DSS

DSS 2026 / Day 2 / Amber Hall

Auditing the Chain Itself: Pointing an AI at Consensus, Mempool, and P2P

Igor Gulamov

14:37 · Recording

Watch on YouTube ↗Open recording viewerOriginal programme entry
Download plain textDownload timestamped JSON

Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.

Transcript

0:01Uh hi. I am Igor from Seventh Chat,

0:05and I tell you about

0:08auditing not just smart contract with

0:10AI, but general software,

0:14on-chain and off-chain infra.

0:17Uh

0:19um so

0:21uh

0:24um first let's talk about Harmony one

0:28hack.

0:30What happens in

0:32the August?

0:34There are two root causes. First is

0:38malleability of cross-shard nodes, and

0:42second is

0:45zero of K

0:47consensus exploit. And

0:51attacker combine it together and

0:55uh

0:56mint

0:57a lot of one tokens

1:00that deal damage to Harmony ecosystem.

1:04Uh

1:04how such kind of hacks

1:08could happens? And we we haven't seen

1:12such hacks before. Uh

1:16these hacks are multi-domain. So you to

1:21to make this hack you should

1:24know well

1:26general cryptography.

1:29So uh

1:32um

1:33programming and

1:36consensus,

1:38game theory. And usually

1:42there is not so many specialists who

1:44know all these areas.

1:46Also in summer there are some

1:50vulnerabilities

1:51in zero knowledge protocols

1:55that's also

1:57related to the

2:00thing

2:01that to to hack zero knowledge, you

2:03should

2:05be strong in cryptography, in very

2:09experimental cryptography, of course,

2:11and in low level assembly like things

2:16and in consensus and so on.

2:19So

2:21this hacks could happen because AI

2:23becomes smarter, AI provides PhD level

2:29mass

2:30to everyone and

2:34hackers can now hack

2:37the things.

2:39In 2025

2:41the question was will AI replaces

2:44auditors? And now we know that

2:48the market

2:51is transforming and we have all

2:54processes, but it's just inertia and the

2:57process will be rebuilt around AI.

3:02And now we have a lot of news

3:07most of news are crazy like

3:10private private K equals four or

3:13Binance hack or CV with DES SMS and many

3:20others.

3:21But I think in the

3:25during a year

3:27this will be very small part of the

3:29list.

3:30So take a look at this picture. This is

3:34how intelligence is decreasing and

3:39two to five years ago

3:42it

3:43cost of intelligence decreasing twice

3:46per three months.

3:47Now it decreasing twice per 20 days.

3:52So it's

3:53decreasing not

3:56exponentially, it's super exponentially.

3:58So

4:00models become smarter, but

4:02the cost of inference for these models

4:06decreasing with acceleration,

4:09super-exponential acceleration. So

4:11singularity is

4:14very near.

4:16So

4:17next

4:19now we know that no layer

4:23is out of reach. So

4:24there are multiple hacks of unverified

4:28bytecode.

4:29Hackers can use

4:31decompilers and

4:34LLMs to make

4:37reverse engineering of smart contract

4:39and next they can

4:42audit, check the smart contracts, find

4:44bugs on them and hack the smart

4:47contracts.

4:48Also

4:50legacy sophisticated code is not a mode.

4:55Code could be very sophisticated

4:58for humans, but AI will read it,

5:01understand and find the things inside

5:04like

5:08PHP to shell exploit.

5:10And

5:12of course

5:13cross-domain things like Harmony hack or

5:18ZK hacks.

5:21Number of CVs growing and

5:25findings become more complex, dangerous

5:29and interesting.

5:33Valuable part of hacks during last year

5:36is not just

5:38smart contract hacks, but also

5:42blockchain infra hacks.

5:47So here are examples

5:51how it is going.

5:55Here is

5:58root cause

6:00find it

6:02by

6:03our

6:05AI agent Seven Chats. So we reproduce

6:10the Harmony Hack. How? We we load the

6:13code inside it and it finds two root

6:16causes from which the original attack

6:19was built. Of course, agent has no

6:23access to internet and

6:25no access to new commits of code. So it

6:31find out this from

6:34prefix code before on the state before

6:37anything happens.

6:39Uh

6:44Yeah. And

6:45speaking about resources needed to find

6:49the scenes, it's not so

6:53huge resources.

