1 00:00:00,000 --> 00:00:05,760 Imagine treating your company's customer data not like a valuable asset, but like 2 00:00:05,760 --> 00:00:07,220 weapons-grade plutonium 3 00:00:07,220 --> 00:00:10,160 Oh, wow, that is a heavy comparison, right? 4 00:00:10,160 --> 00:00:13,180 But think about it highly toxic 5 00:00:13,180 --> 00:00:18,180 Incredibly dangerous and I mean virtually impossible to clean up once it leaks 6 00:00:18,180 --> 00:00:22,640 Yeah, you really can't just put the genie back in the bottle exactly and today 7 00:00:22,640 --> 00:00:26,500 We are exploring why sending your proprietary data to standard 8 00:00:26,500 --> 00:00:31,690 You know off-the-shelf AI tools is basically the equivalent of carrying that plutonium 9 00:00:31,690 --> 00:00:33,080 in a leaky cardboard box 10 00:00:33,080 --> 00:00:37,480 Which is terrifying for any business totally, but before we unpack the solution to 11 00:00:37,480 --> 00:00:38,940 this massive vulnerability 12 00:00:38,940 --> 00:00:43,440 Let's introduce the supporter making today's deep dive possible safe server, right? 13 00:00:43,440 --> 00:00:46,710 Because if you are running a business an association or really any kind of 14 00:00:46,710 --> 00:00:47,920 organization, you know 15 00:00:47,920 --> 00:00:52,090 The struggle relying on proprietary AI tools and cloud services from massive 16 00:00:52,090 --> 00:00:53,960 vendors like Microsoft or Google 17 00:00:53,960 --> 00:00:57,320 It often means locking yourself into an expensive black box 18 00:00:57,320 --> 00:00:57,760 Oh, yeah 19 00:00:57,760 --> 00:01:01,760 And beyond the unpredictable costs just handing over your sensitive data to these 20 00:01:01,760 --> 00:01:04,300 tech giants raises some incredibly serious 21 00:01:04,300 --> 00:01:07,000 legal regulatory and 22 00:01:07,000 --> 00:01:11,300 Compliance concerns. Yeah data sovereignty becomes critical here 23 00:01:11,300 --> 00:01:15,760 I mean when we talk about email retention financial records audit trails and strict 24 00:01:15,760 --> 00:01:17,520 data protection under the law 25 00:01:17,520 --> 00:01:21,960 Relying on those massive proprietary platforms means losing control 26 00:01:21,960 --> 00:01:25,400 You don't know where your data lives or who might be looking at it, right? 27 00:01:25,400 --> 00:01:29,450 Keeping your data on your own terms is a strict legal requirement in a lot of 28 00:01:29,450 --> 00:01:31,880 industries and that's where safe server comes in 29 00:01:31,880 --> 00:01:37,370 They help organizations replace those expensive opaque tools with secure open 30 00:01:37,370 --> 00:01:38,460 source solutions 31 00:01:38,460 --> 00:01:42,300 And you know often at a fraction of the cost. Yeah, which is a huge win 32 00:01:42,300 --> 00:01:46,870 Definitely they guide businesses from the initial consulting phase all the way 33 00:01:46,870 --> 00:01:48,320 through to operation 34 00:01:48,480 --> 00:01:52,880 Running everything on servers located right in the EU giving you that total control 35 00:01:52,880 --> 00:01:53,700 exactly 36 00:01:53,700 --> 00:01:55,940 So if you want to take back control of your infrastructure 37 00:01:55,940 --> 00:02:00,280 You can find more information at safe server dot DE and that perfectly sets up our 38 00:02:00,280 --> 00:02:01,980 mission for today. It really is 39 00:02:01,980 --> 00:02:05,810 We're looking at documentation from the cybersecurity firm confident security along 40 00:02:05,810 --> 00:02:10,100 with the github repository for an open source framework called open PCC 41 00:02:10,100 --> 00:02:15,720 So our goal here is to figure out how you can easily jump into using powerful AI 42 00:02:15,720 --> 00:02:16,300 models 43 00:02:16,480 --> 00:02:20,810 Without handing over all your confidential data. Yes, and even if you are a total 44 00:02:20,810 --> 00:02:22,140 beginner to cloud architecture 45 00:02:22,140 --> 00:02:27,630 Just stick with us by the end of this deep dive. You will understand exactly how to 46 00:02:27,630 --> 00:02:29,040 secure your AI workflows 47 00:02:29,040 --> 00:02:32,620 Let's start with the sheer scale of the liability we are dealing with 48 00:02:32,620 --> 00:02:37,060 We have to completely rethink