WEBVTT 1 00:00:00.320 --> 00:00:03.200 Hey, Carl. Here, we're on a mission to find new 2 00:00:03.240 --> 00:00:05.919 fans of the show. I mean, where else can you 3 00:00:05.960 --> 00:00:10.240 get real life comedy and horror in one podcast. You 4 00:00:10.279 --> 00:00:12.960 can do your part by leaving us a review on iTunes, 5 00:00:13.039 --> 00:00:16.719 Spotify or wherever you get Security this Week. Also, you 6 00:00:16.719 --> 00:00:19.079 can get an ad free feed by becoming a five 7 00:00:19.120 --> 00:00:22.039 dollars a month patron. Go to Patreon dot Security this 8 00:00:22.079 --> 00:00:25.320 week dot com and sign up, and don't forget Discord. 9 00:00:25.600 --> 00:00:28.519 Discord dot Security this week dot com. Lots of good 10 00:00:28.559 --> 00:00:32.320 stuff happening there. Yeah, I think that's it. So on 11 00:00:32.359 --> 00:00:35.240 with the podcast. Hey, did you guys know that I 12 00:00:35.280 --> 00:00:36.880 took a workshop with Melbrooks? 13 00:00:37.079 --> 00:00:37.200 No? 14 00:00:37.560 --> 00:00:39.320 Yeah, comedy workshop. How do you think we got all 15 00:00:39.320 --> 00:00:44.000 those Monday's episodes? No? Basically I was sitting down with 16 00:00:44.039 --> 00:00:48.759 Melbrooks and he told me the secret to comedy. Yeah, 17 00:00:48.880 --> 00:00:51.240 Dwayne asked me, what's the secret to comic? Go ahead? 18 00:00:51.320 --> 00:00:55.039 Yeah, absolutely, what's the what's the secret timing? I don't 19 00:00:55.079 --> 00:00:55.439 get it? 20 00:00:55.880 --> 00:00:56.600 He that is Carl. 21 00:01:05.400 --> 00:01:07.280 Okay, we got a lot of stuff to get through 22 00:01:07.319 --> 00:01:10.400 today on Security This Week, I'm Carl Franklin, It's Dwayne 23 00:01:10.439 --> 00:01:15.439 Leflotte and Patrick Hines. Let's start with identity crisis, novel 24 00:01:15.640 --> 00:01:20.959 vulnerabilities leading to Kerberos downgrade, DOS, denial of service, and 25 00:01:21.040 --> 00:01:23.680 full domain takeover. This one's awesome. 26 00:01:23.719 --> 00:01:26.680 So this is the new new thing that's happening. We're 27 00:01:26.680 --> 00:01:30.079 going to talk a lot about the identity as a place, 28 00:01:30.120 --> 00:01:35.120 that place that gets attacked, and because that's the place 29 00:01:35.159 --> 00:01:39.640 that matters the most, we implemented two factor authentication because 30 00:01:39.799 --> 00:01:43.799 passwords weren't strong enough. Regular authentication was falling down. So 31 00:01:43.840 --> 00:01:45.799 now this is the logical place where the hackers have 32 00:01:45.840 --> 00:01:48.040 to attack, because otherwise what are they going to do, 33 00:01:48.359 --> 00:01:49.760 like just consider things secure? 34 00:01:49.879 --> 00:01:51.879 Yeah, once they have your identity, it's game over. 35 00:01:51.959 --> 00:01:53.920 Yeah, because now they can they can log in, right. 36 00:01:54.000 --> 00:01:55.760 This one's so good though. So good. 37 00:01:56.599 --> 00:01:59.519 Directory is something that we've kind of moved away whatnot, 38 00:01:59.560 --> 00:02:02.920 kind of dramatically moved away from, and many clients moved 39 00:02:02.920 --> 00:02:07.680 away from active directory is great, but it makes things convenient, 40 00:02:08.039 --> 00:02:09.879 so which is really a problem. 41 00:02:10.120 --> 00:02:13.560 Yeah, I hear what you're saying, but from a corporate standpoint, 42 00:02:14.439 --> 00:02:16.240 one hundred percent of the companies out there aren't moving 43 00:02:16.240 --> 00:02:18.639 away from active directory. They're just not doing it. It's 44 00:02:18.960 --> 00:02:21.199 too much of a pain in the ass. Normal users 45 00:02:21.199 --> 00:02:23.280 aren't gonna understand how to Well, I get to log 46 00:02:23.360 --> 00:02:25.680 into my local box with a user and passer that's 47 00:02:25.719 --> 00:02:28.080 not used anywhere else and has access to nothing, and 48 00:02:28.120 --> 00:02:30.520 then every resource I need to access, I'm gonna use 49 00:02:30.520 --> 00:02:33.039 some sort of password management. Like nobody's gonna do it, right, 50 00:02:33.080 --> 00:02:35.639 She's not gonna happen, right, And there are all sorts 51 00:02:35.680 --> 00:02:37.560 of other ways you can do. You can implement things 52 00:02:37.560 --> 00:02:38.439 like zero trust. 