DSS 2026 — Security Starts Before the Signature Steffen Kux Source video: https://www.youtube.com/watch?v=4fTqBpstVdM Automatic captions — not human-reviewed. Complete spoken-audio coverage is unverified. Source language tag: de. This tag differs from the handoff’s English description; language review is pending. Timing gaps requiring review: 8:09–10:10, 10:10–11:45, 14:16–15:01 Source export SHA256: a6264ea903553f5f7da5bea2f2a9f5cbc44db74259bb878db9deb64241b58a4c [0:12] So [0:13] this was this is just in Uniswap [0:15] transaction. I did [0:18] change one ether in another currency [0:22] and what I'm supposed to sign is [0:26] this data. [0:27] And even if I use my [0:30] wallet, my hardware wallet, it doesn't [0:33] make it much better, ne? I I have a big [0:36] display on my hardware wallet, but when [0:38] I look at it, it's exactly the same [0:40] information. [0:41] So what can we really know when we have [0:46] such information to make [0:49] a educated education that oder educated [0:52] decision that this is a transaction I [0:54] want really to sign. [0:57] A real problem. [1:01] Let's go back one year. [1:03] This time this was the biggest hack, [1:06] but this changes quickly. We had some [1:09] bigger hacks now, [1:11] but last year in February there was a [1:13] hack. It's [1:15] a little bit [1:16] not everything directly correlates to [1:19] the topic we will discuss today, but at [1:21] least we see the problem. [1:23] So in the Babil hack the problem was [1:25] there was a manipulated user interface. [1:29] And when designers signed [1:33] a very normal transaction moving [1:35] something from the cold storage to the [1:37] hot storage oder to the hot wallet, [1:41] they signed something which was not the [1:43] transaction they really [1:46] intended to sign. [1:48] And the result was [1:51] there were 1.5 billion dollars in tokens [1:55] lost because there was a transaction [1:58] which did not do transaction as [2:01] intended, but changed the logic of the [2:03] smart contract oder auf der auf der [2:05] Multisig. [2:07] Now we can say okay, this is one special [2:09] problem. There was a very direct attack [2:13] to do this in in this application, but [2:17] this is a problem we are confronted [2:19] everywhere. So we have transactions [2:24] and [2:25] when we want to be sure what this [2:27] transaction is doing, we need somehow to [2:31] simulate it. So in the former version of [2:34] the [2:35] Gnosis Safe you had the possibility to [2:39] simulate your transaction, ne? There was [2:41] a connection to Tendely. You could [2:44] simulate it. It was sent to their [2:45] servers. They did the simulation, [2:47] returned it to you. [2:49] But actually this is also only a shift [2:52] of trust, ne? When I ask someone else on [2:55] an OBC again a centralized service, I [2:58] have to trust that the simulation [3:00] results is ja really what I'm intended [3:04] to do. [3:05] So and it's [3:07] actually also a privacy issue because I [3:09] ja [3:10] reveal data before I do my transaction [3:13] to another party. So [3:16] I don't say that we cannot trust those [3:19] companies, but we also know that even [3:22] big companies sometimes are compromised. [3:24] Na, even if I trust this company in [3:27] person, but if the data is going there [3:29] and if it happens that they are [3:31] compromised, I would leak my data. [3:36] So signing is not enough, ne? Today [3:39] mostly we created transaction. [3:42] For instance, I'm using a DeFi tool or [3:46] exchange. Let's say we are using [3:47] Uniswap. I want to change one Ether to [3:51] USDC. Then I create a transaction. I see [3:54] it on my screen and then in my wallet I [3:57] have to sign something and I sign it and [4:00] hope that the transaction is correct. [4:03] What we need instead of doing it in this [4:06] simple way, we need first to create a [4:09] transaction of course, then we need to [4:11] verify that every information which is [4:14] used to create a transaction is really [4:17] correct. We have to simulate it and [4:19] analyze what are the [4:21] results and then I have to explain the [4:23] results so that I really understand [4:25] these results belong to this transaction [4:28] and then I really know what I'm going to [4:30] sign and then I can do an educated [4:33] decision what I really sign. [4:38] So our topic from Colibri [4:41] we are building a stateless client and [4:43] we say don't trust any information you [4:45] get, always verify before you use it. [4:49] And now we want to use this this topic [4:53] also to [4:55] to validate transaction before we send [4:57] them. [5:01] When I prepare a transaction normally I [5:04] need some data, I need some balances, I [5:06] need some information from the contract, [5:08] some state information. [5:10] So if I don't validate this information [5:13] I'm even