DSS 2026 — Ill Bloom Joaquin Pereyra Source video: https://www.youtube.com/watch?v=OgkZzTOhU0Q Automatic captions — not human-reviewed. Complete spoken-audio coverage is unverified. Source language tag: en (English). Source export SHA256: e5aab74893af58b3cf2b10e2cb5b156c121f9e95cd19de36213b70c3a34abe24 [0:00] security there. Um, [0:01] hello. [0:03] While I'll wait for the slides. Okay, [0:06] quick. I was going to introduce myself, [0:08] but I have slides for that. I'm Joaquin [0:09] Pereira. I've been a researcher for more [0:11] than 10 years and I work at Coinispect, [0:14] which is a security firm that's been [0:16] providing services [0:18] in the blockchain space since 2014. So [0:21] that's basically forever in blockchain [0:24] times. [0:25] Um, [0:27] this talk [0:28] we are [0:31] we're going to talk about Earbloom. [0:33] Does someone know what Earbloom was? [0:36] Yes, you, because you read my yeah. [0:39] That's that's actually surprising, but [0:44] it speaks about the state of security [0:46] today. [0:49] The talk will have two objectives. One, [0:51] it's going to be a very technical [0:53] talk. I'm going to try to present [0:55] a lot of details about what a [0:57] pseudo-random number generator flow [0:59] looks like in the real world. [1:01] And [1:02] it's also going to be a story about how [1:04] to deal with an incident of that [1:06] of that style. It's not a common [1:09] incident response kind of [1:11] kind of thing because it's very [1:13] distributed. So I want you to imagine [1:15] you're [1:16] contacted to respond to an incident. [1:20] You it's a normal project for you. I [1:22] guess most people here work in security, [1:24] so you've done something like that. [1:26] And you realize the flow that you're [1:28] looking at [1:29] does not affect the single protocol that [1:31] contacted you. [1:33] It affects [1:34] you don't know how many accounts, you [1:36] don't know how many blockchains, you are [1:38] not sure how many millions of dollars [1:40] could be at stake here. [1:43] Also, you see evidence that it's being [1:45] exploited. It's [1:46] continuously being exploited while [1:48] you're researching [1:50] after the the first incident was [1:51] reported. [1:53] And [1:54] there's no easy way to actually know [1:56] who's affected and who isn't [1:58] because the [1:59] the vulnerability in a library. [2:02] It's actually, [2:04] as we'll come to find out, it's in a [2:05] library of libraries. So it's deep, [2:08] deep, deeply nested in many, many, many [2:11] applications. [2:13] So that's basically the talk and we're [2:15] going to try to [2:17] um [2:19] discuss how how we dealt with it so [2:21] hopefully you can learn about it. [2:23] And if you're you're faced with a [2:25] similar problem ever in an incident [2:27] response, you learn from our ideas and [2:31] our mistakes and what we tried and what [2:33] we [2:34] uh did. [2:35] So [2:37] El Bloom, for those that don't know, was [2:40] a problem in a pseudo random number [2:42] generator in crypto JS, which is [2:46] probably a library you are not aware. [2:48] You don't know, [2:49] but it's been uh live since 2014 [2:53] and it was really, really heavily used [2:56] like from 2022 14 [2:59] maybe till 2020 2021. [3:03] And uh it was really, really used before [3:07] native [3:08] um crypto was available in JavaScript [3:11] because remember there was a time when [3:12] the crypto module wasn't as available as [3:15] it is now. [3:17] And uh [3:18] the peak exposure was $12 million. [3:21] At least [3:23] almost $6 million were stolen, which [3:26] unfortunately in our industry that [3:28] sounds like [3:30] not a lot. [3:32] Um [3:33] but that's [3:35] that speaks about other things in the [3:36] industry. [3:37] Um we ended up fearing that almost at [3:40] least almost 3,000 accounts were [3:42] vulnerable [3:44] and uh [3:45] we analyzed 14 chains for signs of [3:47] exploitations and we found at least [3:51] five chains that were affected. So as I [3:53] said, this is all [3:55] we figured all this out afterwards. At [3:57] the beginning we didn't know if it was [3:58] $5 million, [4:00] $100k, [4:02] $1 billion, [4:03] you have [4:04] no way to know. [4:06] And um just so we understand the [4:08] vulnerability and we [4:11] can understand what we did to figure out [4:13] who was affected, let's do a quick intro [4:16] into BIP 39. Who is familiar with BIP [4:19] 39? [4:21] Wow, almost. Okay, so the quick intro is [4:23] worth it. [4:25] BIP 39 is a way that um [4:28] basically [4:30] you can have the same seed [4:33] to [4:34] control all your accounts in uh in [4:37] blockchains, right? It's a way that you [4:38] have when you initiate your wallet, it [4:41] tells you, hey, write down these 12 or [4:43] 14 [4:44] 24 words [4:46] and you can have accounts in Ethereum [4:48] and blockchain and and Bitcoin and third [4:51] chain and whatever you want and have [4:53] multiple addresses with the same kind of [4:56] 12 or 24 words or 18 words, but no one [4:59] ever used 18 words for some reason. [5:02] The way that works is uh first your [5:04] wallet will gather some entropy from [5:06] somewhere that uh it will gather [5:09] a random number. [5:11] I will do the examples with 128 bits [5:13] because it's easier, but there are [5:15] several lengths available. [5:18] And uh once your wallet has that [5:20] entropy, it will use that random number [5:23] to index into some words. So basically, [5:26] let's do the worst version possible. [5:28] Let's say the random number is 67 [5:32] and uh the the dictionary is food and [5:34] potato. [5:36] So your [5:37] index number six is food, index number [5:39] seven is potato, your words will be food [5:42] potato. [5:43] Something that not many people know [5:46] and it's a weird quirk of BIP 39 [5:50] is that it supports many languages. [5:53] Uh you can have your words in English [5:56] or Chinese or Spanish or French and [6:00] probably more. I don't know how many [6:02] languages exactly are supported. [6:05] And the weird part, [6:07] if you see the whole process, [6:10] is that [6:12] the language you have [6:14] will determine the master seed. [6:17] So the same entropy, if you have the [6:20] same random number [6:21] but a different language, [6:24] the words will be different and the [6:26] words it's [6:29] what the input for [6:31] PBKDF2 is. [6:34] So PBKDF2 [6:36] will have a different output if you have [6:37] different languages, [6:39] which is kind of surprising. Languages [6:41] seems like a UX thing, not a [6:45] master generating thing, [6:48] detail kind of thing. But that's how it [6:50] works. [6:51] I will try not to say PBKDF2 never and [6:54] again in my life if I can avoid it. It's [6:56] a mouthful, so I will just say hash. [6:58] It's not a hash really, but [7:00] similar enough. [7:02] And well, from there you have your [7:04] master key and there's some process to [7:07] derive infinite accounts from that [7:09] master key. [7:10] And the important part here is that the [7:12] hash function is designed to be [7:14] extremely slow and that's a feature. [7:16] That's to avoid [7:18] brute forcing even if you have a bad [7:20] random number generator and that will [7:22] also be very relevant. [7:24] So that's B 39. Let's get situated in [7:28] time here and there's something very [7:30] interesting with the timeline. [7:32] It's a very reduced timeline. [7:35] The crypto JS [7:37] issue was introduced in 2014, so very [7:39] old. [7:40] And it was [7:42] fixed in 2020. [7:45] And I realized [7:46] just an hour ago when I was reviewing [7:48] the slides that I kind of lied to you [7:52] because the title of this talk is [7:54] discovering Illum, but it was fixed in [7:57] 2020, so we really didn't discover it. [7:59] We rediscovered it. Or we were the first [8:02] good guys to rediscover it after the [8:05] hackers started draining funds. [8:07] The [8:09] weird part here [8:11] is that the first people to realize the [8:13] the flaw [8:15] didn't realize the potential of the flaw [8:18] or didn't realize how easy it was to [8:21] actually [8:24] crack the seeds [8:26] from the seed or from the bad seed or [8:28] random number generator. So [8:30] the maintainers of the library kind of [8:32] knew [8:33] that their pseudo random number [8:35] generator was bad. [8:37] They just thought it wasn't [8:39] as bad [8:40] as it turned out to be. [8:42] And that's [8:43] a very interesting point. Sometimes you [8:46] see this kind of phenomenon [8:48] where people know there's a [8:49] vulnerability and they will basically [8:53] trust that it's not as bad as it looks. [8:56] It ends up ends up being extremely bad. [8:59] And uh on May 2026 we were made aware of [9:02] a hack that was happening. We were [9:04] contacted to find the root cause. We [9:07] found it [9:08] and uh [9:10] all through the following months [9:12] until we were able to [9:14] basically warn as many people as as we [9:16] could [9:17] uh that hacker kept draining funds [9:19] unfortunately. [9:20] And uh the