53 00:02:38.560 --> 00:02:41.520 Yeah, I was gonna say, Dane, what is the answer 54 00:02:41.639 --> 00:02:45.919 to shoring up active directory in the cloud for these companies. 55 00:02:46.560 --> 00:02:49.080 Should have been built better? Maybe I don't know. I mean, 56 00:02:49.080 --> 00:02:53.120 the problem is it's it's very active directory. Is a 57 00:02:53.280 --> 00:02:57.039 Patrick trite is convenient, right, but it's it's also been 58 00:02:57.080 --> 00:03:02.400 around longer than I don't know, twenty five thirty years, 59 00:03:03.120 --> 00:03:06.800 well before just before it was stolen from Novelle and 60 00:03:06.919 --> 00:03:10.360 Microsoft Ticket so so anyways, it's been around a long time. 61 00:03:10.960 --> 00:03:12.879 So the problem with that is you keep just adding 62 00:03:12.919 --> 00:03:14.520 on to it and adding on to it and adding 63 00:03:14.520 --> 00:03:16.120 on to it and adding on to it, and there's 64 00:03:16.159 --> 00:03:19.439 so much technical debt in there. But unfortunately, you know, 65 00:03:19.520 --> 00:03:21.840 thirty years of companies are now built on top of it, 66 00:03:22.000 --> 00:03:23.159 so you can't just rip it out. 67 00:03:23.199 --> 00:03:26.680 I mean it's it's literally single sign on, which we've 68 00:03:27.120 --> 00:03:30.360 decided is a bad thing. Single sign on is not 69 00:03:30.840 --> 00:03:33.639 a good thing. We need zero trust. Do you see 70 00:03:33.680 --> 00:03:37.280 active directory participating in zero trust? Now? It's the athenma 71 00:03:37.319 --> 00:03:37.879 to it, isn't it? 72 00:03:38.199 --> 00:03:41.080 With all these AI tools, couldn't Microsoft just whip up 73 00:03:41.120 --> 00:03:44.280 its genie AI guys and say, hey, we want you 74 00:03:44.319 --> 00:03:47.319 to rebuild active directory from the bottom up, using the 75 00:03:47.360 --> 00:03:50.199 same interface to the customer, but make it more secure 76 00:03:50.240 --> 00:03:52.360 and take out all the baroque bs that we don't 77 00:03:52.360 --> 00:03:53.039 need anymore. 78 00:03:53.120 --> 00:03:55.520 Sure, yeah, I think you could. I think you could. 79 00:03:55.560 --> 00:03:57.360 I think it takes time. I don't know. I don't 80 00:03:57.360 --> 00:04:00.319 know that anybody would trust, you know, vibe code up 81 00:04:00.319 --> 00:04:01.280 active directory yet. 82 00:04:01.319 --> 00:04:03.919 Well no, I don't mean vibe coding, I mean assisted coding. 83 00:04:04.120 --> 00:04:06.319 I could see that. I could see that. But this so, 84 00:04:06.439 --> 00:04:10.120 this particular case, the way that this was exploited, Right, 85 00:04:10.159 --> 00:04:11.960 we talked about denial of service and all sorts of 86 00:04:12.000 --> 00:04:14.520 other things. The way this is exploited is you have 87 00:04:15.159 --> 00:04:20.240 service principal names in active directory, and those SPNs you know, 88 00:04:20.279 --> 00:04:22.959 give you access to resources and all sorts of weird stuff, right, 89 00:04:24.279 --> 00:04:27.399 and they're supposed to be unique, right, it'd be weird 90 00:04:27.439 --> 00:04:30.600 if my account resolved to my account and Patrick's account 91 00:04:30.600 --> 00:04:33.639 at the same time. And what a researcher found is 92 00:04:33.680 --> 00:04:36.199 there's I'd have to look back at the article to 93 00:04:36.240 --> 00:04:38.160 see the number of it's like three hundred and eighty 94 00:04:38.160 --> 00:04:43.040 two unprincipal Unicode characters. They're just they're just they look 95 00:04:43.160 --> 00:04:47.279 like spaces. And of those, there's like one hundred and 96 00:04:47.360 --> 00:04:52.000 twenty six of them that active directory visa vl tapp 97 00:04:53.800 --> 00:04:58.399 doesn't actually recognize but are uniquely different characters. 