DSS 2026 — The 0 to 1 guide to AI powered Auditing Methodology Viraz Malhotra Source video: https://www.youtube.com/watch?v=2XkdKecZ-js Automatic captions — not human-reviewed. Complete spoken-audio coverage is unverified. Source language tag: en (English). Source export SHA256: f49fe9d70ba6cb1d432e6f7df1a391e8fa5b00096b672aeb357876d1b544d836 [0:00] Hello. [0:01] Yeah. [0:03] Hey, hi guys. My name is Viraj. I'm a [0:06] blockchain engineer and a security [0:07] researcher and today I'm going to be [0:09] speaking about a zero to one drive to AI [0:12] powered auditing methodology. And as [0:15] we've seen like this year, AI has been a [0:18] big revolution in terms of acting as a [0:20] catalyst for security research work. And [0:23] so I'm going to be just talking about [0:25] the basic pitfalls that we need to avoid [0:27] when you're integrating AI into your [0:28] workflow. A high level overview of the [0:31] harness that I've been using to help [0:34] me like find issues. And then just a [0:37] overview on like different methodologies [0:39] from some OG researchers that I've seen [0:42] and [0:43] and yeah, the different tools that you [0:45] can try and integrate as well. So [0:51] Yeah, so basically we start off with the [0:54] pitfalls. So I think [0:55] AI is only good as [0:58] can only be good as good as your own [1:00] manual experience and knowledge. So I [1:02] think the first bit is you can't really [1:06] basically just tell the AI to give the [1:09] give the AI a code base and then tell [1:11] the LLM to just like find issues. So [1:13] that is the most low utility thing thing [1:16] that you can do. I think it's more about [1:18] acting [1:20] about asking different questions to it [1:22] and in terms of like a particular state [1:24] transition or a particular [1:27] you know, [1:28] particular invariant and things like [1:30] that and not just like a high level [1:31] question that you know, just [1:33] dump the code base and ask it to find [1:36] issues. So that doesn't really work [1:37] because a bug is what is a bug? A bug is [1:40] a actual deviation from the intended [1:43] behavior that the developer has [1:45] in the protocol. And that's how you [1:49] should you know [1:51] approach the whole [1:53] you know, process of hunting issues with [1:57] you know, using LLMs, right? And then [1:59] the other bit which is also important [2:01] that a lot of times that I've seen this [2:04] and we've seen it in the past as well [2:06] that AI can also invent issues as well [2:09] on its own. So at all times you need to [2:11] have a very solid POC verification [2:15] process as well in place where you don't [2:17] want the AI to basically just fabricate [2:19] issues and things like that and also not [2:21] allow the AI to for example your harness [2:24] or whatever to basically reject valid [2:27] issues as well. So you need a lot of [2:28] like manual overview on those things as [2:30] well. [2:32] Yeah. So my setup sort of high level [2:35] setup looks like this where these are [2:37] like five stages where you have the [2:39] setup stage where we basically freeze [2:40] the code base scope, [2:42] you know, give the context about the [2:44] known issues, the in scope impacts [2:47] assuming it's a but bounty and then also [2:50] like you know things which might be out [2:52] of scope for example like permission [2:54] paths and things like that. And then we [2:56] start off with the hunt stage where you [2:57] have like different agents for example [3:00] and they can you know have like a [3:02] focused approach on different attack [3:04] variants and then so you it's kind of a [3:07] divide and conquer approach in that [3:09] sense and [3:10] and then then we go on to the third [3:12] stage which is the break stage where we [3:14] basically verify [3:16] issues you know through the use of like [3:18] POCs, [3:19] basically verifying the on-chain [3:22] you know contracts verifying against the [3:24] deployed code and things like that. [3:27] And then [3:28] we have the proof stage which is the [3:30] human de-duplication process where you [3:33] you know your harness basically lists [3:35] down all the leads that is all the [3:38] promising leads that it has generated, [3:40] but then it's up to you to basically [3:42] manually you know verify each lead and [3:45] see which you know issues are sort of [3:48] like duplicates and then [3:51] you know basically make a list of the [3:52] issues to further basically have a full [3:55] POC verification using the harness and [3:58] then have the judge basically give its [4:00] verdict and then obviously a manual step [4:03] is again to basically [4:05] you know figure out which issues do you [4:07] want to submit actually. [4:09] And this is a high level architecture [4:12] which is again about the five stages [4:14] that I talked about but it's just like a [4:15] bit more expanded in terms of [4:18] yeah like the different approaches again [4:20] the setup approach which I spoke about [4:23] about the context and you know the [4:26] threat modeling and things like that [4:28] which is you know passed as an input to [4:30] the hunters [4:32] and then you have [4:34] the POC verification state state that I [4:37] spoke about where you you know verify [4:39] stuff with actual POCs and then you have [4:43] the de duplication process which is [4:45] again a human stage right and then you [4:48] have the POC verification stage again [4:51] where you [4:53] where the human basically de duplicates [4:55] all the issues and passes it to the LLM [4:57] or the harness and then you and then [4:59] that stage basically verifies those [5:02] issues and it passes it on to the [5:03] judging side of things. [5:05] So that's how the whole things work [5:07] thing works and then you you know [5:08] basically select the issues manually [5:10] yourself. [5:11] Then what we have is like these are like [5:13] some different tools that I've seen [5:15] around or I've used which are pretty [5:17] popular [5:18] or I've seen they've been popular [5:20] around. So you can also like try this [5:22] I've listed down the patch of AI web 3 [5:24] security repository it lists all these [5:26] issues there are a lot more tools. So [5:28] you can you know [5:30] try and [5:31] play around with these and maybe use [5:33] some methodologies from these [5:35] repositories into your approach as well. [5:38] And then I was recently part of the Y [5:40] audit rocket fellowship where I [5:42] interacted with a lot of like OG [5:44] researchers and things like that and [5:45] they and I picked up some of the [5:48] things that they told me about which [5:50] have been [5:51] you know for example these different [5:53] approaches around the sync approach or [5:55] looking between looking for seams [5:57] between different components or you know [5:59] writing down where you took shortcuts [6:01] and things like that or try to expand [6:03] the low severity issues that you found [6:05] into a bigger attached service. So yeah, [6:08] I think that's been pretty much it and [6:10] I'm I'm going to be around if there's [6:12] any questions or anything. So yeah, [6:14] thanks a lot. Yeah.