07,340 always tell from the language the user used that what they were trying to do was a jailbreak. 323 00:20:07,340 --> 00:20:13,640 And so you need AI, separate AI that's not influenced by the user's instructions, watching 324 00:20:13,640 --> 00:20:20,280 both the input of the prompt into the LLM and then the output from the LLM and say, 325 00:20:20,280 --> 00:20:21,800 do these two things match? 326 00:20:21,800 --> 00:20:23,920 Does any of this look suspicious in any way? 327 00:20:23,920 --> 00:20:26,800 And it's not just one-on-one, but like the whole conversation. 328 00:20:26,800 --> 00:20:28,960 And you're looking at things like intent and semantics. 329 00:20:28,960 --> 00:20:30,800 And so that's where we're heading with it. 330 00:20:30,800 --> 00:20:35,240 Like we need AI to protect AI or protect users and AI interactions. 331 00:20:35,240 --> 00:20:42,000 So Richard, you mentioned earlier on about the different layers of an AI application. 332 00:20:42,000 --> 00:20:48,600 But for the security folks out there, how does that translate into something that we 333 00:20:48,600 --> 00:20:52,960 can do something with in terms of say defenses or mitigations that we put in place? 334 00:20:52,960 --> 00:20:55,200 Can you run us through that? 335 00:20:55,200 --> 00:20:56,480 Absolutely. 336 00:20:56,480 --> 00:21:00,720 So if we started the usage layer, the first thing to think about is that we still have 337 00:21:00,720 --> 00:21:08,480 all the same cybersecurity risks from things like an insider, which is growing in popularity. 338 00:21:08,480 --> 00:21:12,320 Then there's social engineering and phishing links and phishing documents. 339 00:21:12,320 --> 00:21:15,440 They're all going to be amplified by the use of AI by the attackers. 340 00:21:15,440 --> 00:21:21,240 They're going to find better ways to do that in localized languages or targeting individuals. 341 00:21:21,240 --> 00:21:26,600 And then we have this not a new concept, but we're resurfacing it with poisoned content. 342 00:21:26,600 --> 00:21:30,520 And that idea being that you might trust the content is benign. 343 00:21:30,520 --> 00:21:34,280 Let's say I'm a hiring manager and I'm going to look at a bunch of CVs. 344 00:21:34,280 --> 00:21:39,480 And so I might want to take 10 CVs from my top candidates and have the language model 345 00:21:39,480 --> 00:21:40,840 review them all for me. 346 00:21:40,840 --> 00:21:45,800 What I don't know is that somebody might have hidden some kind of instructions inside of 347 00:21:45,800 --> 00:21:49,960 that Word document or even an image or an audio file, whatever it might be. 348 00:21:49,960 --> 00:21:54,560 So at that usage layer, there's a whole set of things that we can do to protect ourselves, 349 00:21:54,560 --> 00:21:59,400 like identity security, make sure we're not logging in as admin so that if we do trigger 350 00:21:59,400 --> 00:22:03,120 some kind of phishing link, we're not causing more damage later on. 351 00:22:03,120 --> 00:22:08,300 We can do content filtering, some basic things like just know where the content came from. 352 00:22:08,300 --> 00:22:13,200 Looking for white text on white background is such an obvious one. 353 00:22:13,200 --> 00:22:16,440 And also special characters, smiley faces. 354 00:22:16,440 --> 00:22:19,720 There's lots of telltale signs, very basic and simple you could do. 355 00:22:19,720 --> 00:22:23,160 I think those defenses will get more mature in time. 356 00:22:23,160 --> 00:22:28,200 Once we create the prompt and then we create the content and we add it all together, we're 357 00:22:28,200 --> 00:22:29,640 now into the application. 358 00:22:29,640 --> 00:22:32,780 So we've sent that in when we're waiting for a response. 359 00:22:32,780 --> 00:22:34,200 That's when prompt filtering comes in. 360 00:22:34,200 --> 00:22:39,840 This is something that in Microsoft Copilot, we build all this in and we look after all 361 00:22:39,840 --> 00:22:42,880 that, we make sure it's best defenses we can make. 362 00:22:42,880 --> 00:22:46,560 But if you're building your own, you need to make sure that your prompt filtering is 363 00:22:46,560 --> 00:22:49,840 kind of top of the game to prevent some of the obvious harms. 