98 00:04:58.399 --> 00:04:59.399 But some of them they do. 99 00:04:59.519 --> 00:05:02.360 So I can take Patrick's account and I can add 100 00:05:02.519 --> 00:05:06.120 one of these characters at the end, and an active 101 00:05:06.160 --> 00:05:08.199 director will go, yeah, this is a unique object. But 102 00:05:08.240 --> 00:05:12.759 then when it goes to verify, right, oh, does Dwayne 103 00:05:12.839 --> 00:05:15.279 have access to this account? It goes, well, I got 104 00:05:15.279 --> 00:05:19.360 two users here, er uh, and then denial of service 105 00:05:19.399 --> 00:05:23.279 happens because I was able to make my you know, 106 00:05:23.839 --> 00:05:27.000 for lack of a better term, my unique identifier in 107 00:05:27.120 --> 00:05:30.560 active directory look exactly like Patrick's for the filtering it 108 00:05:30.639 --> 00:05:34.079 uses for certain functions. Wow, so this is it's a 109 00:05:34.680 --> 00:05:37.639 it is a novel attack. There's three different attacks, and 110 00:05:37.680 --> 00:05:41.160 here one of them is a denial of service. One 111 00:05:41.160 --> 00:05:44.879 of them is a pre vesque, uh, where you can 112 00:05:45.000 --> 00:05:48.079 just yeah, you can jump straight to DA. And that's 113 00:05:48.120 --> 00:05:50.079 what we're usually looking for, is right, can you get 114 00:05:50.079 --> 00:05:53.839 to DA and how quickly can you get there? So this, 115 00:05:53.839 --> 00:05:56.879 this is an interesting find, honestly, this this I really 116 00:05:56.920 --> 00:05:59.319 love these clever kind of finds. 117 00:05:59.360 --> 00:05:59.519 There. 118 00:05:59.560 --> 00:06:02.399 There was a while back actually print Nightmare was one 119 00:06:02.439 --> 00:06:04.120 of the ones that was just a really kind of 120 00:06:04.120 --> 00:06:08.240 interesting cleverifying right. There was another one in the past 121 00:06:08.279 --> 00:06:11.800 where any any domain user could go and add five 122 00:06:11.920 --> 00:06:14.720 computers for some random reason to active directory and there 123 00:06:14.759 --> 00:06:17.439 was all sorts of ways to manipulate that that so 124 00:06:17.480 --> 00:06:20.959 there's a lot of cool exploits that can go give 125 00:06:20.959 --> 00:06:23.639 you privileged escalation or denial a service or something along 126 00:06:23.720 --> 00:06:26.160 those lines. And this, this actually is really kind of unique. 127 00:06:26.319 --> 00:06:29.839 It's not being seen in the wild right now. This 128 00:06:29.839 --> 00:06:33.839 this researcher obviously, you know, announced it to Microsoft. Microsoft 129 00:06:33.920 --> 00:06:38.240 has patched it, so if you're running the latest patches 130 00:06:38.240 --> 00:06:42.879 on Microsoft, you are good here. But yeah, no, it's kudos. 131 00:06:43.120 --> 00:06:44.480 That's a that's a good find. 132 00:06:45.040 --> 00:06:47.759 And this is a little bit of a precursor some 133 00:06:47.839 --> 00:06:49.519 other things we're going to talk about Oh. 134 00:06:49.480 --> 00:06:54.959 Yeah, yeah. So if identity is the biggest target, Hacker 135 00:06:55.000 --> 00:06:58.800 News says device cod fishing is the fastest growing threat 136 00:06:58.879 --> 00:07:01.319 of twenty twenty six. And here are six reasons why. 137 00:07:01.560 --> 00:07:04.920 Device code phishing, device code and so is that like 138 00:07:04.959 --> 00:07:08.959 when you know you're using two FA and somebody sends 139 00:07:08.959 --> 00:07:11.279 a text to you with a code that you have 140 00:07:11.319 --> 00:07:13.560 to enter that kind of phishing? 141 00:07:13.759 --> 00:07:17.480 Uh yeah, absolutely, yeah, it's uh and it's it's uh, 142 00:07:18.079 --> 00:07:20.120 you know you what, We've all done this, right, yeah, 143 00:07:20.240 --> 00:07:21.839 And we talked about this, i think last week because 144 00:07:21.839 --> 00:07:26.759 there was another device fishing sort of token based attack 145 00:07:26.839 --> 00:07:30.040 we were talking about. You go to Netflix on your 146 00:07:30.079 --> 00:07:31.879 TV and it says, hey, do you want to log 147 00:07:31.920 --> 00:07:34.439 in with your username and really long password because hopefully 148 00:07:34.480 --> 00:07:36.839 you're using a password manager m M. You go, nah, 149 00:07:36.839 --> 00:07:38.560 I don't want to do that, and it goes okay, cool. 150 00:07:38.600 --> 