CVE was published in August [9:23] 2026. CVE publishing was [9:26] quite challenging but very important [9:28] because once you get the CVE [9:30] and many applications get automatically [9:33] warn that there's something wrong with [9:35] their dependencies. So that was [9:37] important for us. [9:39] Why did people kept using a library that [9:41] was [9:43] known to be vulnerable? I mean there was [9:44] a fix available and everything. This was [9:47] a time uh where mobile wallets were [9:50] being really popular and there were lots [9:52] of mobile wallets. [9:53] The market wasn't as concentrated as it [9:56] is now. [9:57] And it's very interesting because you [10:00] can go and see the GitHub issues. [10:02] You can go and read the discussions. [10:05] And the discussions are basically hey I [10:07] updated [10:08] and it it stopped working. My app is not [10:10] working anymore. [10:11] And the fix that was popular was hey, [10:14] just downgrade and it will work again. [10:16] Uh so many applications [10:18] just downgraded and uh were like okay, [10:21] works again. I can build my app. [10:23] Uh so everything's fine. [10:27] Um the maintainer even [10:29] kind of supported like uh [10:32] kept the zero random generator [10:35] in a new major version [10:37] saying hey, yeah, well, I can't update [10:39] the zero random number generator because [10:41] it breaks people up. [10:43] So here's a new version with other fixes [10:46] and that still works with React Native. [10:49] Uh so again, [10:51] kind of interesting to see how [10:55] if you don't put security first or if [10:56] you don't [10:58] try to prioritize security, [11:00] the fixes can be available, the [11:01] vulnerability can be known, [11:03] everything can be basically spelled out [11:05] for you. [11:06] You will make mistakes still. [11:09] So this this part was very surprising to [11:11] find. [11:12] And [11:13] also another surprising part of the [11:14] timeline, [11:19] 2020 was the year the fix was posted. [11:22] 2026 was the year [11:24] the first identified drain happened. [11:27] And um [11:29] that's a long time. [11:31] And that's actually actually a long time [11:33] even after [11:34] AI uh wasn't the scene. So it took [11:38] quite long for someone to to actually [11:40] start hacking [11:41] with the problem. [11:43] Now [11:44] let's go [11:46] to the problem at hand. [11:48] In the first beta version of this talk, [11:50] I was going to go by by line by line and [11:52] read the code. [11:54] Then I remembered that when a speaker [11:55] does that, I usually sleep. I don't like [11:59] it. I think it's very boring, very [12:00] confusing. There's always someone that [12:02] cannot read the lines because they're [12:04] too far away. [12:05] So what I encourage you to do, [12:08] go to the GitHub and go with your agent [12:11] through the code base. And in this talk, [12:14] instead of doing that, let's go through [12:16] a simplified conceptual [12:18] um version of the flow instead of going [12:20] line by line [12:22] like nerds. [12:23] Okay. [12:24] So the problem is [12:26] you want some randomness, as we said, to [12:28] generate a seed, right? You're a wallet, [12:30] you need to generate a seed, and your [12:31] user comes to you, hey, I want to use [12:34] blockchain. [12:35] Uh okay, here are your 12 words. What [12:37] you had to do was generate some [12:41] random number. [12:42] And this library promised you to give [12:44] you a random number. [12:46] What the library did was [12:48] take some input [12:51] and um [12:53] take the input was simply the amount of [12:55] randomness you wanted. [12:57] I don't know if you know, but randomness [12:59] is measured in [13:01] bits. So we say one bit of randomness, [13:04] 18 bits, whatever. [13:06] To be cryptographically safe, you need [13:07] 128 at the very least, at the very [13:10] least. That's why during the 2010s [13:13] um 12 words were very popular. Nowadays [13:16] we use 24 words. 