364 00:22:49,840 --> 00:22:52,320 It's like malware filtering today. 365 00:22:52,320 --> 00:22:57,600 Then you have all your applications, security controls, it's the normal things like SQL 366 00:22:57,600 --> 00:23:01,480 commands or API security, et cetera. 367 00:23:01,480 --> 00:23:03,240 Eventually you're going to build a meta prompt. 368 00:23:03,240 --> 00:23:07,400 So a system meta prompt is where you take what the user said they wanted you to do, 369 00:23:07,400 --> 00:23:10,200 you've cleaned it up and then you might add additional guardrails. 370 00:23:10,200 --> 00:23:15,320 Like I want you to act like a professional health and safety representative and I want 371 00:23:15,320 --> 00:23:18,000 you to answer the user's questions in a certain way. 372 00:23:18,000 --> 00:23:21,200 Then you give a whole bunch of instructions that comes along with whatever the user asked 373 00:23:21,200 --> 00:23:22,200 for. 374 00:23:22,200 --> 00:23:24,540 Now we package that up and then we send that to the LLM. 375 00:23:24,540 --> 00:23:30,200 So now we're in the platform and in the platform again we've got layers of security here but 376 00:23:30,200 --> 00:23:34,440 in that deep safety system we're certainly filtering for harmful content and there's 377 00:23:34,440 --> 00:23:35,760 a bunch of categories for that. 378 00:23:35,760 --> 00:23:43,160 We've just updated the Azure AI Content Safety Filter now has customizable filters. 379 00:23:43,160 --> 00:23:47,400 So if you don't want certain words or certain kinds of language being used, you can change 380 00:23:47,400 --> 00:23:48,400 it. 381 00:23:48,400 --> 00:23:54,280 But we've also got political problems and racist problems and harmful content like 382 00:23:54,280 --> 00:23:57,540 bio weapons, all of that can get filtered out. 383 00:23:57,540 --> 00:23:59,680 And then we ask the LLM the question. 384 00:23:59,680 --> 00:24:03,560 So it's taking a long time to get there, it's milliseconds in the real world, but by the 385 00:24:03,560 --> 00:24:06,720 time it now gets the LLM it's going to process whatever information it's given and it's going 386 00:24:06,720 --> 00:24:08,880 to send it back out. 387 00:24:08,880 --> 00:24:11,720 That's your on the way back out, we're back into the application layer. 388 00:24:11,720 --> 00:24:18,640 We can look at what data is if I used content from HR and content from finance and content 389 00:24:18,640 --> 00:24:21,560 from legal, what is my outcome? 390 00:24:21,560 --> 00:24:23,560 Is it a mix of all three? 391 00:24:23,560 --> 00:24:24,560 Is it none? 392 00:24:24,560 --> 00:24:27,640 Because we've cleared all the sensitive data out like what is it? 393 00:24:27,640 --> 00:24:32,400 And so you have to do some kind of sensitive labeling on the data going out. 394 00:24:32,400 --> 00:24:35,800 But you also want to check it again to see if it's appropriate for the user that's using 395 00:24:35,800 --> 00:24:36,800 it. 396 00:24:36,800 --> 00:24:41,480 So are you giving out customer pricing lists when it's the customer you're sending it to? 397 00:24:41,480 --> 00:24:43,280 No, I want to send that to my sales reps. 398 00:24:43,280 --> 00:24:45,100 I don't want to send it to my customer. 399 00:24:45,100 --> 00:24:48,140 And so you can check all these things before you then send it out. 400 00:24:48,140 --> 00:24:51,700 And then we're back into the first layer again, which is the usage layer. 401 00:24:51,700 --> 00:24:55,960 If the user isn't on a corporate managed device or on the right kind of application, if it's 402 00:24:55,960 --> 00:24:59,560 anything that's suspicious, you might not want to send that data to them. 403 00:24:59,560 --> 00:25:05,640 That's some of the normal layers of security we put in place, like don't send it to Dropbox, 404 00:25:05,640 --> 00:25:08,720 just only send it to OneDrive or only send it to an office application. 