00:07:41.160 Then just log in on your laptop or on your 151 00:07:41.199 --> 00:07:46.920 phone and type this code in right to Netflix. You go, okay, cool, Right, 152 00:07:46.959 --> 00:07:49.120 and you go to you go to Netflix and you 153 00:07:49.160 --> 00:07:51.759 go to the slash ad device or whatever it is, 154 00:07:51.839 --> 00:07:54.160 and you type in the code right, And what you've 155 00:07:54.439 --> 00:08:00.759 just done is you have authorized the TV to gain 156 00:08:00.920 --> 00:08:05.319 access to your account, right, right, So you've passed the 157 00:08:05.439 --> 00:08:09.120 authentication stage. You've already authenticated. I used my username and 158 00:08:09.120 --> 00:08:12.399 password on my phone, I used my multi factor authentication I. 159 00:08:12.439 --> 00:08:15.000 Used and you goun to bridge it to this other device, I. 160 00:08:14.959 --> 00:08:18.279 Used my my you know, secure keys, whatever. And once 161 00:08:18.319 --> 00:08:21.639 I got access to whatever that device is, whatever that 162 00:08:21.759 --> 00:08:24.879 service is, I then gave another device access by just 163 00:08:24.920 --> 00:08:26.920 typing in six little, six little digits. 164 00:08:26.959 --> 00:08:28.000 Yeah, I mean, I get it. 165 00:08:28.120 --> 00:08:30.199 This isn't quite as bad, but it's it's kind of 166 00:08:30.240 --> 00:08:33.919 like the later evolution of you go over to your 167 00:08:34.000 --> 00:08:36.279 router and you hit a button and there's a four 168 00:08:36.320 --> 00:08:39.919 digit number that you can enter and join your Wi Fi. 169 00:08:40.600 --> 00:08:40.759 Right. 170 00:08:40.799 --> 00:08:44.840 Well, hackers love that because it dramatically weakens the security 171 00:08:44.840 --> 00:08:46.879 of having to provide the Wi Fi password. Yeah, and 172 00:08:46.919 --> 00:08:49.799 in this case, like in the Wi Fi situation, WPS 173 00:08:49.840 --> 00:08:52.360 like we may brute force that, right, there's only so 174 00:08:52.399 --> 00:08:55.159 many codes in this particular case, you don't even need 175 00:08:55.200 --> 00:08:56.480 to brute force it. All you need to do is 176 00:08:56.480 --> 00:08:59.919 con somebody into taking those you know, six little DIDs 177 00:09:00.480 --> 00:09:04.080 and typing it into the account. And what's super interesting 178 00:09:04.200 --> 00:09:07.519 is when these attacks are executed against you know, mirror 179 00:09:07.519 --> 00:09:12.440 mortals right, who aren't like digging into Well, I wasn't. 180 00:09:12.279 --> 00:09:14.799 Actually logging in. Why would I ever type this code in? 181 00:09:14.919 --> 00:09:15.440 Whatever it is? 182 00:09:16.399 --> 00:09:18.519 But at the social engineering level, when you click on 183 00:09:18.559 --> 00:09:20.919 the link in an email that comes in or a 184 00:09:21.000 --> 00:09:24.240 tax that comes in, or whatever, the website you go 185 00:09:24.320 --> 00:09:30.559 to is legitimately Microsoft or legitimately Google or legitimately Netflix 186 00:09:30.679 --> 00:09:34.080 or whatever it is. So your password manager is not 187 00:09:34.120 --> 00:09:35.720 going to flag it. It's not going to say, oh, 188 00:09:35.759 --> 00:09:37.679 I shouldn't type in the username password here because you 189 00:09:37.960 --> 00:09:38.879 are at Microsoft. 190 00:09:38.960 --> 00:09:41.000 But if I go, if I send a message to 191 00:09:41.080 --> 00:09:44.039 somebody who's not on guard, who's not savvy to what's 192 00:09:44.080 --> 00:09:46.679 going on, and I say, oh, you know, your Netflix 193 00:09:46.759 --> 00:09:50.480 is coming up for a review and we'd like to 194 00:09:50.519 --> 00:09:54.279 give you a deeper discount. Sure, can you validate that 195 00:09:54.320 --> 00:09:57.480 you're still using this account? Here? Go log in and 196 00:09:57.600 --> 00:10:00.120 use these six characters to validate it. 197 00:10:00.159 --> 00:10:02.360 And then you redirect them and then you give them 198 00:10:02.399 --> 00:10:05.320 the link that brings them straight to like Netflix dot com, 199 00:10:05.360 --> 00:10:08.200 slash you know, TV or whatever link I think is that, 200 00:10:08.320 --> 00:10:10.960 or TV or whatever, and they go, okay, cool, click 201 00:10:10.960 --> 00:10:12.399 on it, and then they do the log in thing 202 00:10:12.440 --> 00:10:14.960 and then they just type in the code and click okay, right, 203 00:10:15.039 --> 00:10:18.399 and it will say we've added a device to your account. 