DSS 2026 — What a libp2p Gossipsub Crash Taught Us About AI Security and DeFi Infrastructure Manu Sheel Gupta Source video: https://www.youtube.com/watch?v=o2syYDTNAAc Automatic captions — not human-reviewed. Complete spoken-audio coverage is unverified. Source language tag: en (English). Source export SHA256: bdcfddc89f8a144ef7043e0f7a70370d2859cef91f4d73fcfd885feee782f735 [0:01] Good afternoon, everyone. I'm amazed to [0:03] see the kind of work everyone is doing [0:06] in security before the conference. [0:09] It's one of the responsibilities when [0:10] you are running an afternoon session to [0:12] kind of figure out ways through which we [0:14] can get everyone energetic. And I was [0:16] actually wondering that over the course [0:18] of last set of hours, [0:20] I think the everyone who was speaking [0:22] over here, they were like really [0:23] energetic. The kind of [0:26] thoughts and the kind of discussions we [0:27] had, it actually gave me the idea that [0:30] sure, Binance attack has happened, but [0:32] that's not the end of the world. I think [0:34] if you all come together, we can ensure [0:36] that we can have a better DeFi security [0:39] and a protocol system. So my talk is [0:41] essentially about the same. I'm Manu [0:43] Gupta from Li. P to P, [0:46] one of the maintainers and the technical [0:47] lead. And it is all focused on [0:50] preventive security measure. Preventive [0:52] security measure means a place where we [0:54] have prevented some kind of Binance hack [0:57] not to happen. So I am taking that into [0:59] consideration. [1:01] My talk is on gossip sub crash and the [1:03] two very key developments which I'll try [1:05] and cover over the course of talk. [1:07] One is on DeFi AI infrastructure and [1:10] second on AI security, which all of you, [1:11] my wonderful friends, you have been [1:13] working so well. [1:15] Before we get into some of these [1:16] glitches, I'll try and see if [1:18] things work out well. [1:20] Just one second. [1:24] Yeah, okay, fantastic. So a very quick [1:26] 101 on the infrastructure level I wish [1:28] to share. All of us have been discussing [1:31] about how we can ensure safeguard of [1:33] security measures at the smart contract [1:35] level and at the application level. [1:37] Fantastic. I mean, that's great. But I [1:39] also wish to share with you, there is a [1:40] very interesting surface beneath every [1:43] application layer in a decentralized [1:45] finance protocol. And that is [1:47] essentially what we call as the [1:49] infrastructure level. And infrastructure [1:51] level [1:52] is one of those things where you do not [1:54] have to attack. You do not have to [1:56] figure out keys to figure out where the [1:58] problem is. You do not have to go for [2:00] All you have to do basically is ensure [2:02] that the network stops talking. [2:04] And if the network stops talking in [2:06] distributed systems, we have this very [2:08] wonderful concept like there are three [2:09] attacks which I wish to share with you. [2:11] One is network partition, second is [2:13] Cible, third is Eclipse. If any one of [2:16] these attacks they happen, what we are [2:17] basically saying is [2:19] the entire assets that we have in DeFi, [2:22] they will not be hacked, you'll not be [2:24] moving across wallets. Basically the [2:26] network will stop working. And that [2:28] essentially is one of the things which [2:30] I'm worried about when I look into some [2:32] of these recent happenings at the [2:35] Binance attack case as well as some of [2:38] the previous ones. So what I'll be [2:39] trying to unfold is probably a [2:41] collective discussion and a [2:43] collaborative method. I think that's [2:45] what we should try and do during [2:46] afternoon sessions. We'll try and kind [2:49] of figure out what we can do better at [2:51] the infra layer and also at the [2:53] application layer using some of the [2:54] existing innovations that have happened [2:57] in AI and without AI, which means using [2:59] fuzzing as was been mentioned previously [3:01] by one of our friends. I would also kind [3:03] of share some of the key things we need [3:06] to understand at the [3:08] network level where we are talking about [3:10] very core methods of publish and [3:12] subscribe and also of peer [3:14] communication. So very quick 101 at the [3:17] infra level what we'll be discussing in [3:19] the course of talk one would be peer [3:21] communication. Peer communication via [3:23] gossip sub, flood sub and grid topology. [3:26] Second we'll be discussing peer routing. [3:28] So peer routing is via Academlia