405 00:25:08,720 --> 00:25:13,320 So there's controls, that's 10 controls I've just told you about in the diagram I'm staring 406 00:25:13,320 --> 00:25:14,800 at. 407 00:25:14,800 --> 00:25:19,880 But there's many different ways that you can try to prevent, expect it's going to go wrong 408 00:25:19,880 --> 00:25:23,440 and then do detection and response on the way back in real time. 409 00:25:23,440 --> 00:25:27,880 Richard, you're talking to loads and loads of customers and I'm sure people who are listening 410 00:25:27,880 --> 00:25:29,920 are thinking about this too. 411 00:25:29,920 --> 00:25:34,400 But I'd like to ask people, where do you think, because there's a lot to do here, where do 412 00:25:34,400 --> 00:25:39,840 you think is the best place people can start and be the most impactful with these controls 413 00:25:39,840 --> 00:25:42,880 and mitigations and things they should be worried about? 414 00:25:42,880 --> 00:25:47,640 Because of course, we'd love to say do everything straight away, but we all know that that's 415 00:25:47,640 --> 00:25:49,600 not really possible, right? 416 00:25:49,600 --> 00:25:50,600 Absolutely. 417 00:25:50,600 --> 00:25:54,280 I can plug my book here, I think, and say go read my book. 418 00:25:54,280 --> 00:25:58,280 So I put all this together just a couple of months ago, I decided to write a book on this 419 00:25:58,280 --> 00:25:59,600 topic. 420 00:25:59,600 --> 00:26:03,040 And through the chapters, I try to structure that of like all the things you need to think 421 00:26:03,040 --> 00:26:04,040 about. 422 00:26:04,040 --> 00:26:09,960 But apart from reading the book, I would say that at a bare minimum, start drawing diagrams 423 00:26:09,960 --> 00:26:15,400 is my personal offering to the world is like without a good diagram, you don't know what 424 00:26:15,400 --> 00:26:16,680 people don't know. 425 00:26:16,680 --> 00:26:18,060 And you don't know what you don't know either. 426 00:26:18,060 --> 00:26:22,060 So by putting the diagram together of what you think the system looks like, then trace 427 00:26:22,060 --> 00:26:27,440 the data through that system and see if everyone else in the room agrees or as what ifs, you 428 00:26:27,440 --> 00:26:32,480 know, and just keep asking what if what happens if I interject a command here or what if I 429 00:26:32,480 --> 00:26:33,480 leak the data there. 430 00:26:33,480 --> 00:26:37,080 And so they're going back to threat modeling and that idea of getting more people involved 431 00:26:37,080 --> 00:26:38,280 in threat modeling. 432 00:26:38,280 --> 00:26:43,280 A diagram for me is why I'm that's why I moved into marketing is I want to make sure that 433 00:26:43,280 --> 00:26:47,640 we we provide people with the tools to go and make a difference. 434 00:26:47,640 --> 00:26:52,920 And that's a very nice segue into our next into what I was going to ask you about next, 435 00:26:52,920 --> 00:26:53,920 Richard. 436 00:26:53,920 --> 00:26:58,800 As you said, you've written a book, both Michael and I have also written books. 437 00:26:58,800 --> 00:27:04,440 And I know there's a lot of people in industry who quite like the idea of doing it. 438 00:27:04,440 --> 00:27:11,140 So although a little bit of a segue, I wanted to ask you what what's your advice for for 439 00:27:11,140 --> 00:27:17,520 people who might want to get their name on a book and be an author to show it to the 440 00:27:17,520 --> 00:27:18,520 nearest and dearest. 441 00:27:18,520 --> 00:27:19,920 I mean, I'm not going to lie. 442 00:27:19,920 --> 00:27:24,840 That's basically my was my motivation originally for doing it, that I just wanted to say I 443 00:27:24,840 --> 00:27:27,920 was an author more than anything. 444 00:27:27,920 --> 00:27:28,920 That's about it. 445 00:27:28,920 --> 00:27:29,920 I'll be honest with you. 446 00:27:29,920 --> 00:27:30,920 You don't do it for the money. 447 00:27:30,920 --> 00:27:34,760 I would say you do it because the best reason I found is to learn. 