what data actually is right because people always say 49 00:02:37,060 --> 00:02:38,740 data is the new oil exactly 50 00:02:38,740 --> 00:02:42,380 They view it as fuel but the author and activist Cory Doctorow 51 00:02:42,540 --> 00:02:48,020 He's the one who provides that striking weapons-grade plutonium quote found in our 52 00:02:48,020 --> 00:02:49,560 sources such a great analogy 53 00:02:49,560 --> 00:02:54,170 It really is he argues that personal data is dangerous. It's long-lasting and once 54 00:02:54,170 --> 00:02:55,200 it has leaked 55 00:02:55,200 --> 00:02:57,180 There is absolutely no getting it back 56 00:02:57,180 --> 00:03:02,050 Yeah, you cannot unleak a database of confidential customer information or you know 57 00:03:02,050 --> 00:03:03,680 proprietary source code that a developer 58 00:03:03,680 --> 00:03:07,140 Accidentally pasted into a public AI chatbot. Oh, man 59 00:03:07,140 --> 00:03:10,600 We've all read those horror stories somebody just trying to debug a script and 60 00:03:10,600 --> 00:03:12,940 suddenly the company's IP is in the training data 61 00:03:12,940 --> 00:03:18,440 Exactly and cryptography professor Matthew D green makes a pretty blunt observation 62 00:03:18,440 --> 00:03:20,040 about this in our source material 63 00:03:20,040 --> 00:03:21,080 What does he say? 64 00:03:21,080 --> 00:03:26,540 He basically notes that in practice if you aren't running an AI model locally on 65 00:03:26,540 --> 00:03:27,740 your own personal device 66 00:03:27,740 --> 00:03:33,660 Your alternative is to ship private data to open AI or someplace sketchier 67 00:03:33,780 --> 00:03:37,900 Where who knows what might happen to some place sketchier? Yeah, that's reassuring 68 00:03:37,900 --> 00:03:39,980 right security technologist Bruce 69 00:03:39,980 --> 00:03:44,760 Schneier refers to this exact vulnerability as the pollution problem of the 70 00:03:44,760 --> 00:03:46,700 information age the pollution problem 71 00:03:46,700 --> 00:03:51,400 Yeah, we are producing toxic runoff every time we interact with these cloud models 72 00:03:51,400 --> 00:03:55,130 Just dumping it into the digital ecosystem and hoping it doesn't poison the well 73 00:03:55,130 --> 00:03:55,560 Wow 74 00:03:55,560 --> 00:04:00,070 He argues that protecting privacy is the environmental challenge of our time and 75 00:04:00,070 --> 00:04:01,700 right now the infrastructure 76 00:04:01,700 --> 00:04:06,050 Most companies use is just fundamentally flawed. Okay, but hold on. Let's look at 77 00:04:06,050 --> 00:04:07,020 this practically for a second 78 00:04:07,020 --> 00:04:11,640 A traditional enterprise server has firewalls end-to-end encryption strict identity 79 00:04:11,640 --> 00:04:13,240 and access management roles 80 00:04:13,240 --> 00:04:16,700 If shipping data to the public cloud is so sketchy 81 00:04:16,700 --> 00:04:20,960 Shouldn't a business just self host their own AI models on their own private 82 00:04:20,960 --> 00:04:23,580 servers? That's the logical next question, right? 83 00:04:23,580 --> 00:04:27,720 Like how could a bad actor possibly dump the memory if the IT department has done 84 00:04:27,720 --> 00:04:28,500 their job? 85 00:04:29,380 --> 00:04:32,940 Configuring those firewalls and access controls. Well the source material from 86 00:04:32,940 --> 00:04:33,940 confidence security 87 00:04:33,940 --> 00:04:36,780 Explicitly addresses this assumption 88 00:04:36,780 --> 00:04:41,780 Traditional self-hosting solutions are actually insufficient for strict compliance 89 00:04:41,780 --> 00:04:45,660 Wait, really why because they still rely on human trust 90 00:04:45,660 --> 00:04:49,810 Those firewalls and access management roles you just mentioned they are 91 00:04:49,810 --> 00:04:51,180 administered by humans 92 00:04:51,180 --> 00:04:55,750 Oh in a traditional self hosted setup the root system administrator always has 93 00:04:55,750 --> 00:04:58,060 ultimate access if a bad actor 94 00:04:58,420 --> 00:05:01,690 Compromises those admin credentials or you know if the server hardware is 95 00:05:01,690 --> 00:05:04,660 physically accessed they can dump the memory straight from the RAM 96 00:05:04,660 --> 00:05:06,580 They can read the logs 97 00:05:06,580 --> 00:05:11,470 So it's a single point of failure the human element exactly from a legally binding 98 00:05:11,470 --> 00:05:13,060 compliance standpoint 99 00:05:13,060 --> 00:05:17,700 Trusting the IT department is not a measurable security metric. I mean that makes 100 00:05:17,700 --> 00:05:19,340 sense. You can't audit a promise, right? 