24 words are 256 bits. [13:20] 12 words was uh 128. Well, basically you [13:24] told the library, hey, I want 28 uh 20 [13:27] 128 bits of [13:29] randomness, give them to me. [13:32] And the library did some [13:35] weird [13:36] math. [13:37] But basically, if you [13:39] reduce it, [13:41] you'll see that the library ended up [13:43] doing one call to math.random, [13:45] which is 32 bits of randomness, [13:48] and then a coin flip basically [13:50] every [13:52] four-byte chunk. So for 128 bits, that's [13:56] four bytes, four chunks. Yeah? [13:59] So we end up and the last for reasons [14:02] that are [14:03] not really relevant, [14:05] the last uh the last chunk had would [14:06] only one one coin flip, not two. [14:09] So we end up with only two to the 39 [14:13] space. [14:14] We wanted a space of two to a 128. We [14:17] end up with two to the [14:20] 39. That's way smaller, way way smaller. [14:24] Do the math. [14:26] It's it's actually worse than that [14:27] because for some reason there's a [14:29] multiplication in the in the code that [14:31] multiplies the high bits and the low [14:34] bits of a number. [14:36] If you do that, you'll get collisions [14:38] because imagine you have the number [14:40] bytes [14:41] two and four [14:42] that has the same output that if you [14:44] multiply the high and low bits of four [14:46] and two, which is kind of obvious when I [14:49] say it like that, but it was missed for [14:51] some reason. So after some math, you can [14:54] see this leaves you with 37 bits of [14:57] randomness. 37 bits of randomness is [15:00] awful. [15:01] By the way, I I do want to mention [15:03] something about the codebase. [15:06] I already have the slide, so why not? [15:09] If you go through this, you will see [15:11] math.random [15:13] and if someone [15:16] read the slide and saw that [15:18] maybe they thought, oh, this talk is [15:20] going to be math.random is bad. [15:22] Don't use math.random. Talk is over. [15:25] Great. Congrats. [15:26] Well, [15:27] the interesting part here is that [15:30] math.random wasn't the bad the the the [15:32] bad part. [15:33] I I think it would have been even best [15:36] to use [15:38] just use math.random instead of doing [15:40] all that they do because they introduced [15:42] like collisions and they introduced [15:46] one call to math.random no matter how [15:49] many bytes you wanted. So in the end it [15:53] was even worse [15:55] than just calling math.random. So yeah, [15:57] math.random is bad. Don't use it for [15:59] cryptography, but it can be even worse. [16:02] That's the good news. You're not the [16:04] worst if you use math.random. [16:08] So yeah, that's basically how you end up [16:10] with only [16:12] 37 bits of randomness when you wanted [16:15] 128. [16:17] For 256 bits, the math is the same. You [16:21] end up with 49 bits of randomness. [16:24] Um, [16:25] 47 if you end up [16:27] taking into account the collisions. [16:30] So, [16:33] remember we were doing an incident [16:35] response during this time. We found [16:37] this. Uh, we wanted to confirm if this [16:39] was actually what was causing the hacks. [16:43] People were being drained while we [16:44] researched this topic. [16:46] And if you ever participated in an [16:48] incident response, [16:49] you know it's quite stressful. You're [16:52] trying to tell it like logically, and [16:54] yeah, we found this, and then we found [16:55] that, and then we did that. [16:57] But the reality is much more chaotic. [17:00] So, we found this, [17:03] and we wanted to make sure um [17:05] that um [17:07] that this was actually the issue that [17:08] was [17:10] um being leveraged by the attacker. And [17:12] as I said, [17:14] um [17:15] sorry, yeah. [17:17] And we [17:20] kind of got lucky here [17:22] because we were able to confirm the the [17:25] issue because a user, a victim [17:27] actually shared their seed with us. [17:30] What's what was so important about this? [17:33] If you remember the BIP 39 slides, [17:36] the expensive part of the process [17:39] is the [17:40] PBKDF2 part. I said it again. Damn. Um, [17:47] if you can skip that part, [17:50] iterating over 2 to the 37 space is [17:54] quite cheap. You can do it quite easily. [17:58] So, once a user and [18:00] again, if you remember the slides of BIP [18:03] 39, [18:05] the hashing comes [18:07] after the seed. So, if you have a seed [18:09] like a raw seed, like the words, you can [18:12] iterate over that without doing the [18:14] hashing, the expensive hashing part, and [18:16] you can do it much more quicker. [18:19] So, once the user shared their seed with [18:21] us, [18:23] um the user was had already been [18:25] drained, so they had nothing to lose. [18:28] We found it matched [18:29] our [18:31] reproduction of the weak seed of random [18:32] number generator, [18:34] and we were quite happy because we found [18:36] the the actual root cause. We were [18:37] convinced at that point. [18:40] So what to do now? [18:43] Well, you have to find who is affected, [18:45] and not all the users