DSS 2026 — Evergreen: continuous AI-augmented security by Lido x Nethermind Eugene Mamin Source video: https://www.youtube.com/watch?v=_yvcOTN6XWQ Automatic captions — not human-reviewed. Complete spoken-audio coverage is unverified. Source language tag: en (English). Source export SHA256: d042187b25c471b464f1217bcc01f1ed7cc3f20b36212f05cc77b8310e05ffc3 [0:02] It might be me. Okay. [0:13] Hi everyone. [0:17] Okay, here's my presentation. So, [0:21] I'm going to talk about Evergreen [0:23] concept [0:24] we built together with [0:26] Nethermind as [0:29] the thing that [0:30] probably every protocol should have in [0:32] the industry at the moment. [0:36] I'm Eugene. I'm look for engineering at [0:39] Lida Labs. [0:40] You can follow me on Twitter. [0:43] Nice to meet you all. [0:49] Sorry, I [0:52] yeah, technology. [0:54] So, agenda is simple. I will start with [0:56] numbers, then what we [0:59] did with about that numbers and what you [1:03] can start doing on Monday as a protocol. [1:08] I'll start with audits [1:10] and continue the line Yuri started. [1:15] Lida did 30 audits this year [1:18] with eight independent security firms. [1:22] We have zero critical, zero highs audits [1:25] work, but that's still not enough [1:28] because [1:29] like Yuri said, when you finish audit, [1:33] it's just the beginning, not the end. [1:36] Next one is Lida bug bounty. [sighs] [1:40] So far [1:42] from the very beginning of the program, [1:45] we paid [1:46] $410,000 [1:50] for white hats submitted. [1:52] However, you can follow [1:54] this particular arrow that initially [1:58] white hats claimed to have like 180 [2:03] critical security findings. [2:06] After triage on our side, [2:09] it ended up with three critical payments [2:12] actually made. So, most of the work [2:15] here, unfortunately, is saying no, but [2:18] still carefully. [2:23] What changed recently? So, that's the [2:26] whole graph of intake we have in bug [2:29] bounty. [2:30] You can see that starting from 2025, [2:33] mostly late 2025, [2:35] intake has [2:37] largely increased and you can't just [2:40] basically automatically beat up every [2:43] submission. [2:44] You have to have [2:47] a senior engineer who [2:49] who should say yes or no and why. And [2:52] ideally, you should have a second [2:54] opinion about each [2:56] particular about each material finding. [3:03] What else? So, [3:06] it's not only the code that should be [3:09] audited and reviewed and put on scope of [3:12] bug bounty. It's also external world [3:15] change while your code stays the same [3:18] and it might be even the state of code [3:21] remain the same. [3:23] What I mean? It might be dependency [3:26] change. So, suppose you rely on bridge [3:28] aggregation and one of the bridges [3:30] decided to change the [3:33] approach for the for the relay [3:34] mechanism. [3:36] It might be your off-chain stack [3:38] dependency. Of course, you can think [3:39] about Oracles first, but that's not the [3:42] only off-chain tooling you might have in [3:44] your protocol. [3:46] I mean it might might be auxiliary [3:47] service which could also have its own [3:51] weaknesses [3:53] like race conditions, bugs, and things [3:55] pretty hard to debug. [3:57] And last but not least, you still rely [4:00] on something when you build a protocol. [4:02] For Lida staking, it's Ethereum itself. [4:04] It's consensus spec. [4:06] So when Ethereum undergoes hard forks, [4:10] we must be sure that protocol [4:14] would work correctly, would work [4:16] according to the spec, and we won't be [4:18] affected both from consensus layer side [4:20] and from execution layer side. [4:23] So that we start testing every hard fork [4:25] early with devnets, with our very own [4:28] devnets, then public devnets and public [4:30] testnets and so on and so forth. [4:32] Lately we found pretty interesting [4:34] vulnerability in Glamsterdam hard fork [4:37] and consensus layer spec. [4:39] The thing is that if you had a deposited [4:44] like not yet deposited but pending [4:46] validator deposit, [4:48] it could have been attributed to a [4:50] builder with the same pub key resulting [4:53] in [4:54] fund loss for this deposit. [4:57] Luckily this was fixed. Huge shout-out [5:00] to Core devs. Huge shout-out to folks [5:03] who supported our effort to make it [5:05] fixed. And on mainnet this year we [5:08] expect Glamsterdam it will be fixed. [5:13] So that's why we started to build [5:15] Evergreen. [5:17] It's the idea of audit that never ends [5:21] and that is [5:24] always aware of [5:26] everything happening around. [5:29] We started with two tracks. One is [5:32] external track. We had a pilot with [5:34] Nethermind. And another track is [5:37] internal. We built very like our very [5:40] own tool Lida Lens. [5:45] Starting from pilot, [5:48] Nethermind team helped us to build [5:52] a concept for Lida Core codebase. It's [5:55] like 19 [5:58] K of code. [6:00] It was four weeks with two security [6:03] researchers, two AI engineers. [6:06] Uh and the most of work was not about [6:08] like picking the the best model. Of [6:11] course, the most work was about [6:13] properly [6:15] uh gathering curated documentation and [6:18] context. [6:21] Uh so [6:22] uh as a part of this experiment uh it [6:25] was measured then when you have a proper [6:27] context [6:29] you have a real leverage over the model [6:32] quality. [6:33] So both uh detection rate increases and [6:36] false positive rate decreases [6:39] uh with you gathering context about your [6:41] protocol, your dependencies and so on. [6:45] Uh [6:47] three lessons we learned from this [6:49] pilot. [6:50] The first one as I said context is the [6:52] lever. So [6:54] pay most attention about like uh what [6:57] your design intent. [6:59] Write it down because [7:02] design intent is something that [7:06] uh not agent, not security auditor, not [7:09] a white hat can understand pretty well [7:12] from [7:13] out of nowhere [7:15] unless you written it out. [7:17] Uh and of course don't rely on [7:20] on any particular engine. Use bunch of [7:23] tools. They might be overlapping but if [7:26] you have [7:28] the whole family of tools um [7:31] working together you get the most of the [7:33] coverage. [7:36] Uh you can learn about this uh [7:39] particular pilot uh on Nazar Mine blog. [7:42] Numbers, [7:44] particular results, everything shared. [7:46] So uh [7:49] it [7:50] it's available for like async [7:53] investigation. [7:54] Uh for our internal track we built Lidar [7:57] Lens. Uh [7:59] it's the [8:01] over like it's the engine that works [8:04] over the clock. [8:06] It has 39 [8:08] uh repositories or projects at scope. [8:12] And it has like just four [8:15] separate AI models underneath [8:19] that allows us to [8:21] sleep better than we slept before. [8:27] Okay. So [8:30] the follow-up about Lidar Lens will be [8:33] provided by Sergey. [8:35] Please stay tuned. It will be in an hour [8:38] or so. [8:40] What you should start doing on Monday? [8:43] Like prepare your context pack. [8:46] If you don't have a team to spin up this [8:49] particular like never-ending audit, then [8:52] try to gather external team [8:55] and [8:57] of course agents propose, people [9:00] approve. [9:01] Approach works here. [9:04] Audits is a snapshot. Security must be [9:07] evergreen. [9:09] That's pretty it. Thank you.