DSS 2026 — Sword vs. Shield: Navigating the AI Security Chaos in 2027 Davit Kobakhidze Source video: https://www.youtube.com/watch?v=Cyn9zfVyVtU Automatic captions — not human-reviewed. Complete spoken-audio coverage is unverified. Source language tag: en (English). Source export SHA256: 48f1109120a4055b7d9233b3a75414d6d5ec3e609d188ccce4de2314cf4f3013 [0:00] So okay. [0:02] Hey, my name is David. [0:04] I'm a BD guy from Consensus Diligence. [0:06] So not a security auditor. [0:08] Sorry about that, but I do think that I [0:11] can share some interesting perspective [0:14] because I talk a lot with our clients [0:16] and [0:17] with our security researchers, which can [0:19] help me to provide some interesting [0:21] information for you guys. [0:22] So the the topic of my presentation is [0:25] Sword versus Shield. It's an ancient [0:27] concept about the competition between [0:29] offense and defense team. I believe [0:31] originated in [0:33] China [0:34] thousand years ago. So I do think that [0:37] what we have today [0:38] on the market it [0:40] this competition, who will outsmart who? [0:44] The topic says navigating the AI [0:46] security chaos in 2027, but also it [0:49] backs down to how do you keep your [0:51] protocol and funds secure in 2027? [0:55] And honestly [0:57] after last couple of weeks with all [0:59] these bugs exploited, with all the [1:01] protocols got hacked, [1:02] with all the incident we have [1:05] I think you already have an answer. Do [1:08] you? [1:09] Anyone? [1:11] You simply can't. [1:14] That's it. That's it for my talk. Thank [1:16] you so much. [1:18] Okay. [1:20] The timer says I still have 14 minutes [1:22] to go. [1:23] So I perhaps should stop with bombing [1:25] with dead jokes. [1:27] And let's try to at least figure out how [1:29] can we minimize exposure [1:31] moving forward. [1:32] Before that, let's just have a quick [1:34] recap [1:35] on what I've heard from the market [1:38] throughout this last couple of years. [1:41] So first [1:42] actually in 2024 I already start hearing [1:45] from protocols and security researchers [1:47] themselves that oh [1:49] AI promises like a lot of efficiency. I [1:52] do think it will replace security [1:54] researchers moving forward. Back then it [1:56] was like [1:57] vague timeline three to five years. [2:00] But as early as 2026, actually during [2:03] the Denver conference, I said here in a [2:05] lot from different protocols that you [2:08] know what? [2:09] We think AI made security simple. We'll [2:12] probably cut off our spending and do a [2:15] lot of stuff in house. [2:16] And throughout 2026 we see this idea [2:19] manifested in actions. [2:21] So you see like job market for security [2:23] researchers in US dropped by 26%. [2:26] Lots of security companies, both Web two [2:29] and Web three start cutting off their [2:31] security teams, auditing teams. [2:34] And bug bounties have been paused or [2:36] canceled or closed. [2:38] But [2:39] actually I do think that in reality [2:42] right now in November 2026 [2:45] and moving forward to 2027 [2:48] AI actually created way more work for [2:50] security researchers. [2:52] It might be a bit different work. It [2:54] probably evolved. [2:55] But still there is a long [2:59] way to go. [3:00] So why did that happen? In my opinion [3:03] there was a wrong perception of AI [3:05] capabilities and the sword, meaning [3:07] hackers and offense got sharper first. [3:10] And the shield, meaning the capabilities [3:12] like a silver bullet, magic pill, super [3:15] app who closes all your security [3:17] questions [3:19] is still being built or potentially will [3:21] not [3:22] be built ever. [3:24] So I do think that defenders should pick [3:26] up the sword too. [3:28] Because again, remember then when you [3:30] are trying to [3:32] hack the website or the app [3:34] you can test one bag across multiple [3:36] targets, but when you work as a defender [3:39] in a defense team [3:41] you have to test for every box in a [3:42] single target. That's a bit different [3:45] work. [3:46] And I don't think that AI made offense [3:50] or hackers smarter [3:52] or way more powerful. But what I do [3:53] think AI did, it made them faster and [3:57] exploits way cheaper. So here are the [3:59] like