448 00:27:34,760 --> 00:27:38,520 So I thought I knew a lot about AI as like confident enough to write a book. 449 00:27:38,520 --> 00:27:39,520 Why not? 450 00:27:39,520 --> 00:27:42,520 And so I started to track like, what does it take to write? 451 00:27:42,520 --> 00:27:45,520 I've written books before, but I've written them on Windows and Sentinel. 452 00:27:45,520 --> 00:27:49,040 They're very technical, and I wanted to make sure this book was to a broader audience. 453 00:27:49,040 --> 00:27:50,320 So I just started writing it. 454 00:27:50,320 --> 00:27:52,160 But I don't know how you both approached it. 455 00:27:52,160 --> 00:27:54,640 Maybe you can share your experiences here. 456 00:27:54,640 --> 00:28:00,880 But the easiest way to get started, if anybody wants to do it, is you create your 10 chapters 457 00:28:00,880 --> 00:28:04,280 and inside each of your 10 chapters, you write 10 bullet points. 458 00:28:04,280 --> 00:28:06,480 And that gives you the bones of your book. 459 00:28:06,480 --> 00:28:13,840 Now with that, it will take you, I'd say 140 hours of writing, and then 60 hours of perfecting, 460 00:28:13,840 --> 00:28:20,400 which is research and diagrams and editing and getting reviews from 10 more people. 461 00:28:20,400 --> 00:28:26,160 But eventually, you're aiming to write somewhere between 60 and 70,000 words, and it will take 462 00:28:26,160 --> 00:28:28,200 you about 200 hours of effort. 463 00:28:28,200 --> 00:28:32,040 So depending on how many books you sell and how much you sell the book for and whether 464 00:28:32,040 --> 00:28:36,280 your publisher takes most of the profit or not, you're probably not going to get minimum 465 00:28:36,280 --> 00:28:37,280 wage for writing a book. 466 00:28:37,280 --> 00:28:44,120 But what you do get is all the lessons learned of having to, the two hardest chapters for 467 00:28:44,120 --> 00:28:50,600 me, has to be the ethical framework and then AI governance. 468 00:28:50,600 --> 00:28:55,080 So these are two things that don't normally think about ethics in a security world, some 469 00:28:55,080 --> 00:28:56,440 people do, but not everybody. 470 00:28:56,440 --> 00:29:00,120 When you're from a technical background, ethics weren't the thing I studied. 471 00:29:00,120 --> 00:29:04,120 And then AI governance is so new that took a lot more research to understand where are 472 00:29:04,120 --> 00:29:05,920 we at in the world of governance. 473 00:29:05,920 --> 00:29:09,860 And so you have to dig into the bits that you don't know very well. 474 00:29:09,860 --> 00:29:13,720 But now I can say I'm more confident in being able to have these conversations just because 475 00:29:13,720 --> 00:29:16,360 I spent those 200 hours in writing. 476 00:29:16,360 --> 00:29:19,600 And then it is nice to feel it in your hands and you've got a book and you can put it on 477 00:29:19,600 --> 00:29:20,600 the shelf. 478 00:29:20,600 --> 00:29:23,720 And every now and again, someone says, Hey, I bought your book and I'm reading it. 479 00:29:23,720 --> 00:29:24,720 That's pretty cool. 480 00:29:24,720 --> 00:29:25,720 Yeah. 481 00:29:25,720 --> 00:29:31,240 I don't know how many hours we spent on the last book, which are designing and developing 482 00:29:31,240 --> 00:29:33,640 Secure ASI solutions. 483 00:29:33,640 --> 00:29:36,520 As you don't know, I don't really track it to be absolutely honest with you, but I do 484 00:29:36,520 --> 00:29:44,600 agree with agreeing on basically the chapter outline and then the main points in each chapter 485 00:29:44,600 --> 00:29:47,660 on the understanding that you're probably going to change. 486 00:29:47,660 --> 00:29:49,440 We were pretty lucky. 487 00:29:49,440 --> 00:29:56,080 I think that the outline that we had for the book, I think we deleted one chapter and we 488 00:29:56,080 --> 00:30:00,560 took the content of that chapter and infused it into other chapters. 489 00:30:00,560 --> 00:30:04,560 So for example, that actual chapter was going to be a whole chapter on Defender, Defender 490 00:30:04,560 --> 00:30:06,640 for cloud. 