101 00:05:19,340 --> 00:05:24,060 Traditional self-hosting relies on policies. It's essentially a sticky note on the 102 00:05:24,060 --> 00:05:26,340 server saying please don't look at this data 103 00:05:26,340 --> 00:05:30,340 Yeah, that's not gonna hold up in court. No, if an auditor comes knocking you 104 00:05:30,340 --> 00:05:33,720 cannot mathematically prove that the data remained unseen 105 00:05:33,720 --> 00:05:38,260 We have to remove human trust from the equation entirely and build a system that 106 00:05:38,260 --> 00:05:40,500 relies on math instead of policies 107 00:05:40,500 --> 00:05:43,890 Which brings us to the open source blueprint that's attempting to solve this 108 00:05:43,890 --> 00:05:44,340 problem 109 00:05:44,340 --> 00:05:49,870 Open PCC. Yes. It's a framework designed specifically for provably private AI 110 00:05:49,870 --> 00:05:50,500 inference 111 00:05:50,500 --> 00:05:54,560 It's written mostly in the go programming language and the github repository 112 00:05:54,560 --> 00:05:55,460 already boasts 113 00:05:56,140 --> 00:05:59,520 928 stars, which is pretty impressive traction 114 00:05:59,520 --> 00:06:03,590 Definitely and the architecture was heavily inspired by Apple's private cloud 115 00:06:03,590 --> 00:06:04,100 compute 116 00:06:04,100 --> 00:06:07,380 Which millions already rely on for secure AI features? 117 00:06:07,380 --> 00:06:12,210 But open PCC takes those core principles and makes them fully open auditable and 118 00:06:12,210 --> 00:06:14,060 deployable on your own infrastructure 119 00:06:14,060 --> 00:06:18,270 And people in the industry are definitely taking notice a user named abalone on 120 00:06:18,270 --> 00:06:21,140 hacker news called this server engineering 121 00:06:21,420 --> 00:06:26,330 Insanely next level insanely next level. Yeah, they compared the magnitude of this 122 00:06:26,330 --> 00:06:29,180 shift to the massive industry transition to 123 00:06:29,180 --> 00:06:34,660 Stateless architecture 30 years ago or the move to microservices 15 years ago. Wow, 124 00:06:34,660 --> 00:06:36,200 those were massive paradigm shifts 125 00:06:36,200 --> 00:06:38,900 Let's break down why abalone makes that comparison 126 00:06:38,900 --> 00:06:43,350 Because microservices completely changed how applications were built right by 127 00:06:43,350 --> 00:06:44,820 breaking massive 128 00:06:45,180 --> 00:06:49,910 Monolithic applications into small independent pieces, right? So if a one piece 129 00:06:49,910 --> 00:06:52,220 failed the whole system didn't just crash 130 00:06:52,220 --> 00:06:57,700 Right and open PCC is attempting a similar foundational shift, but for privacy it 131 00:06:57,700 --> 00:06:59,060 breaks the assumption of trust 132 00:06:59,060 --> 00:07:03,340 Exactly instead of trusting a single monolithic server and its administrator with 133 00:07:03,340 --> 00:07:04,080 all your data 134 00:07:04,080 --> 00:07:08,340 Open PCC distributes and cryptographically secures the process 135 00:07:08,340 --> 00:07:12,260 So no single entity holds the keys not even the machine's owner 136 00:07:12,260 --> 00:07:16,220 It enforces this through encrypted streaming unlinkable requests and something 137 00:07:16,220 --> 00:07:19,220 called hardware attestation. Okay, let's pause right there 138 00:07:19,220 --> 00:07:21,820 hardware attestation 139 00:07:21,820 --> 00:07:26,600 That is a very dense technical concept is yeah for the listener who doesn't have a 140 00:07:26,600 --> 00:07:27,660 background in cryptography 141 00:07:27,660 --> 00:07:30,420 Let's do an explain like I'm five breakdown 142 00:07:30,420 --> 00:07:35,410 How does hardware actually attest to something and why does that replace the need 143 00:07:35,410 --> 00:07:36,620 to trust the IT guy? 144 00:07:37,220 --> 00:07:42,700 Okay, think of hardware attestation like a digitally enforced