four examples. [4:01] I won't go like deep into them, but just [4:04] to give you a glimpse that, for example, [4:06] for ConnectWise, [4:10] hackers exploited by the way the [4:12] security framework which was published [4:13] by security researcher, [4:15] and they managed to as quick as 2,900 [4:19] logins in a minute to try to steal [4:22] ConnectWise clients data. [4:25] Same similar stuff happened in South [4:26] Korea, like in a matter of days, 16 [4:28] financial institutions, banks and [4:31] crypto companies were exploited. [4:33] Speaking about the [4:35] ease and the like from resource [4:37] perspective, the cheapness of hack. [4:39] So [4:40] Nord Forge VR routers created a great [4:44] research and [4:46] traced actually one hacker's job who [4:48] left a lot of like footprints during his [4:51] hacks and turned out that he hacked [4:55] attempted to hack 1,200 websites and his [4:57] budget was only about $5,000, which [5:00] means that it was less than five bucks [5:03] per attempt [5:04] and [5:05] $50 per successful hack actually, [5:09] which is [5:10] quite impressive. I do believe that [5:12] maybe moving forward it will be even [5:13] cheaper and we should consider that as [5:15] well. [5:16] And a bit more speculative stuff, but [5:18] Palo Alto Unit 42 says that in PHP [5:22] breakout like hackers used AI during [5:27] their further trying to stealing data [5:29] and exploiting [5:32] targets. [5:34] Interestingly enough, then I asked our [5:36] security fellows [5:38] what do they think about the AI usage. [5:43] Don't read it as a study. It's not a [5:45] formal study. It was just like a [5:47] set of conversations [5:48] with my buddies and with the network as [5:50] well. But surprisingly so, [5:54] I had like three almost equal groups of [5:57] people [5:58] saying that [6:00] they have entirely different perspective [6:02] or similar but different perspective on [6:04] AI capabilities. Like the first group of [6:06] people said that actually [6:08] we don' think that AI can find anything [6:11] that I wouldn't have been able to [6:12] identify. Maybe zero to five percent [6:15] additional [6:16] highs and [6:17] criticals. And I'm talking about highs [6:19] and criticals, no about low and [6:20] informational obviously. [6:22] The second group was a bit more moderate [6:24] and they said that [6:25] 10 to 20 percent of highs and criticals [6:28] way faster identifiable with AI than [6:31] without, but still you have to pay [6:33] attention [6:35] at least one out of three [6:37] highs and creates are missed by AI and [6:38] you have to go through the code base, [6:40] especially if it's like a [6:42] major update to the protocol or [6:45] major feature. [6:47] The third group was even more optimistic [6:49] and they said that they've seen a huge [6:51] development of AI capabilities [6:52] throughout the year. If in the beginning [6:54] it was 10 15 percent additional [6:57] value for them, now they see it like 25 [6:59] 35 percent on creates and majors across [7:01] the board. It doesn't mean that they [7:03] wouldn't be unable to find these bugs, [7:04] but it means that they would spend a lot [7:07] more time [7:08] on it. By the way, I'm really curious to [7:11] talk to you after the talk and [7:13] see which group is [7:16] resembles your view on the AI, so feel [7:18] free to approach. [7:20] On the first positive side of things, we [7:22] have to admit that [7:23] AI produces not only the the useful [7:27] stuff, but also slope. [7:29] And the the sort of on that is quite are [7:31] quite interesting. So on first positive [7:33] side of things, AI researchers told me [7:35] that it's not necessarily bad stuff. If [7:38] we get from 30 to 50 percent false [7:40] positive rate and we know that we have [7:42] to go through this findings manually, [7:44] this is a part of the engagement, that's [7:46] fine because we get a lot of [7:48] interesting leads, interesting security [7:50] cues and that will help us investigate [7:53] further and understand protocol better. [7:55] And maybe this particular issue is not [7:58] real, but it will lead me to something [8:00] interesting. [8:03] On the client side of things regarding [8:05] the false positive, Uh the answer was a [8:07] bit different. So