491 00:30:06,640 --> 00:30:11,800 And we realized it was actually better to not make it a separate chapter because it 492 00:30:11,800 --> 00:30:18,960 was changing so much and rather just infused the best practices into the other chapters 493 00:30:18,960 --> 00:30:19,960 instead. 494 00:30:19,960 --> 00:30:22,440 Just things that you should be aware of when it comes to, for example, in the Key Vault 495 00:30:22,440 --> 00:30:25,520 chapter, we talk about Defender for Key Vault, right? 496 00:30:25,520 --> 00:30:30,400 In the storage stuff and the database chapter, for example, is a bit better example. 497 00:30:30,400 --> 00:30:35,480 We talk about Defender for databases and we talk about sort of the overall security score 498 00:30:35,480 --> 00:30:37,920 just conceptually. 499 00:30:37,920 --> 00:30:38,920 But I agree with you. 500 00:30:38,920 --> 00:30:43,000 I think getting that structure in your head helps substantially. 501 00:30:43,000 --> 00:30:45,760 So, but yeah, I agree with that a hundred percent. 502 00:30:45,760 --> 00:30:47,400 As for timing, I don't know. 503 00:30:47,400 --> 00:30:50,020 I'm terrible when it comes to predicting timing. 504 00:30:50,020 --> 00:30:51,520 Some people know this story. 505 00:30:51,520 --> 00:30:58,280 When I wrote the crypto chapter, it was supposed to be 25 pages long and ended up being 88. 506 00:30:58,280 --> 00:31:04,280 So yeah, that'll give you an example of how good I am at working out what we should be 507 00:31:04,280 --> 00:31:07,200 doing in terms of time and length. 508 00:31:07,200 --> 00:31:08,200 Yeah. 509 00:31:08,200 --> 00:31:12,560 That's why I recorded all my times to just work out how long does it take. 510 00:31:12,560 --> 00:31:17,920 It's about 3, 530 words per hour and that's when you'll spend 140 hours. 511 00:31:17,920 --> 00:31:19,680 So if you go to 60,000 words. 512 00:31:19,680 --> 00:31:20,680 But you're right. 513 00:31:20,680 --> 00:31:23,560 And that's actually why a lot of people don't write a book because they don't know how long 514 00:31:23,560 --> 00:31:27,880 it's going to take or what efforts involved or they get two chapters into it and it's 515 00:31:27,880 --> 00:31:28,880 amazing. 516 00:31:28,880 --> 00:31:30,120 And now you've got the grunt work to do. 517 00:31:30,120 --> 00:31:31,600 You've got to finish it. 518 00:31:31,600 --> 00:31:34,240 So well, it's more than just that. 519 00:31:34,240 --> 00:31:35,240 You're absolutely correct. 520 00:31:35,240 --> 00:31:36,520 But it's more than that. 521 00:31:36,520 --> 00:31:38,040 You're under a time constraint. 522 00:31:38,040 --> 00:31:40,080 Like you can't take forever to write a book. 523 00:31:40,080 --> 00:31:44,040 You just can't, you know, I mean, people, you know, they're holding your feet to the 524 00:31:44,040 --> 00:31:45,040 fire. 525 00:31:45,040 --> 00:31:47,480 You can't just lollygag your way through writing a book. 526 00:31:47,480 --> 00:31:49,840 You know, you've we actually hit our schedule. 527 00:31:49,840 --> 00:31:53,880 In fact, we were actually about about a week and a half ahead of schedule when we were 528 00:31:53,880 --> 00:31:54,880 done. 529 00:31:54,880 --> 00:31:57,400 We were very proud of ourselves, but you know, we put in the hours. 530 00:31:57,400 --> 00:32:02,760 My guess is we probably put in more hours than we thought to make that schedule though. 531 00:32:02,760 --> 00:32:04,560 But that's just the nature of the beast. 532 00:32:04,560 --> 00:32:09,360 I mean, you bring up a really interesting point about learning stuff. 533 00:32:09,360 --> 00:32:12,640 You will learn stuff as you write the book. 534 00:32:12,640 --> 00:32:16,360 Even if you're an expert in the topic, you will still learn things. 535 00:32:16,360 --> 00:32:22,280 And in fact, one thing I've found with doing, you know, sort of the stuff that I write about, 536 00:32:22,280 --> 00:32:24,360 you'll find bugs in products too. 537 00:32:24,360 --> 00:32:28,520 So there's a product from the Key Vault team called Managed HSM, Managed Hardware Security 538 00:32:28,520 --> 00:32:29,520 Module. 