204 00:10:18.399 --> 00:10:21.519 But most people aren't even reading that deeply. So and 205 00:10:21.600 --> 00:10:23.279 a lot of times what will happen is on the 206 00:10:23.360 --> 00:10:27.159 back end, the attackers have automated systems waiting for access, 207 00:10:27.360 --> 00:10:30.639 and the moment you give them access, they run. It starts, 208 00:10:30.799 --> 00:10:33.360 It starts adding different devices, it starts pulling down in 209 00:10:33.399 --> 00:10:36.960 from it. I mean, faster than you can deny access 210 00:10:37.000 --> 00:10:37.759 to that account. 211 00:10:38.039 --> 00:10:40.240 Well, and most people don't know how to deny it. 212 00:10:40.320 --> 00:10:43.159 Apparently. This is number one is it defeats every form 213 00:10:43.200 --> 00:10:46.159 of MFA, including passkeys. But number two in this article 214 00:10:46.440 --> 00:10:51.440 is that the fast ecosystem has fully industrialized it. In 215 00:10:51.480 --> 00:10:55.080 other words, it's it's standard practice. 216 00:10:54.600 --> 00:10:57.360 Now oh yeah, yeah yeah, and for you convenient. The 217 00:10:57.519 --> 00:11:00.559 fast ecosystem, by the way, is the fishing as service, 218 00:11:01.360 --> 00:11:03.480 So you know there are a lot of frameworks. Heck, 219 00:11:03.519 --> 00:11:05.039 there's a lot of services out there. I used to 220 00:11:05.080 --> 00:11:07.879 talk about this, especially on lectures I'd give where I'd 221 00:11:07.879 --> 00:11:11.039 say hacking is now commercialized. 222 00:11:10.559 --> 00:11:13.159 Shouldn't it be pass per fishing fishing? 223 00:11:13.360 --> 00:11:19.399 Well, I mean, if you're gonna go with it as 224 00:11:19.759 --> 00:11:27.720 PHAs So here from ban Bam so hacking is definitely commercialized. 225 00:11:27.759 --> 00:11:31.399 It's definitely you can pay to get ransomware created, you 226 00:11:31.440 --> 00:11:33.120 can pay to have it deployed, you can pay to 227 00:11:33.120 --> 00:11:35.279 have a troubleshoot, you can pay to have somebody answer 228 00:11:35.360 --> 00:11:38.360 the phone when Grandma calls to send bitcoin, et cetera. 229 00:11:38.559 --> 00:11:41.440 And reading this article, which we linked here in Hacker News, 230 00:11:41.919 --> 00:11:45.759 I find it most disturbing. And I'm just guessing I'm 231 00:11:45.799 --> 00:11:49.200 inferring from what I see on this page that if 232 00:11:49.240 --> 00:11:52.960 you get fished like this, not only will it turn 233 00:11:53.080 --> 00:11:55.679 all of your guitars from right handed to left handed, 234 00:11:56.399 --> 00:11:59.840 but it will put a soft rectangular red eye patch 235 00:12:00.120 --> 00:12:03.360 floating eye patch on your head with a little white 236 00:12:03.440 --> 00:12:03.960 arrow on it. 237 00:12:03.960 --> 00:12:06.320 It's definitely definitely. That's yeah, that's fair. 238 00:12:08.200 --> 00:12:10.799 So I will ask that we change our order of 239 00:12:10.840 --> 00:12:15.360 operations because this specifically calls out Collie three sixty five, 240 00:12:15.840 --> 00:12:18.879 and one of our later stories is about exactly that tool. 241 00:12:19.120 --> 00:12:21.759 Oh let's talk about it then, Okay. 242 00:12:21.440 --> 00:12:23.679 If you skip down the road a little bit, we 243 00:12:23.799 --> 00:12:26.879 have an article also by the Hacker News, Cali three 244 00:12:27.000 --> 00:12:31.480 sixty five weaponizes Microsoft authentication against US company's new enterprise risk. 245 00:12:31.759 --> 00:12:35.000 So this is part of that basket of risks. And 246 00:12:36.799 --> 00:12:40.080 when I see Collie three sixty five, is this a 247 00:12:40.080 --> 00:12:45.120 an actual like tool on Collie the Linux distribution or 248 00:12:45.159 --> 00:12:47.799 is this just a name that crosses over? I think 249 00:12:47.799 --> 00:12:48.399 it's the second. 