DHT. [3:31] Third we'll discuss about some of the [3:32] transport methods which we are using in [3:34] Ethereum. The most notably is quick [3:36] which is now now heading towards [3:38] becoming an ITF standard. We'll also [3:40] discuss about TCP, our connection [3:42] oriented system, some of the connection [3:44] less systems like UDP, web socket [3:46] secure. We'll discuss on WebRTC direct [3:49] which definitely Chromium deprecated, [3:51] but we have got a WebRTC direct V2. And [3:54] I would love to have some of you please [3:55] contribute at the specs of lib P2B to [3:57] try and see how we can ensure a better [4:00] participation in that. And then we'll [4:01] discuss how quick which is one of the [4:04] key methods which is used by all the [4:06] consensus clients of Ethereum. How we [4:09] can ensure a better security using [4:11] encrypted client hello and ECH ITF [4:14] standard which is going to be discussed [4:16] two weeks from now at ITS San Francisco [4:18] and I would love to meet some of you [4:20] over there and definitely looking [4:23] forward to catching up on that. [4:25] So first wish to share with you about [4:28] one of the [4:29] key learnings of D5 security management [4:31] and infra level. What we see basically [4:33] is [4:34] one of the basic problems in the way we [4:37] are writing code and the way a lot of AI [4:40] work and agentic work is happening is [4:42] that sometimes the training of data is [4:45] happening from what I'll call it as not [4:47] a sane data which means that agents [4:50] sometimes they're trying to do under [4:52] fitting and over fitting which is the [4:53] same problem of [4:55] like what we see in machine learning [4:57] they're trying to have that same set of [4:59] problems in infra level security [5:01] management which means if I look in the [5:03] last three months [5:05] the number of security issues that have [5:07] come to me at the infra level that is [5:09] equivalent to the total number of [5:11] security issues that I've received over [5:12] my career working at the infra level. So [5:15] that is the number that I'm seeing in [5:17] terms of the participation of security [5:20] issues. Now the point is what are these [5:22] security issues? I'll first mention that [5:24] and second I'll mention why many of [5:27] those issues are not relevant and why [5:29] many of those issues are like me too [5:31] issues and how we a maintainer how a [5:33] developer who is working at the infra [5:35] level they can ensure that they look for [5:37] those things which means like a in a [5:39] funnel operation they look only for [5:41] those issues where it's a make or break [5:43] scenario. So we'll try and work out the [5:45] optics for the same. So the first thing [5:47] essentially is like if you look through [5:50] this diagram what I have essentially [5:52] explained is that if your nodes are out [5:54] of sync and if there is a panic in one [5:56] of the nodes. So we have a distributed [5:58] network. [5:59] Just a very quick question. You are all [6:02] familiar with validator nodes and [6:04] consensus nodes and how execution and [6:06] consensus clients work. I mean, if there [6:08] is any problem, we'll definitely discuss [6:09] on that. But I'm [6:11] assuming one particular part while I'm [6:13] discussing this diagram. So we have a [6:15] distributed system. We have the [6:16] consensus nodes. They're all talking to [6:18] each other and what happens is there is [6:21] a panic that is getting triggered in one [6:23] of the nodes. Now why this panic would [6:25] happen? This panic would happen by one [6:27] of the very basic things, integer [6:28] overflow, which was explained in the [6:30] last to last talk very nicely in the [6:32] symbolic execution talk also. So that is [6:35] one of the key things which causes a [6:37] panic. And once that panic happens, [6:39] people like as a long-lived network or [6:42] as a fault-tolerant network, we would [6:43] expect that Ethereum would kind of come [6:46] back to its original state very quickly. [6:48] Well, that is what we are trying to do [6:50] and that's why I'm saying that my talk [6:52] is a preventive security measure. But if [6:54] that does not happen, if there is a [6:56] quantum attack like what Justin very [6:58] recently mentioned, he's a good friend, [7:00] he shared about let's get into bunkers [7:02] mode. So just to ensure that we don't [7:04] get in even if you're in a bunkers mode, [7:06] we do the security well at the at the [7:09] bunker level. I do wish to share what [7:11] really happens is p2p state sync stalls, [7:14] the