539 00:32:29,520 --> 00:32:32,480 And there was a bug signing data. 540 00:32:32,480 --> 00:32:33,880 There's a bug in their REST API. 541 00:32:33,880 --> 00:32:35,880 It wasn't a bug in the HSM. 542 00:32:35,880 --> 00:32:39,200 But I, you know, I was doing some experiments and I found this bug calling the REST API. 543 00:32:39,200 --> 00:32:42,600 It was brand, by the way, it wasn't even shipping when I was using it. 544 00:32:42,600 --> 00:32:44,480 So you know, bugs are expected. 545 00:32:44,480 --> 00:32:49,600 But yeah, you'll probably find bugs and discrepancies in products and features as well when you're 546 00:32:49,600 --> 00:32:52,160 writing a book because you're being very disciplined about it when you're writing. 547 00:32:52,160 --> 00:32:55,320 I hope, you know, I hope people are disciplined when they're writing. 548 00:32:55,320 --> 00:32:56,680 And yeah, you'll probably find issues. 549 00:32:56,680 --> 00:32:58,680 Is that something you came across too? 550 00:32:58,680 --> 00:33:02,720 Yeah, I started to come across the what if scenarios. 551 00:33:02,720 --> 00:33:06,600 So I'm not in the red team, but I get to work with them so often that I work, you get to 552 00:33:06,600 --> 00:33:10,280 see how they work or some of the thought patterns they go through. 553 00:33:10,280 --> 00:33:11,880 And it's really just what if. 554 00:33:11,880 --> 00:33:17,920 And so I thought about things like what would a poisoned well or a honey pot look like if 555 00:33:17,920 --> 00:33:21,640 you reverse what we're using for today, the honey pot is to attract the attackers. 556 00:33:21,640 --> 00:33:23,960 Well, what if we try to attract AI? 557 00:33:23,960 --> 00:33:25,240 What if we get AI to go on? 558 00:33:25,240 --> 00:33:26,880 And there's been some examples in the news recently. 559 00:33:26,880 --> 00:33:31,480 I won't repeat them now, but there's some interesting stories of AI telling users to 560 00:33:31,480 --> 00:33:33,960 do things because it found it on Reddit. 561 00:33:33,960 --> 00:33:37,800 And so if you can attract the AI to go and use something as a source of truth, and it's 562 00:33:37,800 --> 00:33:41,360 really not, that's a great scenario. 563 00:33:41,360 --> 00:33:45,960 And on your point like about being under the gun, when normally I would be and I'd have 564 00:33:45,960 --> 00:33:46,960 been told you're on a schedule. 565 00:33:46,960 --> 00:33:51,080 So if you go to a publisher, you'll be under a schedule and you'll say, I'll submit X chapter 566 00:33:51,080 --> 00:33:55,240 by a certain date and then version two X number weeks later. 567 00:33:55,240 --> 00:33:57,240 I didn't, I went self-published. 568 00:33:57,240 --> 00:34:00,440 And that's another angle is you do everything yourself. 569 00:34:00,440 --> 00:34:02,760 It means you've got the freedom of time, but you're right. 570 00:34:02,760 --> 00:34:06,800 If you start writing and you take two years, how relevant is the information you wrote 571 00:34:06,800 --> 00:34:07,800 two years ago? 572 00:34:07,800 --> 00:34:09,140 And you'll end up rewriting it so many times. 573 00:34:09,140 --> 00:34:11,680 So I gave myself six weeks. 574 00:34:11,680 --> 00:34:13,680 I wrote it in six weeks. 575 00:34:13,680 --> 00:34:15,720 So I don't encourage anybody to do that. 576 00:34:15,720 --> 00:34:17,880 I had a lot of spare time on my hands. 577 00:34:17,880 --> 00:34:22,320 I just uninstalled Instagram and spent all the spare time writing a book instead. 578 00:34:22,320 --> 00:34:23,320 Yeah. 579 00:34:23,320 --> 00:34:27,240 Uninstalling social media is probably just a good idea anyway. 580 00:34:27,240 --> 00:34:28,240 Yeah. 581 00:34:28,240 --> 00:34:29,240 But that's interesting. 582 00:34:29,240 --> 00:34:31,560 So you did it all in six weeks. 583 00:34:31,560 --> 00:34:32,560 That's pretty good. 584 00:34:32,560 --> 00:34:36,960 I mean, all things considered, you know, but again, and just for people who are listening 585 00:34:36,960 --> 00:34:40,960 need to understand, we're not paid as Microsoft employees to write these books. 586 00:34:40,960 --> 00:34:43,780 I mean, they're basically deemed moonlighting. 