DSS 2026 — Auditing the Chain Itself: Pointing an AI at Consensus, Mempool, and P2P Igor Gulamov Source video: https://www.youtube.com/watch?v=SzeRuQUkV9I Automatic captions — not human-reviewed. Complete spoken-audio coverage is unverified. Source language tag: en (English). Source export SHA256: c23573452a9a623d89cd5e2cb05585ebc679ca7d097e360d160e2c4eb695fd8d [0:01] Uh hi. I am Igor from Seventh Chat, [0:05] and I tell you about [0:08] auditing not just smart contract with [0:10] AI, but general software, [0:14] on-chain and off-chain infra. [0:17] Uh [0:19] um so [0:21] uh [0:24] um first let's talk about Harmony one [0:28] hack. [0:30] What happens in [0:32] the August? [0:34] There are two root causes. First is [0:38] malleability of cross-shard nodes, and [0:42] second is [0:45] zero of K [0:47] consensus exploit. And [0:51] attacker combine it together and [0:55] uh [0:56] mint [0:57] a lot of one tokens [1:00] that deal damage to Harmony ecosystem. [1:04] Uh [1:04] how such kind of hacks [1:08] could happens? And we we haven't seen [1:12] such hacks before. Uh [1:16] these hacks are multi-domain. So you to [1:21] to make this hack you should [1:24] know well [1:26] general cryptography. [1:29] So uh [1:32] um [1:33] programming and [1:36] consensus, [1:38] game theory. And usually [1:42] there is not so many specialists who [1:44] know all these areas. [1:46] Also in summer there are some [1:50] vulnerabilities [1:51] in zero knowledge protocols [1:55] that's also [1:57] related to the [2:00] thing [2:01] that to to hack zero knowledge, you [2:03] should [2:05] be strong in cryptography, in very [2:09] experimental cryptography, of course, [2:11] and in low level assembly like things [2:16] and in consensus and so on. [2:19] So [2:21] this hacks could happen because AI [2:23] becomes smarter, AI provides PhD level [2:29] mass [2:30] to everyone and [2:34] hackers can now hack [2:37] the things. [2:39] In 2025 [2:41] the question was will AI replaces [2:44] auditors? And now we know that [2:48] the market [2:51] is transforming and we have all [2:54] processes, but it's just inertia and the [2:57] process will be rebuilt around AI. [3:02] And now we have a lot of news [3:07] most of news are crazy like [3:10] private private K equals four or [3:13] Binance hack or CV with DES SMS and many [3:20] others. [3:21] But I think in the [3:25] during a year [3:27] this will be very small part of the [3:29] list. [3:30] So take a look at this picture. This is [3:34] how intelligence is decreasing and [3:39] two to five years ago [3:42] it [3:43] cost of intelligence decreasing twice [3:46] per three months. [3:47] Now it decreasing twice per 20 days. [3:52] So it's [3:53] decreasing not [3:56] exponentially, it's super exponentially. [3:58] So [4:00] models become smarter, but [4:02] the cost of inference for these models [4:06] decreasing with acceleration, [4:09] super-exponential acceleration. So [4:11] singularity is [4:14] very near. [4:16] So [4:17] next [4:19] now we know that no layer [4:23] is out of reach. So [4:24] there are multiple hacks of unverified [4:28] bytecode. [4:29] Hackers can use [4:31] decompilers and [4:34] LLMs to make [4:37] reverse engineering of smart contract [4:39] and next they can [4:42] audit, check the smart contracts, find [4:44] bugs on them and hack the smart [4:47] contracts. [4:48] Also [4:50] legacy sophisticated code is not a mode. [4:55] Code could be very sophisticated [4:58] for humans, but AI will read it, [5:01] understand and find the things inside [5:04] like [5:08] PHP to shell exploit. [5:10] And [5:12] of course [5:13] cross-domain things like Harmony hack or [5:18] ZK hacks. [5:21] Number of CVs growing and [5:25] findings become more complex, dangerous [5:29] and interesting. [5:33] Valuable part of hacks during last year [5:36] is not just [5:38] smart contract hacks, but also [5:42] blockchain infra hacks. [5:47] So here are examples [5:51] how it is going. [5:55] Here is [5:58] root cause [6:00] find it [6:02] by [6:03] our [6:05] AI agent Seven Chats. So we reproduce [6:10] the Harmony Hack. How? We we load the [6:13] code inside it and it finds two root [6:16] causes from which the original attack [6:19] was built. Of course, agent has no [6:23] access to internet and [6:25] no access to new commits of code. So it [6:31] find out this from [6:34] prefix code before on the state before [6:37] anything happens. [6:39] Uh [6:44] Yeah. And [6:45] speaking about resources needed to find [6:49] the scenes, it's not so [6:53] huge resources. [6:56] It could be done with [6:59] small budget during one day. [7:05] How [7:07] we can [7:09] audit [7:12] nodes and off-chain infra and [7:16] other things? [7:18] Smart contracts are small, so we can [7:22] split the code base to not so many [7:26] number of context windows and put it [7:29] into AI and AI will get a lot of [7:32] information from this context window and [7:36] it's working well without [7:41] complex enough [7:44] pipelines. But if we are speaking about [7:47] 120k [7:49] lines of code, it's not working. [7:52] So we need to [7:59] read this code with different [8:04] tools and [8:06] cheap [8:08] AI agents [8:10] smart enough to understand the code, but [8:14] this agent shouldn't be smart enough to [8:16] find the things. Why? To find point of [8:20] interest and context windows. [8:25] For example, we can just download [8:29] tens of thousands repositories, mark [8:33] with frontier LLMs how different parts [8:37] of code connected together. Then we can [8:41] run AI agent with prompt write the scene [8:46] that can find deterministically the same [8:49] connections between the code [8:51] based on for example Tree-sitter. And [8:55] next [8:56] we get [8:57] heuristic [8:59] tool that allow us to [9:02] split the code base to semantically [9:05] correct [9:07] context windows. And we can load this [9:12] context windows to thousands or tens of [9:14] thousands parallel agentic pipelines. [9:17] These pipelines will generate [9:18] hypotheses. Usually [9:21] it generate [9:23] 10 or 20 hypotheses per each [9:28] overlapping context window. [9:31] And next we triage this hypotheses. [9:34] First just reading the code and the [9:38] hypothesis. Next executing [9:42] the [9:43] unit test showing the [9:46] proof of concept in isolated sandbox. [9:55] Here are [9:58] all major incidents [10:01] we reproduce it [10:04] from August 2025 [10:07] in seventh chat. [10:10] So in in some cases [10:13] there there are only root causes. So, [10:15] the the the place inside the [10:20] code base where the bug happens. In some [10:24] cases all attack flow or [10:30] exploit. [10:32] But the solution is mostly focused on [10:35] auditing. That's why root cause is [10:38] enough to prevent damage. [10:41] If [10:42] people [10:44] run [10:46] AI audit with enough quality before [10:49] deploy and find out all the [10:53] bugs in the code, then hacker will have [10:56] no [10:59] basement to build the attack. [11:03] Also, this month we solved it even [11:06] bench. [11:07] It is now 96% [11:11] and I think it's [11:15] completely [11:16] solved because GPT-5 is hallucinating [11:20] sometimes and [11:22] it could be 100 but [11:26] there is a lot of random factor. [11:30] Also speaking about [11:33] abilities of AI comparing with humans, [11:37] we can we have results 17 months ago [11:42] on big contests in Sherlock. We out [11:45] perform a lot of human researchers and [11:48] take sixth place. [11:55] Also we reproduce it PHP to shell. [12:00] So, it's it's not blockchain. It's a [12:02] general very sophisticated C++ code. [12:07] But it find out the place where memory [12:11] corruption happens. [12:12] So you can load into such kind of AI [12:17] agents any software, [12:21] not just smart contract, not just node, [12:24] but anything and it will find [12:28] bugs in the code code base and write [12:31] proof of concept as [12:33] unit tests showing [12:36] something going wrong. [12:41] So the industry is changing. [12:45] I think currently models can hack iPhone [12:49] frontier models, but in [12:53] six months we have these models [12:55] executing on iPhone. What does it mean? [12:58] That means that someone can build worm [13:01] that can travel between devices and live [13:04] into forms. First is like spores when [13:08] dump worms [13:10] like classic things infect computers [13:13] with CPU and store the weights [13:17] in [13:20] memory, [13:22] SSD and so on. And another form it if [13:25] it's find something with GPU like [13:27] iPhone, like MacBook, like desktop PC [13:31] with strong GPU and it can execute [13:34] itself sync and [13:37] invent next generation of itself, invent [13:40] next generation of dump worms. [13:43] So [13:45] I think there are a lot of risks like [13:47] this and the way how to prevent it is [13:54] formal verification of everything and [13:57] shifting from [13:59] AI role to describe the [14:02] requirements of theorems for formal [14:04] verification. So then we can prevent [14:07] such things [14:09] because if we [14:11] have formal proof that [14:14] invariance [14:16] holding that [14:19] this invariance cannot be broken by [14:21] superintelligence. [14:26] So thank you for your attention. Let's [14:30] talk on DEFCON [14:32] and DSS 2026.