6:56It could be done with

6:59small budget during one day.

7:05How

7:07we can

7:09audit

7:12nodes and off-chain infra and

7:16other things?

7:18Smart contracts are small, so we can

7:22split the code base to not so many

7:26number of context windows and put it

7:29into AI and AI will get a lot of

7:32information from this context window and

7:36it's working well without

7:41complex enough

7:44pipelines. But if we are speaking about

7:47120k

7:49lines of code, it's not working.

7:52So we need to

7:59read this code with different

8:04tools and

8:06cheap

8:08AI agents

8:10smart enough to understand the code, but

8:14this agent shouldn't be smart enough to

8:16find the things. Why? To find point of

8:20interest and context windows.

8:25For example, we can just download

8:29tens of thousands repositories, mark

8:33with frontier LLMs how different parts

8:37of code connected together. Then we can

8:41run AI agent with prompt write the scene

8:46that can find deterministically the same

8:49connections between the code

8:51based on for example Tree-sitter. And

8:55next

8:56we get

8:57heuristic

8:59tool that allow us to

9:02split the code base to semantically

9:05correct

9:07context windows. And we can load this

9:12context windows to thousands or tens of

9:14thousands parallel agentic pipelines.

9:17These pipelines will generate

9:18hypotheses. Usually

9:21it generate

9:2310 or 20 hypotheses per each

9:28overlapping context window.

9:31And next we triage this hypotheses.

9:34First just reading the code and the

9:38hypothesis. Next executing

9:42the

9:43unit test showing the

9:46proof of concept in isolated sandbox.

9:55Here are

9:58all major incidents

10:01we reproduce it

10:04from August 2025

10:07in seventh chat.

10:10So in in some cases

10:13there there are only root causes. So,

10:15the the the place inside the

10:20code base where the bug happens. In some

10:24cases all attack flow or

10:30exploit.

10:32But the solution is mostly focused on

10:35auditing. That's why root cause is

10:38enough to prevent damage.

10:41If

10:42people

10:44run

10:46AI audit with enough quality before

10:49deploy and find out all the

10:53bugs in the code, then hacker will have

10:56no

10:59basement to build the attack.

11:03Also, this month we solved it even

11:06bench.

11:07It is now 96%

11:11and I think it's

11:15completely

11:16solved because GPT-5 is hallucinating

11:20sometimes and

11:22it could be 100 but

11:26there is a lot of random factor.

11:30Also speaking about

11:33abilities of AI comparing with humans,

11:37we can we have results 17 months ago

11:42on big contests in Sherlock. We out

11:45perform a lot of human researchers and

11:48take sixth place.

11:55Also we reproduce it PHP to shell.

12:00So, it's it's not blockchain. It's a

12:02general very sophisticated C++ code.

12:07But it find out the place where memory

12:11corruption happens.

12:12So you can load into such kind of AI

12:17agents any software,

12:21not just smart contract, not just node,

12:24but anything and it will find

12:28bugs in the code code base and write

12:31proof of concept as

12:33unit tests showing

12:36something going wrong.

12:41So the industry is changing.

12:45I think currently models can hack iPhone

12:49frontier models, but in

12:53six months we have these models

12:55executing on iPhone. What does it mean?

12:58That means that someone can build worm

13:01that can travel between devices and live

13:04into forms. First is like spores when

13:08dump worms

13:10like classic things infect computers

13:13with CPU and store the weights

13:17in

13:20memory,

13:22SSD and so on. And another form it if

13:25it's find something with GPU like

13:27iPhone, like MacBook, like desktop PC

13:31with strong GPU and it can execute

13:34itself sync and

13:37invent next generation of itself, invent

13:40next generation of dump worms.

13:43So

13:45I think there are a lot of risks like

13:47this and the way how to prevent it is

13:54formal verification of everything and

13:57shifting from

13:59AI role to describe the

14:02requirements of theorems for formal

14:04verification. So then we can prevent

14:07such things

14:09because if we

14:11have formal proof that

14:14invariance

14:16holding that

14:19this invariance cannot be broken by

14:21superintelligence.

14:26So thank you for your attention. Let's

14:30talk on DEFCON

14:32and DSS 2026.

Source and accuracy details

Source export SHA256: c23573452a9a623d89cd5e2cb05585ebc679ca7d097e360d160e2c4eb695fd8d