wax seal on an 145 00:07:42,700 --> 00:07:43,220 envelope 146 00:07:43,220 --> 00:07:47,540 But built directly into the physical microchip of the server a wax seal on the 147 00:07:47,540 --> 00:07:48,300 microchip 148 00:07:48,300 --> 00:07:48,540 Okay 149 00:07:48,540 --> 00:07:50,300 so when a server boots up a 150 00:07:50,300 --> 00:07:54,170 Specialized security chip on the motherboard takes a mathematical snapshot of the 151 00:07:54,170 --> 00:07:56,340 exact code running on the machine 152 00:07:56,340 --> 00:08:00,050 It measures the operating system the applications everything like taking a 153 00:08:00,050 --> 00:08:01,940 fingerprint of the software exactly 154 00:08:01,940 --> 00:08:07,420 And if an administrator tries to secretly install malicious software to spy on your 155 00:08:07,420 --> 00:08:09,680 data that mathematical snapshot changes 156 00:08:09,680 --> 00:08:11,820 The fingerprint is different. Oh, I see 157 00:08:11,820 --> 00:08:16,300 So before your phone or your computer sends any sensitive AI prompts to that server 158 00:08:16,300 --> 00:08:21,780 It asks the server for that specific mathematical proof if the proof doesn't match 159 00:08:21,780 --> 00:08:22,780 the publicly audited 160 00:08:22,780 --> 00:08:27,940 Safe region of the code your device simply refuses to send the data Wow 161 00:08:28,260 --> 00:08:31,990 So you are no longer trusting an administrator's promise. Nope. You are verifying a 162 00:08:31,990 --> 00:08:35,940 cryptographic guarantee generated by the physical silicon itself 163 00:08:35,940 --> 00:08:38,380 That is brilliant 164 00:08:38,380 --> 00:08:42,880 Okay, so we've covered how we trust the code running on the machine 165 00:08:42,880 --> 00:08:48,360 But there is another major mechanism mentioned in the github repo called an oblivious 166 00:08:48,360 --> 00:08:50,740 HTTP relay or OH TTP 167 00:08:50,740 --> 00:08:54,550 Yes, this seems to handle how the data actually travels to the server. Let's try an 168 00:08:54,550 --> 00:08:56,620 analogy to visualize this. It's like 169 00:08:57,380 --> 00:09:01,220 Sending a highly confidential letter to a brilliant consulting detective. Okay, I 170 00:09:01,220 --> 00:09:02,100 like where this is going 171 00:09:02,100 --> 00:09:05,380 But instead of taking it yourself you give the letter to a blindfolded courier 172 00:09:05,380 --> 00:09:09,070 The courier knows where the detective's office is but has absolutely no idea what 173 00:09:09,070 --> 00:09:11,820 is written inside the letter because it's locked in a safe 174 00:09:11,820 --> 00:09:13,980 The detective receives the safe 175 00:09:13,980 --> 00:09:18,800 Opens it using a special key reads the problem and writes a solution 176 00:09:18,800 --> 00:09:23,490 But the detective has no idea who the courier works for or who originally sent the 177 00:09:23,490 --> 00:09:25,660 letter that analogy perfectly 178 00:09:25,820 --> 00:09:30,340 Isolates the mechanics of the OH TTP relay you are completely separating who is 179 00:09:30,340 --> 00:09:32,500 asking the question from what they're asking 180 00:09:32,500 --> 00:09:37,990 The who from the what exactly the relay acts as the blindfolded courier when you 181 00:09:37,990 --> 00:09:39,220 send an AI prompt 182 00:09:39,220 --> 00:09:43,750 Your IP address your identity goes to the relay but the pump itself is encrypted, 183 00:09:43,750 --> 00:09:44,060 right? 184 00:09:44,060 --> 00:09:48,340 The relay forwards the encrypted prompt to the server actually running the AI model 185 00:09:48,340 --> 00:09:49,940 that compute server decrypts 186 00:09:49,940 --> 00:09:52,380 The prompt generates the answer and sends it back 187 00:09:52,460 --> 00:09:56,870 So the compute server knows what the prompt was but it only sees the IP address of 188 00:09:56,870 --> 00:09:58,180 the relay not you 189 00:09:58,180 --> 00:10:02,180 Yes, and the relay knows who you are, but only sees encrypted gibberish 190 00:10:02,180 --> 00:10:06,850 Neither party holds the full puzzle making it impossible to link your identity to 191 00:10:06,850 --> 00:10:08,460 your proprietary data 192 00:10:08,460 --> 00:10:12,500 Okay, having an open-source framework with hardware attestation and