Oracle feeds halt, and liquidation [7:17] bots fail. Essentially, what we are [7:18] talking about is DeFi security coming [7:20] under serious threat. Now it's not like [7:23] I'm trying to make anyone of us. Like [7:25] I'm so happy to see the energy over here [7:27] and the kind of discussions and the kind [7:29] of cases we have discussed that I'm not [7:31] trying to share why this example that [7:34] this would happen in Ethereum. What I'm [7:36] trying to share is that these are the [7:37] things that happens when you have an [7:40] integer overflow and that integer [7:42] overflow causes a trouble in the overall [7:44] consensus client. So if a consensus [7:46] client gets broken, then essentially [7:48] what we are seeing is a DeFi problem at [7:51] scale. Now how do you ensure that that [7:53] does not happen is what we'll discuss. [7:56] So we'll take a specific case which is a [7:58] very popular one. It got a lot of [8:00] attention at NIST at all the key places [8:04] where we discuss about Ethereum security [8:08] and it was [8:09] being shared by everyone who was working [8:12] in Ethereum and I was essentially the [8:14] official spokesperson for doing all this [8:17] discussion around CBE. So I wish to [8:20] share first of all that this was one [8:22] issue which we faced in Lighthouse. [8:23] Lighthouse is like 27% of the Ethereum [8:26] or 30% of the Ethereum consensus client [8:28] market. It goes up to 43 now. So in the [8:31] last year you can say that in consensus [8:32] clients it's like 43%. It uses Rust [8:35] libp2p. We also have Prism client which [8:37] uses Go implementation of libp2p and [8:39] then we have some of the others like [8:41] Nimbus, Lodestar [8:43] and some of the lean Ethereum consensus [8:45] clients. So I'm taking one specific [8:47] scenario and I'm sharing this because we [8:49] were able to safeguard well. I mean [8:51] generally where we are not able to [8:53] safeguard well [8:54] it's not easy for me to disclose because [8:57] there's a lot of PR related discussions [8:59] that happen but it's important to [9:01] openly discuss some of those things. [9:03] So I wish to share if you talk about [9:05] this exploit what this essentially was [9:07] that we have a swarm of autonomous [9:10] agents and I'll also share how we can [9:12] create our own swarm of autonomous [9:13] agents. We basically had a very [9:16] interesting integer overflow function [9:18] and that was causing serious trouble. [9:20] Now the point is if you would think like [9:22] what [9:23] would really happen if there is an [9:25] integer overflow? So the funny part is [9:28] it basically causes a node to panic and [9:31] node to then go offline. So [9:33] the trace is basically the following. [9:35] You have an integer, you have the [9:37] arithmetic operations. As was explained [9:39] in the previous talk talk you have [9:40] either the bitwise operations or you [9:42] have the algebraic operations and you [9:44] have a specific node which has not been [9:47] defined or which has had a memory [9:50] overflow or a memory specific issue. So [9:53] if you have that kind of a specific [9:54] problem what happens essentially is a [9:56] fatal panic. A fatal panic means node [9:59] out of sync and node out of sync [10:00] basically means your DeFi protocol doing [10:03] behaving or misbehaving for the lack of [10:06] better word in a way that it should not [10:09] behave. [10:10] And [10:11] sorry, and there is some trouble with [10:17] yeah, [10:18] just one second. I'll just try and see [10:20] if we can move slides. [10:24] Okay. [10:26] Wow, sure. So [10:31] yeah, thank you. [10:32] Sorry, sorry for the hiccup. Yeah, wish [10:34] to share that essentially what that [10:36] basically means is that if a specific [10:38] security operation has happened in a [10:40] manner that it has caused the node to go [10:43] out of sync, what it basically leads to [10:45] is the entire DeFi protocol coming into [10:48] serious problem. And this is what we [10:49] have found with some of our friends [10:51] working at Aave, Morpho, and some of the [10:54] key DeFi organizations that some of the [10:56] key mentioned points or the spotlight [10:58] points that they do say is that if you [11:00] have the nodes in sync and behaving [11:03] correctly and you have to be right 100% [11:06] of the time. I mean, you get one time [11:07] you get incorrect and you have like some [11:09] integers behaving inappropriately, you [11:12] have the nodes going out of sync and you [11:14] have the entire network going down. So [11:16] here is a simple diagram I do wish to [11:17] share and there are two parts to it. One [11:20] is what is mentioned in this diagram. [11:23] There's another thing which I also wish [11:24] to explain. So this is essentially a [11:26] simple workflow on how we ran a swarm of [11:29] agents to do a code trace, to figure out [11:31] the root cause analysis, to do [11:33] coordination across different [11:35] implementations. At Libp2p, we have 17 [11:37] implementations. [11:39] Seven of those implementations are used [11:41] in all the Ethereum consensus clients, [11:43] which means if there is a problem at the [11:45] Ethereum consensus client level, then we [11:47] are to be blamed and we take full [11:48] responsibility of the same. So as a [11:51] person who is responsible for the [11:53] consensus clients, I mean, we do follow [11:55] some traces and this is one of that [11:57] specific configuration. [11:59] There is one thing which I've not [12:00] mentioned over here which I would love [12:01] to explain and I would love to have [12:03] everyone work and that is how we have [12:05] set up these autonomous agents. [12:08] The configuration and the setup is very, [12:09] very important and as we learned from [12:11] the previous talk, it becomes very [12:12] essential if we have to do the right [12:14] kind of analysis. So first, you should [12:16] be looking at agency directory service. [12:18] You should be looking at how you can [12:20] create a directory service of agents [12:23] where they have a routing, discovery and [12:25] a communication system. And I would [12:27] strongly recommend going back to David [12:29] Mazieres' time at Stanford University [12:31] where he was working on Kademlia DHT. [12:33] Use that as one of the key methods for [12:36] intelligent routing. The second part [12:38] which I do wish to share is very very [12:39] significant and important is the [12:41] messaging between agents. And there are [12:44] two different ways of doing that. One is [12:46] in many organizations we see that they [12:47] have problems with non-HTTP protocols. [12:50] So there I do wish to share if you're [12:52] trying to do an A2A or MCP, try and do [12:54] it over HTTP. And in many other cases [12:57] you can create your own secure low [12:59] latency messaging protocols. Sumanjit [13:01] who just presented on one of the [13:03] interesting talks at PII P2P four five [13:05] talks earlier, he very nicely explained [13:08] how we can have these nodes [13:09] communicating and talking to each other [13:11] and kind of sharing in real time. And [13:13] the third thing which I do wish to share [13:15] is identity interoperability of agents, [13:18] which means if you want Ethereum nodes [13:20] to not just function between your [13:21] computer and one or two computers, but [13:23] if you want to really do it as a [13:24] collaborative action operation so that [13:27] anyone who is trying to approach you [13:28] from a nation state is not getting their [13:32] kind of benefits. If we start operating [13:34] together as a team, so the third part [13:36] essentially is having identity [13:38] interoperability of agent taxon. So [13:40] there I would recommend try and use some [13:42] of the IETF standards like OAuth which [13:44] is OAuth 2.1 or you can use OpenID for [13:47] doing identity interoperability. And the [13:49] third fourth piece essentially is the [13:52] observability part. Observability in [13:54] agentic ecosystem is one of the key [13:56] essence. I mean it's not just about [13:58] tracing or OTel or telemetry logs, it's [14:00] the entire action agent state capsule. [14:03] And we would be presenting about it at [14:05] IETF SF and would love to have your [14:08] thoughts on the RFC that we have. [14:11] So moving forward if I can, I think I [14:15] Yes. [14:16] Sorry, I just went one step back. [14:21] Okay, fantastic. So we are moving [14:23] forward. Why I'm saying that we should [14:26] try and use AI security? Why I'm not [14:29] talking just that we should look at [14:30] traditional fuzzing? I think there are [14:31] some very, excuse me, [14:33] there are some very interesting benefits [14:35] from the AI models, the way they have [14:38] been approaching. I think one is that it [14:39] kind of reduces the cognitive load on [14:42] all of us as security researchers. I [14:44] mean, if you are a researcher and we are [14:46] working through massive loads of data, [14:48] one of the key things we face is that [14:51] the cognitive load is something we all [14:53] wish to address. And triage, if you [14:55] really believe my friend, is right now [14:57] the product. [14:58] I