DSS 2026 — The Audit That Never Ends: Evergreen Security in the Age of LLMs Sergey Khomutinin Source video: https://www.youtube.com/watch?v=3zKkdOhDnN0 Automatic captions — not human-reviewed. Complete spoken-audio coverage is unverified. Source language tag: en (English). Source export SHA256: 9301c53d38ebdd6d44c11433032d69f8afc9731560e32556b083539db1bbd240 [0:02] Okay. [0:04] I guess we can start finally. [0:06] Um yeah, I'm Sergey. I'm Solidity [0:08] developer at Trail of Bits and I'll show [0:10] you how we approach the idea of a [0:12] never-ending audit. [0:15] We all need audits, but [0:18] I think we are doing them wrong. We [0:19] still treat audits as a one-time check [0:22] near the end of a development. [0:24] Um [0:25] we write a code, uh freeze it, send it [0:28] to auditors, get a report, [0:30] uh fix the bugs, and then just repeat. [0:33] Uh [0:35] but what about the off-chain [0:36] applications, UI, integrations, and a [0:39] lot of that code [0:41] commonly never gets audited at all. [0:44] Um [0:45] what if audit became a part of a [0:47] development process uh starting from [0:50] development up to the release and even [0:52] further, we can keep looking for [0:53] problems even in the code that is [0:55] already live. [0:57] And uh [0:59] Frontier RLMs allow us to do so. [1:03] Um [1:03] they are pretty much capable of doing [1:05] security work today and finally with [1:08] Frontier RLMs you don't need a complex [1:10] system to start doing this work. [1:14] Uh they can easily check invariants, [1:17] check assumptions, and uh [1:19] check how components interact with each [1:22] other. [1:24] So that's how we built Phalanx, our [1:27] internal defensive system [1:29] that uses just uh [1:32] a common [1:33] coding agent such as Cloud Code and [1:34] Codex under the hood. [1:36] And uh [1:38] we just give it the right context, uh [1:40] ask it the right questions, [1:42] and it tracks our projects, it runs [1:45] checks, [1:46] and it gives us one place to review all [1:49] of the findings. [1:52] Um [1:56] we don't just run uh audits once. We [2:00] um [2:01] we on all of the changes. [2:03] Changes in the code, changes in the [2:05] documents, maybe information about the [2:07] upcoming hard forks, and even just new [2:10] LLM releases, all of that give us [2:11] reasons to check our code again [2:14] and try finding something something new. [2:17] And [2:18] between these checks, [2:20] agents just keep researching your [2:22] questions to go deeper and try to find [2:24] find something deeper. [2:29] We don't just collect findings. They [2:31] have their own life cycle. Something's [2:34] fixed, something is still open, [2:36] something can be reopened to check again [2:39] by humans. [2:41] And [2:45] let's [2:47] continue, please. [2:50] Yeah, [2:51] as well as [2:54] we check all all of the snapshots of our [2:57] code, development branches, upcoming [2:59] releases, as well as production, so we [3:02] can find problems earlier, at earlier [3:05] stages. [3:06] And [3:07] we don't need just to wait for the final [3:09] audit to [3:11] discover security problems. [3:14] So what's Lens actually code for us? [3:18] More than we can cover in two minutes, [3:20] of course, but I'll just show you two [3:22] public examples. First one is [3:25] when we [3:26] checked our [3:28] code for a Gnosis compatibility, Lens [3:30] just found a bug in the consensus specs. [3:33] Basically, validator deposits could go [3:35] to our builder with the same public key. [3:37] And here's the proposed fix, and it's [3:39] already been merged a couple of days [3:41] ago, which is great. [3:43] And this one is [3:46] interesting. It shows us why we still [3:49] need [3:52] checking the code even if it's not [3:54] changed. Basically, [3:56] Wormhole stopped its delivery service [3:58] that our [4:00] cross-chain governance relied on, so we [4:02] just [4:02] had to update the setup and clients call [4:04] this problem. [4:09] So um [4:11] that's what we built in Lighter, but I [4:13] think you can build something similar [4:16] around what matters to you [4:18] in house. [4:20] And here's why. Um first is that you [4:23] know the context around your code [4:25] better. Uh you know all of your internal [4:28] documents, why some design decisions [4:30] were made this way were just previous [4:32] findings. [4:34] Uh you know the the scope and you know [4:37] which contracts and which applications [4:38] need to be checked together in one [4:40] context. [4:41] And you decide when to run the checks. [4:47] So with with that context you should be [4:49] the first to check your own code [4:52] to [4:53] um like researchers and attackers have [4:57] the same LLMs and the same capabilities. [4:59] Uh you just need to [5:01] be sure that your goal is just to find [5:04] them earlier, find security problems [5:06] earlier before the bounty report or [5:08] before the before an exploit. [5:10] And yeah, LLMs keep getting better and [5:13] your security baseline should keep [5:15] getting better too. [5:17] Uh do not hesitate to build your own [5:18] version of [5:20] an audit that never ends. [5:22] Thank you. [5:24] >> [applause]