blindfolded 193 00:10:12,500 --> 00:10:16,050 couriers is incredible for deep tech engineers who want to build custom 194 00:10:16,050 --> 00:10:16,420 infrastructure 195 00:10:16,420 --> 00:10:20,340 Oh, absolutely, but for a beginner developer or midsize business 196 00:10:20,340 --> 00:10:24,300 I mean building that from scratch requires a PhD in 197 00:10:24,300 --> 00:10:29,200 Cryptography we need an easy entry point which brings us the practical application 198 00:10:29,200 --> 00:10:30,060 of this framework 199 00:10:30,060 --> 00:10:35,680 Our sources introduce a managed service called cone FSC operated by a firm named 200 00:10:35,680 --> 00:10:36,740 confident security 201 00:10:36,740 --> 00:10:41,460 All right. This service is built entirely on the open source open PCC standard 202 00:10:41,460 --> 00:10:44,180 operating under a core philosophy of four words 203 00:10:44,180 --> 00:10:48,860 Don't trust verify don't trust verify. I love that 204 00:10:48,860 --> 00:10:52,420 Yeah, and they detail specific technical guarantees to back up that philosophy 205 00:10:52,420 --> 00:10:55,780 because they utilize the open PCC framework 206 00:10:55,780 --> 00:10:59,440 They offer zero logging wait zero logging at all zero 207 00:10:59,440 --> 00:11:02,500 And that doesn't mean they promise to delete your logs at the end of the day 208 00:11:02,500 --> 00:11:06,550 It means the system architecture literally prevents data from being logged in the 209 00:11:06,550 --> 00:11:08,320 first place. This is a huge distinction 210 00:11:08,320 --> 00:11:12,280 It is your prompts are never used for AI training and they are never sent to third 211 00:11:12,280 --> 00:11:13,420 parties and most crucially 212 00:11:13,420 --> 00:11:17,860 Even the operator of the server does not have privilege access to the private 213 00:11:17,860 --> 00:11:18,500 computation 214 00:11:18,500 --> 00:11:22,480 Okay, let's dig into that operator lockout because this directly addresses our 215 00:11:22,480 --> 00:11:24,860 earlier discussion about the flaws of self-hosting 216 00:11:24,860 --> 00:11:29,240 Yeah, if confident security physically owns the server in their data center 217 00:11:29,240 --> 00:11:34,830 How are they physically locked out of reading the data processing on their own 218 00:11:34,830 --> 00:11:35,400 machine? 219 00:11:35,400 --> 00:11:39,640 It comes down to secure enclaves within the processor itself 220 00:11:39,640 --> 00:11:43,100 When your encrypted prompt reaches the server 221 00:11:43,100 --> 00:11:46,990 It isn't decrypted in the standard open memory of the computer where an 222 00:11:46,990 --> 00:11:48,580 administrator could see it 223 00:11:48,580 --> 00:11:49,300 Where does it go? 224 00:11:49,300 --> 00:11:54,100 It is routed into a heavily isolated section of the CPU called an enclave 225 00:11:54,100 --> 00:11:58,180 You can think of it as an impenetrable black box built into the silicon 226 00:11:58,180 --> 00:12:00,980 So the data is decrypted inside that black box 227 00:12:00,980 --> 00:12:01,300 Yes 228 00:12:01,300 --> 00:12:05,940 The AI model generates the response inside that black box and the response is 229 00:12:05,940 --> 00:12:07,340 encrypted before it ever leaves 230 00:12:07,340 --> 00:12:11,550 Incredible. So if the server operator dumps the machine's RAM or even attaches a 231 00:12:11,550 --> 00:12:14,620 physical pro to the motherboard to spy on the data in 232 00:12:14,620 --> 00:12:18,110 Transit like physically hacking the machine exactly all they will capture is 233 00:12:18,110 --> 00:12:18,780 encrypted noise 234 00:12:18,780 --> 00:12:23,010 The administrator of the operating system is entirely blind to what is happening 235 00:12:23,010 --> 00:12:24,080 inside the enclave 236 00:12:24,080 --> 00:12:27,760 You know from a development standpoint not having to rewrite an entire application 237 00:12:27,760 --> 00:12:33,240 Just to integrate a new security standard saves months of engineering time. Oh 238 00:12:33,240 --> 00:12:33,700 without a doubt 239 00:12:33,700 --> 00:12:36,820 This was a striking detail in the confidence security documentation 240 00:12:37,980 --> 00:12:42,040 they provide an open AI compatible API and SDK a 241 