mean, the amount of as a maintainer, [15:00] the number of security issues we are [15:03] receiving in the last six months and I I [15:05] I can definitely suspect in the last [15:07] next six months it will be almost 10 [15:09] times more than what I have received in [15:10] the last three previous six months. I do [15:13] feel that [15:14] one of the key things over here is how [15:17] we set up the ecosystem, how we keep a [15:19] balance between fuzzing that is [15:21] traditional fuzzing and also the agentic [15:24] forms and try and see where it works and [15:27] where it does not work. And I do wish to [15:29] share with you the inventor of fuzzy [15:30] logic was interestingly my advisor [15:32] friend Dr. Professor Zadeh at the [15:34] University of California, Berkeley. And [15:36] he said something very interesting and [15:38] that is that [15:40] fuzzy logic is very similar to is very [15:42] similar to words with computing or [15:44] computing with words. [15:46] If you use the logic systems in a [15:49] systematic manner, you do not need any [15:51] AI model. And he used to openly say that [15:53] and we have seen that it works very [15:54] well. And in many cases where you have a [15:57] logic task, I would still say go for [16:00] traditional fuzzing. [16:02] Agentic systems or agentic workflows are [16:04] fantastic when you are working with [16:06] differentiation, calculus, tensors, [16:08] mathematics, isolated algebra. So there [16:11] I do wish to share the AI systems are [16:14] doing fantastically well. Why? Because [16:16] they are been trained on certain things. [16:19] And on those set of things they have the [16:22] right kind of expected outcomes. In [16:24] reference to logic paths I still feel [16:27] the human mind and human brain works [16:29] better and we have seen that across [16:32] across number of exams. [16:36] Yeah, just a second. [16:42] Yeah, sorry for the hiccup. While this [16:45] comes up, I hope it comes up. I'll [16:48] definitely [16:50] share [16:52] it's just going the opposite. [16:59] Yeah, I'll just move to the next set of [17:01] slides very quickly. [17:03] Sorry for the hiccup over here. There is [17:06] something seriously. Okay. So here I'm [17:09] discussing the security pipeline where [17:12] we are doing coordination with the [17:14] security team at Ethereum. So [17:16] essentially what we have is we have [17:18] Frederick and we have the consensus [17:19] client team who are based out of London. [17:21] And what we have essentially done is [17:23] that since we work across different [17:25] implementations and across different [17:28] modules. So what we are trying to do is [17:30] we are trying to ensure that it's a [17:32] coordinated effort. And coordinated [17:33] effort what it basically means that we [17:35] use AI for doing verification, for doing [17:38] auditing, for doing those set of things [17:41] where we can reduce the overall [17:42] cognitive load of our security [17:44] researchers. But for many things we are [17:46] still kind of embarking or following a [17:49] very simple strategy that anyone who has [17:51] a good interest or has a good experience [17:54] in the deep tech understanding of infra, [17:56] which basically encompasses your [17:58] transport layer, your identity layer, [18:01] your messaging layer, communication [18:03] layer, routing system, secure [18:06] communication channels, which is TLS [18:08] noise and auto TLS. So someone who has [18:10] deep interest on these things, we are [18:12] basically sharing with them that okay, [18:14] you kind of collaborate with us, work on [18:17] a module by module level, try and see [18:19] and understand from the existing CVEs [18:22] which have already been issued for the [18:23] last I would say seven to eight years. [18:25] So we do have a good human knowledge of [18:27] last seven to eight years and I would [18:29] say the first four five years when there [18:30] was no LLMs. I mean even knowledge at [18:33] that particular time in terms of [18:34] triaging, in terms of collaborating. So [18:37] we have that eight to 10 years of [18:39] knowledge very well developed in terms [18:41] of how Ethereum security has been [18:43] managed and how the DeFi security around [18:45] Ethereum has been managed. So keeping [18:47] that into consideration, keeping that [18:48] into perspectives and building on the [18:50] learning that we have so nicely [18:53] I would say [18:54] documented and also made it sure [18:58] also ensure that it is available to [18:59] everyone who wishes to join. We are [19:01] saying to everyone and