250 00:12:48.639 --> 00:12:50.000 No, this is just a phishing kit. 251 00:12:50.159 --> 00:12:52.879 Oh okay, So it's not like offsec has gone rogue. 252 00:12:53.519 --> 00:12:56.600 No they're not. Yeah, No, this is just a phishing kit, 253 00:12:56.639 --> 00:12:58.720 and it's one of the frameworks that you know was 254 00:12:58.759 --> 00:13:01.639 probably vibe coded and just showed up, and it is 255 00:13:01.840 --> 00:13:03.440 probably pretty good at what it does. 256 00:13:04.480 --> 00:13:06.399 All right, So is there anything more to the story. 257 00:13:06.440 --> 00:13:07.879 I mean, we might as well talk about it now. 258 00:13:08.240 --> 00:13:12.240 Callie three sixty five weaponizes Microsoft authentication against US companies, 259 00:13:12.639 --> 00:13:13.919 new enterprise risk. 260 00:13:14.039 --> 00:13:16.000 I think it's just more of what we were just 261 00:13:16.039 --> 00:13:18.320 talking about. I think it's folds into the same unless 262 00:13:18.360 --> 00:13:20.240 it's all Unless there's something specific. 263 00:13:20.960 --> 00:13:22.799 No, but I think one of the things I'd like 264 00:13:22.840 --> 00:13:25.879 to warrant against is it used to be to perform 265 00:13:25.960 --> 00:13:28.639 this level of an attack, you needed a modicum of 266 00:13:29.120 --> 00:13:32.919 understanding of technology, right, You needed to know like, okay, 267 00:13:33.039 --> 00:13:36.759 maybe even a smattering maybe right, a smidgen would not 268 00:13:36.799 --> 00:13:40.399 be trying I'm trying to highbrow this a little bit 269 00:13:41.720 --> 00:13:47.600 it's just for you car. So now you don't right now, 270 00:13:48.200 --> 00:13:51.039 you can take a stand on the shoulders of giants, 271 00:13:51.080 --> 00:13:52.840 as they say, and just take the frameworks that are 272 00:13:52.840 --> 00:13:54.799 out there, or if you don't want to do that, 273 00:13:55.200 --> 00:13:57.080 you can just vibe code your own. You just ask, 274 00:13:57.879 --> 00:14:00.879 you know, claude to hey, I'm doing this practice thing 275 00:14:00.960 --> 00:14:03.240 and can you create me a fishing framework. And for 276 00:14:03.279 --> 00:14:06.600 the most part, most of these ais don't generally do 277 00:14:07.480 --> 00:14:10.519 what you ask it to. Guardrails are easy to bypass, 278 00:14:11.080 --> 00:14:12.440 so which we'll talk about later. 279 00:14:13.080 --> 00:14:16.360 All right, let's move on. Chinese hacker commands deep Seek 280 00:14:16.480 --> 00:14:20.240 via Telegram to launch autonomous attacks. 281 00:14:20.600 --> 00:14:21.919 There's three things I don't trust. 282 00:14:22.519 --> 00:14:24.639 Telegrams went out when the phone. 283 00:14:24.519 --> 00:14:31.320 Became Telegram is a very very popular in Ukraine and 284 00:14:31.320 --> 00:14:35.000 in Russia. It's banned in Russia now because it was 285 00:14:35.000 --> 00:14:37.480 created by somebody who is not a friend of Russia's anymore. 286 00:14:37.559 --> 00:14:38.519 Oh no, what is it? 287 00:14:38.840 --> 00:14:42.519 So telegrams? Like, what's it's up? It's like, yeah, it's 288 00:14:42.519 --> 00:14:45.320 like what's app? It's like signals, It's like any one 289 00:14:45.399 --> 00:14:49.080 of those you know, messaging slash voice, slash video slash 290 00:14:49.519 --> 00:14:52.639 whatever chat apps. So fun. Fact, the founder who now 291 00:14:52.679 --> 00:14:57.000 I think lives in either Dubai or Switzerland has dropped 292 00:14:57.000 --> 00:15:01.840 their Russian citizenship has been convicted as a terrorist in Russia. Oh, 293 00:15:02.320 --> 00:15:06.440 because they've allowed Ukraine to use Telegram and haven't taken 294 00:15:06.519 --> 00:15:09.240 down forms and messages that have been used as part 295 00:15:09.279 --> 00:15:09.600 of the war. 296 00:15:09.720 --> 00:15:12.559 All right, So deep Seek is an AI model, right, 297 00:15:13.120 --> 00:15:16.679 Chinese a Chinese AI model. But you can also download 298 00:15:16.679 --> 00:15:17.559 it and run it locally. 299 00:15:18.440 --> 00:15:21.320 So yeah, there are deeps we can run locally. 