00:12:42,040 --> 00:12:46,440 Developer doesn't have to learn a completely new protocol to use this providing a 242 00:12:46,440 --> 00:12:49,460 standard interface to leading large language models 243 00:12:49,460 --> 00:12:55,380 Drastically lowers the barrier to entry you simply swap your existing endpoint URL 244 00:12:55,380 --> 00:12:58,460 and your API key to visualize that API swap 245 00:12:58,460 --> 00:13:01,180 It's like having a freight train carrying sensitive cargo 246 00:13:01,180 --> 00:13:04,530 You don't need to rebuild the train the tracks or the cargo from scratch to make it 247 00:13:04,530 --> 00:13:05,340 secure, right? 248 00:13:05,340 --> 00:13:09,070 You just flip a digital switch on the tracks routing the exact same train into a 249 00:13:09,070 --> 00:13:10,900 highly secure verified vault 250 00:13:10,900 --> 00:13:13,740 Instead of an open warehouse. That's a great way to put it 251 00:13:13,740 --> 00:13:17,360 The infrastructure does the heavy lifting while your application continues 252 00:13:17,360 --> 00:13:21,560 functioning exactly as it did before and beyond accessing leading LL M's 253 00:13:21,560 --> 00:13:25,770 The documentation notes that users can host manage and sell their own custom models 254 00:13:25,770 --> 00:13:28,760 with those exact same verifiable privacy guarantees 255 00:13:28,760 --> 00:13:31,660 Yes, and this brings us to the cost factor 256 00:13:32,560 --> 00:13:37,190 proprietary black box AI from major vendors is notorious for unpredictable billing 257 00:13:37,190 --> 00:13:37,960 structures 258 00:13:37,960 --> 00:13:41,260 Oh, tell me about it. The bills can just skyrocket overnight, right? 259 00:13:41,260 --> 00:13:45,970 Confident security tackles that by offering transparent pricing where you pay only 260 00:13:45,970 --> 00:13:46,740 for what you use 261 00:13:46,740 --> 00:13:50,700 With base fees pinned to current market prices per model 262 00:13:50,700 --> 00:13:54,900 So organizations aren't forced to pay a massive premium just to secure their data 263 00:13:54,900 --> 00:13:59,400 Exactly, they get the state-of-the-art security standard without price gouging. Let's 264 00:13:59,400 --> 00:14:01,980 transition to the ultimate application of all this 265 00:14:02,180 --> 00:14:07,080 We've mapped out the architecture the secure enclaves and the easy API swap 266 00:14:07,080 --> 00:14:11,970 But for modern businesses the biggest hurdle to adopting AI is navigating the 267 00:14:11,970 --> 00:14:14,580 massive legal headaches around data compliance 268 00:14:14,580 --> 00:14:18,710 Oh, absolutely. Those legal hurdles are defined by strict regulations like GDPR in 269 00:14:18,710 --> 00:14:20,620 Europe CCPA in California 270 00:14:20,620 --> 00:14:24,540 HIPAA and the healthcare sector and the fines for messing those up are no joke 271 00:14:24,540 --> 00:14:27,020 severe penalties for mishandling personal data 272 00:14:27,220 --> 00:14:31,610 But by utilizing verifiable privacy where an organization can mathematically prove 273 00:14:31,610 --> 00:14:34,840 that data is encrypted unseen and unlogged 274 00:14:34,840 --> 00:14:40,450 Businesses can finally leverage powerful AI models on private data while remaining 275 00:14:40,450 --> 00:14:41,880 strictly compliant 276 00:14:41,880 --> 00:14:44,340 Let's put this into a real-world scenario 277 00:14:44,340 --> 00:14:49,830 Imagine an auditor walks into a hospital's IT room to verify hyper a compliance 278 00:14:49,830 --> 00:14:52,700 regarding a new AI diagnostic tool 279 00:14:52,700 --> 00:14:57,700 Okay stressful day for the IT guy right in a traditional setup that involves weeks 280 00:14:57,700 --> 00:14:59,200 of pulling server logs 281 00:14:59,200 --> 00:15:04,620 interviewing IT staff about access controls reviewing I am policies and ultimately 282 00:15:04,620 --> 00:15:04,620 just 283 00:15:04,620 --> 00:15:08,930 Hoping no internal staff member accidentally left the database exposed. It's a 284 00:15:08,930 --> 00:15:09,740 total nightmare 285 00:15:09,740 --> 00:15:14,020 But with verifiable privacy through a framework like open PCC 286 00:15:14,020 --> 00:15:18,220 What actually happens during that audit the audit transforms from a procedural 287 00:15:18,220 --> 00:15:20,580 nightmare into a mathematical certainty? 