will share their [18:47] seed, obviously. [18:49] Uh you have to find [18:51] what wallets were producing the [18:53] vulnerable seeds, [18:55] and you need to figure out if you can [18:57] save the funds. [19:00] How to find who is affected? Well, [19:02] basically we downloaded the whole [19:04] active blockchain address space of every [19:08] blockchain we could think of that was [19:11] more or less relevant. [19:13] And [19:14] uh you can see there are some stats, but [19:16] all your usual suspects are on there. [19:18] And the first idea is basically, well, [19:21] we iterate. [19:22] We go I equals 0 1 2 [19:26] and we generate every possible [19:28] candidate, and we try to see if that's [19:32] on the set of active blockchain spaces, [19:34] and if it is, well, we have a victim, a [19:36] potential victim, but they may not have [19:38] been drained yet. [19:40] There's an issue with that, which is uh [19:43] it's too slow. [19:44] Um [19:46] a search space of 37 bits or [19:49] even worse for to 256 bits of of [19:54] um [19:55] B 39 size. [19:57] It's bad for cryptography, but it's [19:59] still [20:00] quite big for [20:02] for [20:05] easy and response times. We didn't want [20:07] to take [20:08] three years on brute forcing [20:12] uh all the possible seeds. [20:14] Our interest was being as quick as [20:15] possible so we could be faster than the [20:18] attacker. [20:19] And I'll share some slides on how we [20:21] were faster than the attacker, which I [20:23] think it's very interesting at the end. [20:25] But [20:26] the important part for now is that uh [20:29] we thought of an optimization to this [20:32] process. [20:33] And the optimization is [20:35] quite obvious, which is use GPUs. The [20:37] important part is how do we did that. [20:41] GPUs are very good at computing PBKDF2 [20:44] again. [20:46] And [20:47] the problem was was how do we actually [20:50] match against the active [20:54] the active addresses on a blockchain, [20:55] right? There are two alternatives. [20:58] Either we compute everything on the GPU, [21:00] we store it somewhere, [21:02] and then we kind of see what the [21:06] middle of the two circles is. [21:08] And that's uh [21:11] that's the the the space is too big. [21:13] It's at least almost 20 terabytes. [21:17] It was too big to do that. [21:20] So we we what we ended up using was uh [21:23] Bloom filters. [21:25] Do you know who knows what Bloom filters [21:26] are? [21:27] Very few people. They are cool data [21:29] structures, so you should know more many [21:31] more people should know about Bloom [21:32] filters. Um Bloom filters are basically [21:35] a set. [21:36] And [21:37] the important part of this uh data [21:39] structure [21:40] is that it's able to tell you [21:43] it's a Boolean data structure, let's [21:44] say. [21:45] It can tell you two things. [21:48] You ask, hey, is this item on the set? [21:50] And it tells you, no, it's not on the [21:52] set. [21:53] Or it can tell you, meh, it's probably [21:57] on the set. Uh so you're not sure. [22:00] But you get a probably. And probably is [22:03] good enough for optimizations. [22:06] So [22:07] what do you gain because you lost like [22:10] the certainty that the item is in the [22:11] set [22:12] compared to an actual set data [22:13] structure? [22:14] What do you gain? You gain uh that is a [22:16] very compressed data structure. It uses [22:19] much much less space. So by using a [22:21] Bloom filter, we were able to put the [22:24] whole [22:25] the whole um [22:28] number of active blockchain addresses on [22:31] the GPU directly. [22:33] And the final [22:36] the final pipeline looked like this. For [22:38] every blockchain we were interested in, [22:40] we downloaded the set of active [22:42] addresses, which [22:44] by the way, [22:45] defining active is an whole another [22:48] ordeal, especially in chains like [22:50] Bitcoin, but that's [22:53] I could I could talk for that like for [22:54] another hour, so let's not do that. [22:57] Let's just say active. [22:59] Um we Bloom filtered it. We put [23:02] everything on the GPU [23:04] only if the Bloom filter tells us that [23:06] it's likely on the set of active [23:09] blockchain [23:10] addresses, then we leave the GPU [23:13] and confirm on the actual set. So that [23:16] allowed us to be quite quick [23:19] and allowed us to find [23:21] many many many many many vulnerable [23:23] addresses [23:25] quite quickly. [23:27] So that was the key optimization. [23:30] And it's also fun because of