300 00:15:21.440 --> 00:15:24.279 Yeah, absolutely, so, I mean I have one. I'm running 301 00:15:24.279 --> 00:15:27.159 it here. But am I at risk running this locally? 302 00:15:27.279 --> 00:15:29.120 Or is it deep Seek in the cloud that is 303 00:15:29.679 --> 00:15:31.360 susceptible to this attack? 304 00:15:31.519 --> 00:15:31.559 No? 305 00:15:31.879 --> 00:15:34.279 This is think of this in a different way. This 306 00:15:34.559 --> 00:15:40.440 is a Chinese attacker leveraging Telegram as a command and 307 00:15:40.519 --> 00:15:45.519 control to control a deep Seek instance to go out 308 00:15:45.519 --> 00:15:48.120 and do attacking. Okay, so it will go out, and 309 00:15:48.840 --> 00:15:52.240 it will go out and do all of the reconnaissance 310 00:15:52.279 --> 00:15:54.559 on hundreds of companies and it can it'll come back 311 00:15:54.600 --> 00:15:57.120 and say, hey, I found these companies that we can exploit. 312 00:15:57.519 --> 00:16:00.120 And we've seen this type of a report probably and 313 00:16:00.440 --> 00:16:03.799 I think it was September last year where China was 314 00:16:03.919 --> 00:16:08.600 using an anthropic model to break into organizations, and they 315 00:16:08.679 --> 00:16:12.080 had created this harness where they could say, hey, here's 316 00:16:12.120 --> 00:16:13.919 the you know, go out and grab the top one 317 00:16:13.960 --> 00:16:17.039 hundred US companies and and and start profiling them and 318 00:16:17.080 --> 00:16:20.799 see what their exploits are. So Palo Alto's Unit forty 319 00:16:20.840 --> 00:16:24.159 two said, hey, we're actually detecting this this attacker who's 320 00:16:24.679 --> 00:16:30.960 leveraging telegram to communicate with a deep seek. And then 321 00:16:31.039 --> 00:16:36.120 that deep Seek is actually using Hermes to do things. 322 00:16:36.840 --> 00:16:38.799 For those of you who haven't played around with AI 323 00:16:38.919 --> 00:16:44.200 local agents, Hermes. Hermes is actually by a French organization 324 00:16:44.240 --> 00:16:47.120 if I remember correctly, but Hermes. 325 00:16:46.720 --> 00:16:48.279 Is pronounced in French. 326 00:16:52.559 --> 00:16:53.679 I'm not even gonna try. 327 00:16:54.039 --> 00:16:57.679 They'll revote. 328 00:16:57.759 --> 00:17:03.360 I can't do it. I can't do it. Ohmy Hermes, 329 00:17:03.480 --> 00:17:08.279 it's yeah. Anyways, So, needless to say, you all heard 330 00:17:08.319 --> 00:17:10.799 of open Claw, right, which was one of the first 331 00:17:10.799 --> 00:17:14.119 agents that really kind of got a lot of hype around. 332 00:17:14.319 --> 00:17:19.680 It would leverage llms and it could do stuff right hands, Yeah, exactly, 333 00:17:19.799 --> 00:17:20.119 that's a. 334 00:17:20.039 --> 00:17:21.359 Great question harness. 335 00:17:21.400 --> 00:17:23.240 Yeah, so, and Hermes is one of them. 336 00:17:23.319 --> 00:17:25.279 Let me take that one real quick. So, so you 337 00:17:25.359 --> 00:17:27.160 have the model, which is the l M at the 338 00:17:27.160 --> 00:17:28.640 heart of this thing, and there's been a lot of 339 00:17:28.680 --> 00:17:32.279 hype about oh, GPT, this GPT, that claude this, claude 340 00:17:32.279 --> 00:17:36.640 that the model is the prediction engine that says, based 341 00:17:36.680 --> 00:17:38.920 on this, this is the next word, right, and then 342 00:17:38.960 --> 00:17:40.440 that's the next word, and that's next word. 343 00:17:40.519 --> 00:17:41.960 Yeah, and it's stateless and done. 344 00:17:42.240 --> 00:17:46.160 We've been using harnesses since we moved off of chat 345 00:17:46.200 --> 00:17:50.519 and into agentic AI, and the harness is the thing 346 00:17:50.519 --> 00:17:53.039 that controls the loops and the pointing and the tool 347 00:17:53.119 --> 00:17:55.319 use in all the memory and all that, all the things, 348 00:17:55.480 --> 00:17:58.799 all the things like memory that make it useful outside 349 00:17:58.799 --> 00:18:04.480 of answering questions about Wagner and composing odes to grilled 350 00:18:04.519 --> 00:18:05.200 cheese sandwiches. 