if I have [5:15] verified everything else I'm still in a [5:18] trust system. So York other colleague [5:20] from my from me today mentioned this in [5:23] his presentation is that even if we have [5:26] an ordered the system with unordered [5:28] data it's an unordered system. [5:32] So how are we doing this? [5:35] So Colibri a stateless client is [5:39] a layer in between the RPC and my [5:42] application. That means I read data from [5:44] whatever source I want, let's say an RPC [5:47] or any other source and I also get the [5:51] proofs [5:52] so that I then I can verify that this [5:55] information is correct and we are using [5:56] two type of proofs. [5:59] So the first is the execution layer [6:00] proof, so I proof that let's say a [6:02] transaction is part of a block. [6:12] a belongs to a this block hash. And then [6:15] I know this block contains this [6:17] information and now I need also to prove [6:20] that this information belongs to the [6:23] canonical chain. And here I need the [6:25] consensus layer proofs. I need to prove [6:27] that the validators, the signatures of [6:29] the validators are valid signatures. And [6:32] this I can do starting from a checkpoint [6:34] in the past, verify all the transitions [6:37] of the validator committee up to the [6:39] block where I'm now so that I know I [6:42] have the correct signature. Or I can it [6:44] more efficiently have a zero knowledge [6:46] proof which proves this whole chain of [6:51] of [6:52] of transitions of the validator set. And [6:54] with these two proofs I can produce [6:58] verified state. So I don't need to trust [7:01] the information provider anymore. I just [7:03] get data, I get proofs and then I have [7:06] verified state. [7:08] So and based on this verified state I [7:11] can use an EVM. And of course also this [7:14] EVM should not be somewhere in the [7:16] cloud. I should have it in my [7:18] application. So Colibri also brings a [7:21] very small EVM in included so that we [7:24] now can the verified information, [7:26] execute the transaction in the EVM and [7:29] then have all the information which are [7:32] the results of my transaction. [7:35] What calls are done, [7:37] what is transferred, [7:39] how what are the state changes, which [7:42] events are [7:44] given and so on. [7:46] So we we do it directly in the [7:48] application. [7:51] But here we have another problem. So [7:52] when we list this to the user, [7:55] that is the same when I look into a the [7:58] report from Tenderly. So by the way, I [8:00] really love Tenderly. So as developer [8:03] it's a great tool to analyze when I do [8:06] an transaction with complex smart [8:09] contract interactions [10:10] Und hier kommt ein AI Model [11:45] So, ja. [11:57] We celebrity has been doing some years [12:01] before or explain us what what is a [12:03] elephant. We are not interested in our [12:06] wallet to to get this information. We [12:08] just want to have a model which is very [12:11] quick, very small and very smart in one [12:14] topic. And here we are using domain [12:18] small language models. These are small [12:21] LLMs which are trained with information [12:25] only relating to this topic. [12:28] And then we get very [12:31] good models which we can use and our [12:34] models are [12:36] about 2 GB and we also try to make them [12:41] even smaller and they give answers as [12:44] good as these frontier models because [12:47] they are trained with them and can [12:50] ja produce this good this good [12:53] information. [12:55] So what [12:57] is our [12:58] and what are the limits [13:00] we can verify all information we get [13:02] from the [13:05] from the blockchain. We [13:07] we can see if anything in this [13:10] transaction is something which might not [13:13] fit to our intent [13:15] but of course [13:17] we cannot solve everything. If the state [13:19] is changing before we included there [13:22] might be a different things. [13:25] Ja and we also need to understand what [13:29] my intent is. Now this is also something [13:32] we are researching right now not to make [13:35] a clear definition to see this is the [13:38] intent so that it is easy for the people [13:40] to understand that the intent is really [13:43] fitted. [13:44] So and with this we have a full workflow [13:47] from the data making the data trustless [13:50] with Colibri the the core module then [13:53] simulated in the local EVM analyze it [13:56] and extract all the information now for [13:59] instance to give the model the right [14:02] information we need also the code [14:04] the which is really affected from this [14:06] call. [14:08] And in this point it's not feasible to [14:11] give him all the code we find for [14:13] instance if you do a Uniswap transaction [14:16] the whole Uniswap [15:01] Thank you very much. [15:24] Thank you very much.