288 00:15:20,700 --> 00:15:24,410 The auditor doesn't need to interview the IT staff or comb through thousands of 289 00:15:24,410 --> 00:15:26,840 lines of access logs. Really just skip all that 290 00:15:26,840 --> 00:15:27,060 Yeah 291 00:15:27,060 --> 00:15:30,920 They simply verify the cryptographic signature of the hardware attestation in a 292 00:15:30,920 --> 00:15:32,100 matter of seconds 293 00:15:32,100 --> 00:15:35,390 They can run a mathematical proof confirming that the server is running the audited 294 00:15:35,390 --> 00:15:38,120 code and that the secure enclaves are active 295 00:15:38,120 --> 00:15:42,290 the proof demonstrates that no human not even the system administrator could 296 00:15:42,290 --> 00:15:44,060 possibly have read the patient data as 297 00:15:44,340 --> 00:15:49,060 The documentation concisely puts it this technology provides peace of mind for the 298 00:15:49,060 --> 00:15:51,700 business and a piece of cake for the auditors 299 00:15:51,700 --> 00:15:53,340 A piece of cake for the auditors 300 00:15:53,340 --> 00:15:58,030 I bet they'd love that a study referenced in the sources by Mazuma Hassan, Andrei 301 00:15:58,030 --> 00:16:00,200 Kushnaruk and Elizabeth Boricki 302 00:16:00,200 --> 00:16:03,100 Emphasizes this exact point. Oh, yes 303 00:16:03,100 --> 00:16:07,620 They noted that integrating privacy by design technologies into AI applications 304 00:16:07,620 --> 00:16:12,360 Could mitigate the massive challenges of adopting AI and healthcare and healthcare 305 00:16:12,360 --> 00:16:13,980 is the ultimate stress test 306 00:16:14,380 --> 00:16:17,500 He really is patient records are the most sensitive plutonium 307 00:16:17,500 --> 00:16:22,150 There is if we can solve AI privacy for healthcare using these zero-knowledge 308 00:16:22,150 --> 00:16:23,100 environments 309 00:16:23,100 --> 00:16:28,030 We can solve it for banking legal human resources everything and the sources frame 310 00:16:28,030 --> 00:16:29,420 this level of privacy 311 00:16:29,420 --> 00:16:33,860 Not as a luxury add-on but as an essential baseline for the future of the Internet 312 00:16:33,860 --> 00:16:37,880 I mean it should be Gary Kovacs states in the materials that security and privacy 313 00:16:37,880 --> 00:16:41,100 guarantees are strongest when they're entirely technically 314 00:16:41,540 --> 00:16:47,020 Forcible it shouldn't rely on a company's goodwill or a complex legal contract, 315 00:16:47,020 --> 00:16:47,060 right? 316 00:16:47,060 --> 00:16:51,440 Because goodwill changes when profits drop exactly it must be baked directly into 317 00:16:51,440 --> 00:16:52,900 the code and the silicon and 318 00:16:52,900 --> 00:16:56,900 Venture capitalist Fred Wilson predicts that the company's doing the best job 319 00:16:56,900 --> 00:17:00,520 managing user privacy will ultimately become the most successful 320 00:17:00,520 --> 00:17:04,140 Turning privacy into a core competitive advantage 321 00:17:04,140 --> 00:17:10,180 Marco Altea at toast ring AI captures the underlying philosophy perfectly 322 00:17:10,420 --> 00:17:14,200 He argues that privacy is not an option and that it shouldn't be the price we 323 00:17:14,200 --> 00:17:15,840 accept for just getting on the internet 324 00:17:15,840 --> 00:17:19,170 That's a powerful statement. It is we shouldn't have to surrender our right to 325 00:17:19,170 --> 00:17:22,680 digital privacy or expose our company's intellectual property 326 00:17:22,680 --> 00:17:28,170 Just to participate in the modern AI driven economy tools like open PCC provide the 327 00:17:28,170 --> 00:17:30,340 technical means to finally refuse that trade-off 328 00:17:30,340 --> 00:17:35,360 We are witnessing a necessary transition from an era of security by policy to an 329 00:17:35,360 --> 00:17:38,160 era of security by mathematics and architecture 330 00:17:38,980 --> 00:17:43,270 Beautifully said let's briefly recap the journey we've taken today. We started with 331 00:17:43,270 --> 00:17:46,640 the reality of data as weapons grade plutonium 332 00:17:46,640 --> 00:17:49,240 toxic permanent and 333 00:17:49,240 --> 00:17:53,020 Constantly being leaked into opaque proprietary cloud models, right? 