nice figure from [8:10] Coinbase actually. So they reported only [8:13] 4% of their submissions in the first [8:15] half of 2026 were real, which is insane. [8:18] So 96% of submissions is a complete [8:20] fake, [8:21] which is a pity, which is a pity. Uh but [8:24] we have to understand that there are [8:25] different types of noise for clients as [8:27] well. So clients told me that if [8:30] this report with lots of false positive [8:32] stuff goes to the auditing team or to [8:35] the security team and we know that [8:37] they'll go through it and validate it [8:39] and triage it, then it's fine. However, [8:42] if [8:42] false positive report report with like [8:45] lots of false positives goes directed to [8:47] dev team to patches and fixes, that's a [8:49] real problem. And also uh some clients [8:52] told me that they've seen like huge [8:54] development in effectiveness of AI and [8:57] they received around 50% false positive [8:59] rate in the first half of 2026. Now it's [9:02] way less, but still much more uh [9:05] to analyze. [9:07] We've talked about briefly about code [9:09] bugs, but actually [9:11] uh the majority of money we are losing [9:13] right now is not through the code bugs. [9:15] Although in quantity they're still [9:17] number one. [9:18] The majority of uh money lost during [9:22] this year were through operational [9:23] security, suppliers, partners, [9:26] bugs through and not under your like [9:29] code control. Let's put it that way. [9:32] Couple of examples again. [9:34] Like Bitget got exploded because there [9:36] was a hole [9:38] in the security product they purchased [9:40] and they just drained almost $400 [9:42] million. [9:43] Uh another um [9:45] one more time for ConnectWise as well, [9:47] like a reverse example. They were [9:49] holding their clients' data and they got [9:51] exposed and [9:53] 2,250 clients got their data leaked. [9:57] So it just [9:59] matters that [10:01] the threat is not only from your code [10:03] base. There are multiple venues to got [10:05] exposed. [10:07] And that brings me to another thought [10:08] that you have to treat your security not [10:11] as a uh one layer. You have to treat it [10:14] as a multiple like layered approach when [10:17] in the center you have [10:18] your protocol or your uh [10:21] client's funds [10:22] and you have to build as much layer as [10:24] possible around it. [10:26] Obviously, it starts off with a design. [10:28] Bad architecture, bad design, you'll got [10:30] exposed. Then obviously your code should [10:33] be uh uh in a good shape. Audits, AI [10:36] review, you name it, fuzzing, sometimes [10:39] formal verification, everything is super [10:41] important here. [10:42] But then you have to understand that the [10:45] current state of protocols, current [10:47] state of market is interconnectivity [10:49] between like multiple partners and [10:51] vendors. You have to treat all the [10:53] integration with your partners almost as [10:55] rigorously as your own codebase. [10:57] Obviously then goes separation, security [10:59] keys and access. Don't forget about [11:01] that. On top of that is monitoring. And [11:04] uh you have to understand that just [11:06] having some monitoring system uh [11:09] embedded is not enough. Have you tested? [11:13] Did the right person receive uh the [11:14] notification when it should? Uh did you [11:18] cover the whole potential incidents you [11:20] have with your monitoring system? That [11:22] the question you have to ask. But on top [11:24] of that [11:25] you have to have a plan. [11:27] Unfortunately as it sound [11:29] stuff will break [11:31] and in first hours, even days, uh your [11:33] brain will not work as efficiently if [11:36] something goes sideways. So you need to [11:38] have a plan which outlines first steps [11:41] you have to do in first minutes, first [11:42] hours, first days. [11:44] How to stop the drain? How to stop the [11:46] leakage? How to notify your partners? [11:48] How to make a communication to the [11:50] community? Whether you should uh or [11:52] should not engage with uh law [11:54] enforcement in one capacity. All these [11:56] questions should be decided beforehand [11:58] before hack happened and every [12:00] responsible person should