587 00:34:43,780 --> 00:34:46,000 You actually have to get an agreement from your manager. 588 00:34:46,000 --> 00:34:50,160 In fact, in my case, I had to get an agreement from my manager's manager as well to actually 589 00:34:50,160 --> 00:34:55,400 write the book and you actually sign a moonlighting document because you're essentially earning 590 00:34:55,400 --> 00:34:59,360 money outside of, outside of Microsoft and it's not a Microsoft sanction thing. 591 00:34:59,360 --> 00:35:00,760 It's not part of your job. 592 00:35:00,760 --> 00:35:03,500 You are literally moonlighting. 593 00:35:03,500 --> 00:35:04,500 So yeah. 594 00:35:04,500 --> 00:35:07,520 And to your point, you're not going to, you're not going to retire on many books. 595 00:35:07,520 --> 00:35:12,660 That being said, writing secure code and writing secure code second edition were immensely 596 00:35:12,660 --> 00:35:15,240 successful by any measure. 597 00:35:15,240 --> 00:35:21,560 I mean, I'm not talking John Grisham novels successful, but in terms of tech, tech books 598 00:35:21,560 --> 00:35:24,640 successful, they were ahead of the pack. 599 00:35:24,640 --> 00:35:29,560 We sold a lot of those books, a lot of books. 600 00:35:29,560 --> 00:35:32,160 But yeah, for the most part, you're not going to, like you say, you're not going to retire 601 00:35:32,160 --> 00:35:35,040 on the earnings from a, from writing a book. 602 00:35:35,040 --> 00:35:38,180 It's almost a labor of love if nothing else. 603 00:35:38,180 --> 00:35:42,720 It's good to see people are still reading books with the opportunity of AI and search 604 00:35:42,720 --> 00:35:47,880 to just go find it, you know, summarize this for me or, or tell me what you think of X 605 00:35:47,880 --> 00:35:48,880 topic. 606 00:35:48,880 --> 00:35:49,880 You can learn so fast. 607 00:35:49,880 --> 00:35:51,880 That's one of the things I love about AI is I don't search anymore. 608 00:35:51,880 --> 00:35:55,400 I just ask AI for its opinion on something and it does all the searching for me and then 609 00:35:55,400 --> 00:35:56,920 brings you back to top five websites. 610 00:35:56,920 --> 00:36:00,320 Now not always accurate and you got to double check it, but it's easier than doing like 611 00:36:00,320 --> 00:36:02,320 a search and getting 10,000 pages. 612 00:36:02,320 --> 00:36:03,700 I can go look at. 613 00:36:03,700 --> 00:36:10,020 So they're one of the problems we've got in publishing is it's getting swamped with low 614 00:36:10,020 --> 00:36:15,540 quality AI created content, which people are just trying to do for like a quick buck. 615 00:36:15,540 --> 00:36:19,320 So if you are an expert out there and you do know your topic, you should think about 616 00:36:19,320 --> 00:36:23,120 contributing to the proper corpus of the library, right? 617 00:36:23,120 --> 00:36:27,960 And having content that people can read and appreciate because there is an audience out 618 00:36:27,960 --> 00:36:28,960 there. 619 00:36:28,960 --> 00:36:33,020 And this is where you might not self publish and you might go down the path of a publisher 620 00:36:33,020 --> 00:36:36,520 because if you can convince them it's a good book, they can do all the legwork of getting 621 00:36:36,520 --> 00:36:37,520 that book out there. 622 00:36:37,520 --> 00:36:42,080 And we're still selling books we wrote four or five years ago, one or two copies a week 623 00:36:42,080 --> 00:36:46,080 maybe, but it's still a, you know, Windows security and Sentinel. 624 00:36:46,080 --> 00:36:49,280 They're still books people want to pick up today and go and read about. 625 00:36:49,280 --> 00:36:51,980 So they have long lifespans too. 626 00:36:51,980 --> 00:36:55,040 So be careful what you write in your book because it will still be around in five and 627 00:36:55,040 --> 00:36:56,040 10 years time. 628 00:36:56,040 --> 00:36:59,680 But you bring up an interesting point there though. 