the name. [23:32] The name includes Bloom [23:34] uh and I'll share more about the name [23:35] later. [23:37] So once you get who is affected, you [23:40] need to understand which wallets were [23:42] the ones producing the vulnerable seeds. [23:46] Very hard. [23:47] Very hard. [23:48] We tried three things. [23:50] Again, I'm telling this like it's [23:53] chronologically. It wasn't [23:57] chronologically. [23:58] While we were attempting the the Bloom [24:00] filter thing and researching that, [24:02] other people were trying to find the [24:04] vulnerable wallets and so on and so [24:05] forth. So so keep that in mind. [24:08] The first [24:11] strategy maybe was we knew React Native [24:14] was relevant. [24:16] Uh React Native is very useful mobile [24:17] applications. [24:18] So let's try to reverse engineer APKs. [24:23] Absolutely a waste of time. It's very [24:27] hard not to reverse the IP case, but to [24:30] find [24:31] all the historical post all the history [24:35] of APKs of every wallet forever [24:38] was such a challenge. [24:40] Uh then you have to decompile them like [24:43] try to understand the code base [24:45] too hard. [24:47] Again, this is all with AI and and all [24:49] the whistles and everything. Uh it's [24:51] just [24:52] uh it didn't produce any results and it [24:54] took us a long time. So we quickly [24:56] discarded that. [24:58] In case it's not clear, [25:00] we did have to find every historical [25:02] release of every like [25:05] suspicious wallet [25:07] ever [25:08] because once you produce a vulnerable [25:09] seed, it doesn't matter if you produce [25:11] it in 2015 [25:14] or yesterday. The seed is still [25:16] vulnerable. So if you were vulnerable [25:18] for like a week and then you fixed it, [25:21] you still like had to warn your users. [25:24] So that was why um historical research [25:27] was very important. [25:28] Then uh something that's maybe the most [25:31] obvious idea, [25:33] many wallets are open source. Go to [25:34] GitHub, try to find if they're [25:35] vulnerable. Again, quite hard, quite [25:38] harder than it sounds. Even though this [25:40] produced a couple of results, uh you [25:42] again have to find all the historical [25:43] data. [25:45] And of course, only depending on the [25:46] vulnerable library doesn't mean you [25:48] actually used the vulnerable like code [25:51] path, right? So it was hard to analyze. [25:54] And maybe the most surprising strategy [25:57] worked the best [25:59] um which was [26:01] well, we have the accounts that were [26:03] vulnerable. Let's try to find what those [26:05] accounts have in common, like what [26:07] tokens do they trade, what country they [26:10] are in, [26:11] whatever, what NFTs do they have. [26:14] And then you could figure out the [26:15] community [26:17] and by proxy many times you can find out [26:20] like popular wallets among that [26:22] community. And that actually worked the [26:24] best, which [26:26] wasn't my initial suspicion. I I didn't [26:28] think that would work, but ended up [26:30] working. [26:31] Um, [26:33] I already said this, but uh you cannot [26:35] patch the seed. So, [26:38] again, if you I hope that's clear. If [26:40] you ever produce If a wallet ever [26:42] produced [26:43] um a vulnerable seed, [26:46] the seed is going to be crackable. [26:48] It It doesn't matter if the wallet fixed [26:50] the vulnerability afterwards [26:52] until the user creates a new seed on a [26:56] safe like pseudo random number [26:58] generator, that seed is going to live on [27:01] the searchable space [27:03] and basically be doomed forever. So, [27:06] there's something also very interesting [27:07] here. [27:09] This was a problem since 2014, right? [27:12] 2014 were different times for [27:15] information security, for blockchain, [27:17] for everything. [27:18] Many people were were [27:20] migrated after 2014. So, say you created [27:23] your first seed on 2015 and then you [27:25] migrated to [27:27] top tier wallet in 2018 because you [27:30] learned more about security and you [27:32] said, no, I don't trust this wallet. I [27:34] trust this other one. You bought maybe a [27:36] hardware wallet, but you kept your seed. [27:40] Um, you're doomed anyway. It doesn't [27:42] matter. Your seed is already like [27:45] vulnerable. [27:46] So, it was a hard problem for for for [27:49] wallets. [27:50] Meanwhile, while we are doing all this [27:52] research, the attacker is still [27:56] draining accounts. [27:57] The attacker first drained