have [12:02] this plan [12:03] I would say printed out on his desk. [12:07] Uh another uh important stuff we as a [12:10] security researchers we have to pick up [12:12] the [12:13] I know it might sound like just use my [12:15] AI. [12:16] Actually, it's not. [12:18] Yes and not. [12:19] So, first of all, you have to understand [12:21] how attackers are viewing your codebase [12:24] and how do they use your [12:26] AI to hack you. [12:28] And then you have to build upon that. [12:31] First of all, you have to build a [12:31] process inside the company. How do you [12:34] collect the context for your models? How [12:37] what standards do you have for teams to [12:39] use AI tools? [12:41] What happens if you identify issue, [12:42] where it goes? [12:44] Simple stuff like that, but believe me, [12:46] when you have it processed out, then you [12:47] can go to the next stage. [12:49] What [12:50] what is triage for you? How do you do [12:52] your triage [12:54] in your company? What tools do you use? [12:56] Speaking about tools, [12:58] how many tools do you use? Do you have [13:00] only one in-house tool or you have [13:02] multiple tools like third-party tools, [13:04] your in-house tools? [13:06] How often do you rotate the models you [13:09] use? Because we heard multiple times [13:11] with our customers that even though one [13:14] model could be prominent and one [13:15] framework could work well, [13:17] the combination always beats it. So, if [13:19] you combine this several models, several [13:22] tools, several approaches, it will [13:24] always [13:26] be more efficient than only one [13:28] brilliant model you might have. [13:30] And remember, [13:32] AI alone does not replace an audit or a [13:34] security researcher, but [13:37] done right, it can show you the way the [13:39] attackers will view your codebase [13:41] because they will not audit you, they [13:43] will try to break you. [13:45] And in [13:47] the last thought I would like to share [13:49] is like security should not be stamp, it [13:51] should be a habit, it should be a [13:53] routine [13:54] you perform every day within your [13:57] company. [13:58] And obviously, [14:00] you don't always have to engage with [14:02] third parties and pay a lot of money for [14:04] audits and [14:06] third-party tools. There are stuff you [14:08] can do yourself within your company. [14:10] Everyone has access to AI. You can come [14:12] up with your own tooling within the AI. [14:14] Operational security basics [14:16] as simple as sometimes killing your old [14:18] accounts as well. [14:20] Uh, monitoring, setting up your [14:21] monitoring. No one will set up processes [14:23] and monitoring for you. Even if you have [14:25] like a brilliant report around [14:27] operational security from third party, [14:29] if you do not engage within your [14:31] company, if your leadership is not [14:33] interested in implementing this stuff, [14:35] nothing will work. [14:37] And obviously there is a period of time [14:39] when you have to bring the heavy guns, [14:41] when you have to engage [14:42] with third party companies. I would say [14:45] every feature or every change which [14:47] involves money moving uh, operations, [14:50] new core logic, new features, new bigger [14:52] integrations, [14:54] or if you want to build like a [14:55] operational security playbook for the [14:56] first time, [14:58] this is the time you have to engage uh, [15:00] with third parties because that would be [15:02] way more efficient. And [15:05] remember like not every change need an [15:07] audit, a full audit, but every change [15:09] needs someone to look on it. [15:13] Uh, [15:14] yeah. So getting back to the initial [15:16] question, uh, how do you navigate AI [15:18] security uh, chaos in 2027? [15:22] So the answer is simple. You can't if [15:24] you wait uh, for the shield to be built. [15:27] But if you lower your defense and pick [15:28] up the sword, you have much better [15:30] chances uh, to survive and move forward. [15:33] That's it for my time. Thank you so [15:35] much. [15:36] And please [15:37] please come to me, talk to me in case [15:39] you think I got something wrong. I would [15:41] curious to [15:43] hear your perspective as well. Thank you [15:44] so much.