629 00:36:59,680 --> 00:37:01,080 I love to read. 630 00:37:01,080 --> 00:37:03,960 I read all the time. 631 00:37:03,960 --> 00:37:09,560 I do use AI to summarize some things once in a while, but I do enjoy reading a book. 632 00:37:09,560 --> 00:37:13,440 Look, this is probably more information than anyone needs to know, but I have a whole bunch 633 00:37:13,440 --> 00:37:19,560 of Kindles around, let's just say strategically placed around the house and leave it at that. 634 00:37:19,560 --> 00:37:24,200 And I have one in my laptop bag. 635 00:37:24,200 --> 00:37:26,560 Whenever I go on vacation, I take my laptop with me because of course I have to take my 636 00:37:26,560 --> 00:37:27,600 laptop with me. 637 00:37:27,600 --> 00:37:30,680 I have a Kindle in there as well. 638 00:37:30,680 --> 00:37:35,240 Because they use so little power, you know, the thing can stay on power for a long time. 639 00:37:35,240 --> 00:37:38,280 But yeah, and then that way, you know, if I'm reading something, I can just pick it 640 00:37:38,280 --> 00:37:40,760 up and you know, where I left off. 641 00:37:40,760 --> 00:37:43,560 But yeah, I read all the time. 642 00:37:43,560 --> 00:37:49,080 I probably don't use AI as much as I should to sort of summarize stuff. 643 00:37:49,080 --> 00:37:50,920 I probably should start doing that a little bit more. 644 00:37:50,920 --> 00:37:55,600 The best part of course, you can ask AI to summarize things in a snarky voice, which 645 00:37:55,600 --> 00:37:58,400 is even more interesting. 646 00:37:58,400 --> 00:38:01,840 But yeah, is there anything else you want to add before we wrap this up? 647 00:38:01,840 --> 00:38:06,080 If you're wondering why Sarah's not chiming in, she's actually got a, just right now, 648 00:38:06,080 --> 00:38:09,240 she's got a horrible internet connection and she's just dropped. 649 00:38:09,240 --> 00:38:13,720 So I just sent her a message on Teams and just say, Hey, just hang in there and Richard 650 00:38:13,720 --> 00:38:14,960 and I will wrap this thing up. 651 00:38:14,960 --> 00:38:21,200 And on that very topic, so Richard, one thing we always ask our guests is if they had one 652 00:38:21,200 --> 00:38:25,080 final thought to leave our listeners with, what would it be? 653 00:38:25,080 --> 00:38:29,960 I think as I was writing the book, there's one quote that I came back to time and time 654 00:38:29,960 --> 00:38:34,000 again, and it's from Michael Crichton and Jurassic Park. 655 00:38:34,000 --> 00:38:38,960 And he said, your scientists were so preoccupied with whether or not they could, that they 656 00:38:38,960 --> 00:38:41,560 didn't stop to think if they should. 657 00:38:41,560 --> 00:38:45,640 And when it comes to AI, I think that we could all learn from that one still. 658 00:38:45,640 --> 00:38:50,960 I love the movie, but as we get into the idea of AI, it's like, should you be doing this 659 00:38:50,960 --> 00:38:52,960 with that data and using AI? 660 00:38:52,960 --> 00:38:55,960 And it doesn't mean no, just stop and think. 661 00:38:55,960 --> 00:38:57,240 I love Crichton's work. 662 00:38:57,240 --> 00:39:01,520 I like a lot of Crichton's thinking beyond his books. 663 00:39:01,520 --> 00:39:02,640 All right. 664 00:39:02,640 --> 00:39:05,200 So with that, let's bring this episode to an end. 665 00:39:05,200 --> 00:39:06,960 Richard, thank you so much for joining us this week. 666 00:39:06,960 --> 00:39:11,560 You and I meet somewhat regularly inside of Microsoft, so it's good to sort of interview 667 00:39:11,560 --> 00:39:13,680 you beyond the work that we do. 668 00:39:13,680 --> 00:39:15,960 So again, thank you so much for joining us. 669 00:39:15,960 --> 00:39:19,160 And to all our listeners out there, we hope you found this of use. 670 00:39:19,160 --> 00:39:21,280 Stay safe and we'll see you next time. 671 00:39:21,280 --> 00:39:24,640 Thanks for listening to the Azure Security Podcast. 672 00:39:24,640 --> 00:39:31,480 You can find show notes and other resources at our website, azsecuritypodcast.net. 673 00:39:31,480 --> 00:39:36,320 If you have any questions, please find us on Twitter at Azure Setpod. 674 00:39:36,320 --> 00:39:51,760 Background music is from ccmixtor.com and licensed under the Creative Commons license.