351 00:18:05.359 --> 00:18:08.079 Don't have any idea what you're talking about. He's just 352 00:18:08.160 --> 00:18:12.240 making stuff, Wagner, you. 353 00:18:12.200 --> 00:18:17.680 Know what sandwiches? That okay? Anyway, So the thing is, 354 00:18:17.839 --> 00:18:21.720 in the last round we've gotten diminishing returns on the model, 355 00:18:22.279 --> 00:18:26.200 and so that the hyperscalers and the open weight companies 356 00:18:26.200 --> 00:18:29.240 have been competing on the harness and it's indistinguishable to 357 00:18:29.319 --> 00:18:31.480 most people because they don't see the difference, but the 358 00:18:31.599 --> 00:18:32.559 model can get. 359 00:18:32.559 --> 00:18:35.599 That's where the innovation is going on, is exactly. 360 00:18:35.480 --> 00:18:38.559 And they're finding that just change the same model with 361 00:18:38.640 --> 00:18:41.039 a different harness can make a twenty different point swing 362 00:18:41.119 --> 00:18:44.519 on some of these these standards that they're measuring against. 363 00:18:44.559 --> 00:18:47.680 And of course all these ais are testing against their 364 00:18:47.720 --> 00:18:50.200 training for the test. They're trying to make them better 365 00:18:50.240 --> 00:18:53.400 at the test. But if you find that the AI 366 00:18:53.519 --> 00:18:58.200 that you're using agentically uses tools well, answers questions correctly, 367 00:18:58.680 --> 00:19:04.000 doesn't hallucinate so much, you know, can can correct itself, 368 00:19:04.519 --> 00:19:06.640 then it's the harness that's really making the big difference, 369 00:19:06.640 --> 00:19:08.920 and that's where the competition is. Now. One other thing 370 00:19:08.960 --> 00:19:11.799 I want to say here is Deep Seek and all 371 00:19:11.839 --> 00:19:13.920 these open weight models and open weight models. When you 372 00:19:13.960 --> 00:19:15.920 can download and run on your own stuff. You can't 373 00:19:16.319 --> 00:19:18.720 download mythos and run it on your own stuff. You've 374 00:19:18.720 --> 00:19:20.000 got to run it against the horse. 375 00:19:20.759 --> 00:19:22.839 You have enough money, Okay, let's not. 376 00:19:25.519 --> 00:19:25.640 So. 377 00:19:27.039 --> 00:19:31.000 The latest version of Kimmy Kimmy three by Moonshot, another 378 00:19:31.079 --> 00:19:35.319 Chinese company, it would take a million dollars in GPUs 379 00:19:35.519 --> 00:19:38.960 to run it, So the average person's not downloading that. 380 00:19:39.000 --> 00:19:42.880 They context is kim a model or a harness? 381 00:19:43.279 --> 00:19:46.039 Kimmy is is exactly like Deep Seek. It's another it's 382 00:19:46.039 --> 00:19:49.920 a model plus harness. It's another AI company, but they're 383 00:19:49.960 --> 00:19:51.759 the latest one that came out. They're the they're the 384 00:19:51.759 --> 00:19:55.680 most powerful Chinese model that released, is releasing his open Weight. 385 00:19:56.319 --> 00:19:58.720 But understand, what I'm trying to help you understand is 386 00:19:59.400 --> 00:20:04.759 this hacker either has enormous resources, or is hacking into 387 00:20:04.839 --> 00:20:07.359 and stealing from someone who has enormous resources to run 388 00:20:07.400 --> 00:20:11.759 deep Seek, or they're using the actual official deep Seek 389 00:20:11.920 --> 00:20:15.039 right to hack, or finally, they have millions of dollars 390 00:20:15.039 --> 00:20:17.920 in they're hosting their own modern verse. It's another thing 391 00:20:18.000 --> 00:20:19.839 is they could be using a much older version of 392 00:20:19.839 --> 00:20:22.400 Deep Sea Sure, and it's not as enabled. It's not 393 00:20:22.440 --> 00:20:24.559 as capable as it seems. It's just they happen to 394 00:20:24.599 --> 00:20:25.039 be using it. 395 00:20:25.119 --> 00:20:25.440 Okay. 396 00:20:26.039 --> 00:20:30.039 Interestingly enough, we have found some of the less capable 397 00:20:30.119 --> 00:20:34.920 models are more willing to do bad things. I'll leave 398 00:20:34.960 --> 00:20:39.359 it at that, because they're down and out on everybody's 399 00:20:39.400 --> 00:20:44.680 talking about Methos and you're like, come here, Haiku, put 400 00:20:44.720 --> 00:20:47.440 your armor on. Haiku. You're like, let's do some fun things. 401 00:20:47.559 --> 00:20:48.880