as I said [28:00] $3 million. [28:01] While we were we were doing the [28:02] research, a second wave hit. [28:04] That's almost $2 million. Well, more [28:06] more $2.5 million. And there was a small [28:09] very small drain, 40k [28:12] at the end of July. [28:14] So, this this whole research is being [28:16] done [28:17] kind of on a race with the attacker. We [28:19] were [28:20] trying to figure out what was happening [28:23] and contacting wallets and contacting [28:25] everyone [28:26] as the attacker [28:28] was exploiting stuff. [28:31] I will need to be quick because I only [28:32] have two left. So, let's go. [28:35] One important user had $2 million [28:37] on their account. And the important part [28:39] here is that [28:41] this was the first time we actually [28:42] considered rescuing the funds ourselves. [28:46] Um, $2 million is not pocket money. [28:49] And we wanted to be able to save the [28:52] user's money if possible. [28:54] Unfortunately unfortunately [28:57] it was impossible. Basically, we [28:59] consulted So, this is this is something [29:01] important also to know if you ever are [29:03] faced with a decision like this. [29:05] We consulted several firms, like legal [29:07] firms, [29:08] and none of them [29:10] had a clear path to move forward. [29:13] Like none. Absolutely none of them. [29:17] And we also contacted Seal 911 at the [29:19] beginning of this investigation. [29:21] And kind of [29:23] the idea was to we focus on the research [29:26] as much as possible. [29:28] And Seal 911 tried to [29:30] reach out to to they have lots of [29:32] contacts and were very useful in that [29:33] respect. [29:35] Um, so basically we tried to keep [29:38] kept things to to the research. But it [29:40] was [29:42] it was a hard night to sleep, so to say, [29:44] knowing that there were $2 million [29:46] for [29:47] someone and they were probably about to [29:49] be drained. And they ended up being [29:51] drained and we we suspect they were [29:53] drained. We can't know, of course. We we [29:56] just saw a transaction on the [29:57] blockchain. We we have no idea who they [29:58] are. [29:59] Um, [30:01] how to warn the users very quickly. [30:04] Again, the best way was through the [30:06] community. [30:07] And I [30:10] yeah, this was so [30:13] so hard for us, so frustrating to us [30:15] that we couldn't warn users about this [30:18] that we created Unlock Key. [30:20] Unlock Key is basically have have I been [30:22] puned for for keys. [30:24] So, in a privacy preserving manner, you [30:26] can go paste your address there and see [30:28] if your address is in a known vulnerable [30:31] set [30:32] of [30:33] addresses that are well in a set of [30:36] vulnerable [30:37] Uh and Unlucky is not only for E-Bloom, [30:39] it also contains a cold card addresses. [30:42] And we are sure that more pseudo random [30:45] number generator problems will happen. [30:47] So [30:50] basically we we think efforts like [30:53] Unlucky are very important. [30:55] Red thing times up. [30:58] Um [30:59] these slides I think are very cool. [31:02] There's a break now if you want to know [31:04] why we think we were able to actually [31:06] know about the addresses before the [31:08] attacker because that's also very [31:09] important. Like not to adopt a defeatist [31:12] attitude when you find something like [31:14] this. [31:15] Don't do not think okay attacker already [31:16] knows all is doomed. [31:18] Let's you know do the post mortem and go [31:20] to sleep. [31:22] You can do stuff yet. You can be smarter [31:24] than the attacker at least for a while. [31:26] And uh [31:27] times up. [31:29] But again there's a break now. I'll be [31:31] around if you want to discuss this. Uh [31:33] I'll be happy to do so. [31:34] Uh there's one part that's very [31:37] interesting and very technical about [31:39] pseudo random number generators that we [31:42] figured out before the attacker that [31:44] allowed us to crack even the 256 bits. [31:47] Yes I'm finishing. Um [31:50] but there's a success story for two [31:52] wallets that were able to like warn [31:53] their users and we [31:55] hopefully save people's uh savings. So [32:00] again we think Unlucky efforts like [32:02] Unlucky are very important. [32:04] Let's discuss if you have other idea [32:06] besides Unlucky which by the way is not [32:07] a product. We are not charging for [32:09] Unlucky is an idea. Go and check out the [32:11] site. [32:12] And uh [32:14] fun facts for the break. [32:16] Thank you.