334 00:17:53,020 --> 00:17:57,270 We broke down why traditional self-hosting and firewalls fall short because they 335 00:17:57,270 --> 00:18:00,140 still rely on vulnerable human administrators 336 00:18:00,140 --> 00:18:05,630 We human elements exactly then we explored the open PCC framework discovering how 337 00:18:05,630 --> 00:18:08,280 hardware attestation acts as a digital wax seal 338 00:18:08,280 --> 00:18:14,680 And how oblivious HTTP relays function as blindfolded couriers to separate identity 339 00:18:14,680 --> 00:18:15,260 from data 340 00:18:15,260 --> 00:18:19,800 Which is such a massive leap forward it is and we saw how accessible this has 341 00:18:19,800 --> 00:18:22,100 become through managed services like confident securities 342 00:18:22,100 --> 00:18:26,870 Kind of SCC where securing an application is as simple as flipping a switch on the 343 00:18:26,870 --> 00:18:27,500 API tracks 344 00:18:27,500 --> 00:18:31,140 Utilizing secure CPU enclaves to guarantee operator lockout 345 00:18:31,140 --> 00:18:33,140 It provides the master key for compliance 346 00:18:33,700 --> 00:18:37,600 Allowing organizations to navigate strict regulations like GDPR and hyper through 347 00:18:37,600 --> 00:18:41,710 mathematically enforceable proofs all without sacrificing the efficiency of 348 00:18:41,710 --> 00:18:42,240 artificial intelligence 349 00:18:42,240 --> 00:18:46,940 Or paying exorbitant premiums, which brings us perfectly back to the supporter of 350 00:18:46,940 --> 00:18:49,060 today's deep dive safe server 351 00:18:49,060 --> 00:18:53,580 If the capabilities we just outlined true data sovereignty 352 00:18:53,580 --> 00:18:56,220 mathematically enforceable compliance 353 00:18:56,220 --> 00:19:00,210 Predictable cost structures and protection from massive cloud vendor lock-in if 354 00:19:00,210 --> 00:19:02,500 that aligns with your organization's needs 355 00:19:03,060 --> 00:19:05,720 Safe server is the solution. Yeah, absolutely 356 00:19:05,720 --> 00:19:10,340 Whether you are a business an association or any group looking to replace expensive 357 00:19:10,340 --> 00:19:11,940 opaque AI tools 358 00:19:11,940 --> 00:19:16,210 They provide the necessary expertise safe server is really a partner in your 359 00:19:16,210 --> 00:19:17,220 infrastructure 360 00:19:17,220 --> 00:19:21,050 They can be commissioned for specialized consulting to help find and implement the 361 00:19:21,050 --> 00:19:24,440 exact open source solution for your specific needs, right? 362 00:19:24,440 --> 00:19:28,630 So whether the perfect fit is the open PCC software we explore today or a 363 00:19:28,630 --> 00:19:30,620 comparable open source alternative 364 00:19:30,620 --> 00:19:34,220 They guide you from the planning phase all the way through to secure operation on 365 00:19:34,220 --> 00:19:36,020 servers located right in the EU 366 00:19:36,020 --> 00:19:39,820 You can learn more and take back control of your data at safe server de 367 00:19:39,820 --> 00:19:44,110 Before we wrap up though. There is a final thought. I want to leave you with to mull 368 00:19:44,110 --> 00:19:45,740 over. Okay, let's hear it 369 00:19:45,740 --> 00:19:49,940 We began by discussing the pollution problem of the information age, you know 370 00:19:49,940 --> 00:19:54,100 The toxic runoff of data we leave behind when using standard AI platforms 371 00:19:54,100 --> 00:19:58,990 Right if mathematically enforceable zero-knowledge AI architectures become the new 372 00:19:58,990 --> 00:19:59,480 default 373 00:20:00,100 --> 00:20:05,590 What happens to the massive tech empires build entirely on harvesting and monetizing 374 00:20:05,590 --> 00:20:06,920 our data plutonium? 375 00:20:06,920 --> 00:20:08,620 Oh, that is a fascinating question 376 00:20:08,620 --> 00:20:12,250 If we stop providing the toxic runoff could the information age finally clean up 377 00:20:12,250 --> 00:20:15,340 his pollution problem and force those empires to find a completely 378 00:20:15,340 --> 00:20:19,120 New business model. It's something to think about keep questioning keep exploring 379 00:20:19,120 --> 00:20:20,740 and we'll see you the next deep dive