DSS 2026 / Day 1 / Amber Hall
The Audit That Never Ends: Evergreen Security in the Age of LLMs
Sergey Khomutinin
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:02Okay.
0:04I guess we can start finally.
0:06Um yeah, I'm Sergey. I'm Solidity
0:08developer at Trail of Bits and I'll show
0:10you how we approach the idea of a
0:12never-ending audit.
0:15We all need audits, but
0:18I think we are doing them wrong. We
0:19still treat audits as a one-time check
0:22near the end of a development.
0:24Um
0:25we write a code, uh freeze it, send it
0:28to auditors, get a report,
0:30uh fix the bugs, and then just repeat.
0:33Uh
0:35but what about the off-chain
0:36applications, UI, integrations, and a
0:39lot of that code
0:41commonly never gets audited at all.
0:44Um
0:45what if audit became a part of a
0:47development process uh starting from
0:50development up to the release and even
0:52further, we can keep looking for
0:53problems even in the code that is
0:55already live.
0:57And uh
0:59Frontier RLMs allow us to do so.
1:03Um
1:03they are pretty much capable of doing
1:05security work today and finally with
1:08Frontier RLMs you don't need a complex
1:10system to start doing this work.
1:14Uh they can easily check invariants,
1:17check assumptions, and uh
1:19check how components interact with each
1:22other.
1:24So that's how we built Phalanx, our
1:27internal defensive system
1:29that uses just uh
1:32a common
1:33coding agent such as Cloud Code and
1:34Codex under the hood.
1:36And uh
1:38we just give it the right context, uh
1:40ask it the right questions,
1:42and it tracks our projects, it runs
1:45checks,
1:46and it gives us one place to review all
1:49of the findings.
1:52Um
1:56we don't just run uh audits once. We
2:00um
2:01we on all of the changes.
2:03Changes in the code, changes in the
2:05documents, maybe information about the
2:07upcoming hard forks, and even just new
2:10LLM releases, all of that give us
2:11reasons to check our code again
2:14and try finding something something new.
2:17And
2:18between these checks,
2:20agents just keep researching your
2:22questions to go deeper and try to find
2:24find something deeper.
2:29We don't just collect findings. They
2:31have their own life cycle. Something's
2:34fixed, something is still open,
2:36something can be reopened to check again
2:39by humans.
2:41And
2:45let's
2:47continue, please.
2:50Yeah,
2:51as well as
2:54we check all all of the snapshots of our
2:57code, development branches, upcoming
2:59releases, as well as production, so we
3:02can find problems earlier, at earlier
3:05stages.
3:06And
3:07we don't need just to wait for the final
3:09audit to
3:11discover security problems.
3:14So what's Lens actually code for us?
3:18More than we can cover in two minutes,
3:20of course, but I'll just show you two
3:22public examples. First one is
3:25when we
3:26checked our
3:28code for a Gnosis compatibility, Lens
3:30just found a bug in the consensus specs.
3:33Basically, validator deposits could go
3:35to our builder with the same public key.
3:37And here's the proposed fix, and it's
3:39already been merged a couple of days
3:41ago, which is great.
3:43And this one is
3:46interesting. It shows us why we still
3:49need
3:52checking the code even if it's not
3:54changed. Basically,
3:56Wormhole stopped its delivery service
3:58that our
4:00cross-chain governance relied on, so we
4:02just
4:02had to update the setup and clients call
4:04this problem.
4:09So um
4:11that's what we built in Lighter, but I
4:13think you can build something similar
4:16around what matters to you
4:18in house.
4:20And here's why. Um first is that you
4:23know the context around your code
4:25better. Uh you know all of your internal
4:28documents, why some design decisions
4:30were made this way were just previous
4:32findings.
4:34Uh you know the the scope and you know
4:37which contracts and which applications
4:38need to be checked together in one
4:40context.
4:41And you decide when to run the checks.
4:47So with with that context you should be
4:49the first to check your own code
4:52to
4:53um like researchers and attackers have
4:57the same LLMs and the same capabilities.
4:59Uh you just need to
5:01be sure that your goal is just to find
5:04them earlier, find security problems
5:06earlier before the bounty report or
5:08before the before an exploit.
5:10And yeah, LLMs keep getting better and
5:13your security baseline should keep
5:15getting better too.
5:17Uh do not hesitate to build your own
5:18version of
5:20an audit that never ends.
5:22Thank you.
5:24>> [applause]
Source and accuracy details
- Names, organizations, technical terms, code and numbers may be misrecognized.
- Source wording and disfluencies are preserved; no editorial corrections were applied.
- Export contains whole-second cue starts only; cue ends and subsecond timing are unavailable.
- A full caption-track export does not establish that every spoken word is represented.
- Automatic YouTube captions; not human-verified against the audio.
Source export SHA256: 9301c53d38ebdd6d44c11433032d69f8afc9731560e32556b083539db1bbd240