anyone who is [19:04] interested in doing infra level [19:05] security, join us, collaborate in some [19:08] of these efforts, try and see and make [19:10] it try and see on your system whether [19:13] you can evolve your understanding of how [19:16] things work at a distributed node level, [19:18] how you can fix some of these issues and [19:20] how you can ensure that peer-to-peer [19:22] security really works. And this is one [19:25] of the key learnings which I do wish to [19:27] share that there were a lot of members [19:29] who when I was having coffee outside, [19:32] they mentioned that the work in AI [19:34] security or AI research is now getting [19:38] been done more by AI models and less by [19:40] human beings. I personally believe it's [19:41] quite the opposite. [19:43] In [19:44] in reference to the infra level [19:45] security, we can do a much better job if [19:47] we have some of these models. What I [19:50] essentially wish to share in reference [19:51] to this case was that we use AI agents [19:54] to figure out a panicking of a node. Now [19:57] node panicking is like one of the things [19:59] which we can easily miss. Why? Because [20:02] memory overflow and integer overflow are [20:03] those cases where sometimes a human [20:06] brain will say that okay, these are some [20:08] of the calibrations that the original [20:10] author would have done and I don't have [20:12] to run those kind of code checks or code [20:14] traces. [20:15] And unless you are a symbolic [20:17] researcher, I mean unless you have built [20:18] a symbolic [20:20] system where you are doing an if else [20:21] construct and you're trying to really go [20:23] at that level, doing an arithmetic level [20:25] overflow research is harder. So, what we [20:27] found is that if you are using an AI [20:30] agent and you have a very basic integer [20:32] overflow, [20:33] you can have it fixed in one [20:34] implementation that is rust and then you [20:37] can have a very clear pathway through [20:39] which you can share it across all the 17 [20:42] implementations. So, we have in total 17 [20:44] implementations and the good thing, the [20:47] benefit of AI is that the security team [20:50] remain as it is, which means if there [20:52] are like 20 contributors and 20 security [20:55] researchers in the overall libp2p [20:57] project, the core ones I'm talking [20:58] about, the ones who are being paid for [21:00] doing [21:01] the entire bug fixing, maintenance, code [21:04] management. So, those members are [21:06] basically coordinating faster [21:09] and they are able to coordinate faster, [21:10] they are able to ensure that the final [21:13] code trace that is available, I mean [21:16] they are able to kind of ensure that [21:17] those exploits are not [21:20] very specific to one implementation, [21:22] they're also shared by other [21:23] implementations. Why this becomes very [21:25] important is that the like the attack [21:28] will not come just from one specific [21:29] consensus client that is a lighthouse [21:31] consensus client. It can come from say [21:33] like Teku which is an institutional [21:35] level Ethereum client that uses Java [21:37] libp2p. It can come from Nimbus which is [21:39] using Nim libp2p, another very [21:41] interesting language. It can come from [21:44] Lodestar which is using Zig libp2p or JS [21:46] libp2p. So, the point is that [21:49] the first part is to fix that issue, but [21:51] the second part is also ensuring that [21:53] whether that fix has been propagated [21:55] timely to the other implementations and [21:58] to the other consensus clients because [22:00] if one of those implementations have [22:02] that integer overflow or has that [22:04] specific issue, then this entire [22:06] exercise of using an automat automated [22:09] agentic workflow system, it does not [22:11] reap those kind of benefits. If one of [22:13] the consensus clients is in problem, so [22:15] it's like one of your family members has [22:17] a impact, then you definitely get [22:19] impacted by that. So, it's the same set [22:21] of things and I do wish to share the [22:24] final uh piece [22:26] over here. [22:27] Sorry. [22:28] We had a tough time today, but yeah, [22:31] this is the last slide. I just wish to [22:34] share that [22:35] most important thing is that if we have [22:37] the AI execution done properly, if we [22:40] have the AI security workflow analysis [22:42] done properly, then [22:43] there couple of things to learn from [22:45] this. One is build your triag pipeline [22:48] so that it is not like let me accept AI [22:50] or let me accept AI models, but it is [22:52] like where I need AI. [22:54] I mean, that is essentially the question [22:56] I would like to share with everyone. [22:59] It's not the question of I want AI to [23:01] use security or to do better in [23:02] security. It's like where I need it. And [23:05] where I don't need it, I will use the [23:07] regular human brain. I will use my [23:09] friends and collaborators. I'll work [23:11] with them to ensure that we get to a [23:14] good conclusion to a specific issue. So [23:16] these are like some of the learnings [23:18] which I do wish to share. I do feel that [23:21] many of the things which I shared in [23:22] this discussion, you might be feeling [23:24] that hey, this is very specific to how [23:27] we work at infra level security for the [23:30] DeFi [23:32] for the DeFi project. Yes, definitely it [23:34] is very much related to infra level [23:35] security, but I also feel that it has a [23:38] lot of implications to what we guys are [23:40] doing at the wallet level and at the [23:42] other different levels. That is [23:44] essentially your application level, your [23:46] DeFi protocols. They all are following [23:48] the similar trend and some of the [23:49] learnings we did from the infra layer is [23:51] something which we have kind of [23:53] propagated to the application layer. So [23:55] in case if you are working and you are [23:57] like a do it yourself person who likes [23:59] to run nodes, Ethereum nodes, likes to [24:02] run validators, I would definitely like [24:04] to share two things with you. That's a [24:06] friendly guest. Like that's one of my [24:08] responsibilities. One is you please [24:10] upgrade your gossip sub to 0.49.4 which [24:13] is the latest version. And if you're [24:15] using floods of fantastic, if you're [24:17] using grid topology, I mean fantastic, [24:19] you don't have to worry, but [24:21] do the same. And second essentially is [24:24] that you also prepare your triag [24:26] pipeline for the AI security swarm. So [24:29] essentially try and prepare your AI [24:32] security swarm track pipeline. And yeah, [24:35] if you have any questions, [24:36] please let me know. And also one last [24:40] request, if anyone of you is interested [24:41] in applying for an AI grant or for a [24:44] infra level grant, please do let me [24:46] know. We do have some opportunities. [24:47] Would love to discuss. Thank you. [24:50] [applause] [24:57] You can speak. [24:58] >> I just want to understand that you [25:00] mentioned that there are multiple [25:02] like [25:04] let me do in Java, let me do in [25:07] >> Different implementation languages. [25:09] Yeah, yeah. [25:10] >> So [25:10] I wanted to understand does it make [25:12] sense to have a [25:14] central document [25:15] >> Mhm. [25:15] >> which has the specs written out and that [25:18] is useful for me to verify or [25:21] symbolically [25:22] figure out if the [25:24] >> Absolutely. [25:25] Beautiful question. I do wish to share [25:27] we do have specs. It's called as let me [25:29] to be specs. More than happy to share [25:30] the link. And I do [25:32] so that is that gives a very so that's a [25:34] very interesting question. I mean we are [25:36] working in a decentralized environment, [25:38] but at the same time there are points of [25:39] convergence. Let me to be spec is the [25:41] point of convergence. That's where all [25:43] of us we kind of collaborate. We discuss [25:46] for every specific function how it has [25:48] to be written or for a specific [25:49] architecture how it has to be worked on. [25:51] For any spec like for any spec proposal [25:54] which is shared as an RFP or even if it [25:56] is shared as a PR in our specs repo. So [25:59] we have a very clear path of review [26:01] cycle. So it's generally reviewed by all [26:04] the maintainers of different [26:05] implementations. We have a weekly spec [26:07] meeting. Would love to have you join [26:09] that spec meeting also. We are now [26:11] having it more often than usual because [26:13] of a lot of security issues coming our [26:16] way. Many of them are minor issues, many [26:18] of them are me too issues, but some of [26:19] them are definitely high severity [26:21] issues. And I do wish to share that yes, [26:24] having a common spec is one of the ways [26:26] through which we can align all these [26:28] different implementations. [26:29] Thank you. [26:31] All good? [26:32] >> Thank you. [26:33] >> We are just running short of time. So [26:35] we'll definitely catch up outside for [26:38] coffee if anyone would love to catch up. [26:40] Thank you.