DSS 2026 / Day 1 / Amber Hall
Evergreen: continuous AI-augmented security by Lido x Nethermind
Eugene Mamin
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:02It might be me. Okay.
0:13Hi everyone.
0:17Okay, here's my presentation. So,
0:21I'm going to talk about Evergreen
0:23concept
0:24we built together with
0:26Nethermind as
0:29the thing that
0:30probably every protocol should have in
0:32the industry at the moment.
0:36I'm Eugene. I'm look for engineering at
0:39Lida Labs.
0:40You can follow me on Twitter.
0:43Nice to meet you all.
0:49Sorry, I
0:52yeah, technology.
0:54So, agenda is simple. I will start with
0:56numbers, then what we
0:59did with about that numbers and what you
1:03can start doing on Monday as a protocol.
1:08I'll start with audits
1:10and continue the line Yuri started.
1:15Lida did 30 audits this year
1:18with eight independent security firms.
1:22We have zero critical, zero highs audits
1:25work, but that's still not enough
1:28because
1:29like Yuri said, when you finish audit,
1:33it's just the beginning, not the end.
1:36Next one is Lida bug bounty. [sighs]
1:40So far
1:42from the very beginning of the program,
1:45we paid
1:46$410,000
1:50for white hats submitted.
1:52However, you can follow
1:54this particular arrow that initially
1:58white hats claimed to have like 180
2:03critical security findings.
2:06After triage on our side,
2:09it ended up with three critical payments
2:12actually made. So, most of the work
2:15here, unfortunately, is saying no, but
2:18still carefully.
2:23What changed recently? So, that's the
2:26whole graph of intake we have in bug
2:29bounty.
2:30You can see that starting from 2025,
2:33mostly late 2025,
2:35intake has
2:37largely increased and you can't just
2:40basically automatically beat up every
2:43submission.
2:44You have to have
2:47a senior engineer who
2:49who should say yes or no and why. And
2:52ideally, you should have a second
2:54opinion about each
2:56particular about each material finding.
3:03What else? So,
3:06it's not only the code that should be
3:09audited and reviewed and put on scope of
3:12bug bounty. It's also external world
3:15change while your code stays the same
3:18and it might be even the state of code
3:21remain the same.
3:23What I mean? It might be dependency
3:26change. So, suppose you rely on bridge
3:28aggregation and one of the bridges
3:30decided to change the
3:33approach for the for the relay
3:34mechanism.
3:36It might be your off-chain stack
3:38dependency. Of course, you can think
3:39about Oracles first, but that's not the
3:42only off-chain tooling you might have in
3:44your protocol.
3:46I mean it might might be auxiliary
3:47service which could also have its own
3:51weaknesses
3:53like race conditions, bugs, and things
3:55pretty hard to debug.
3:57And last but not least, you still rely
4:00on something when you build a protocol.
4:02For Lida staking, it's Ethereum itself.
4:04It's consensus spec.
4:06So when Ethereum undergoes hard forks,
4:10we must be sure that protocol
4:14would work correctly, would work
4:16according to the spec, and we won't be
4:18affected both from consensus layer side
4:20and from execution layer side.
4:23So that we start testing every hard fork
4:25early with devnets, with our very own
4:28devnets, then public devnets and public
4:30testnets and so on and so forth.
4:32Lately we found pretty interesting
4:34vulnerability in Glamsterdam hard fork
4:37and consensus layer spec.
4:39The thing is that if you had a deposited
4:44like not yet deposited but pending
4:46validator deposit,
4:48it could have been attributed to a
4:50builder with the same pub key resulting
4:53in
4:54fund loss for this deposit.
4:57Luckily this was fixed. Huge shout-out
5:00to Core devs. Huge shout-out to folks
5:03who supported our effort to make it
5:05fixed. And on mainnet this year we
5:08expect Glamsterdam it will be fixed.
5:13So that's why we started to build
5:15Evergreen.
5:17It's the idea of audit that never ends
5:21and that is
5:24always aware of
5:26everything happening around.
5:29We started with two tracks. One is
5:32external track. We had a pilot with
5:34Nethermind. And another track is
5:37internal. We built very like our very
5:40own tool Lida Lens.
5:45Starting from pilot,
5:48Nethermind team helped us to build
5:52a concept for Lida Core codebase. It's
5:55like 19
5:58K of code.
6:00It was four weeks with two security
6:03researchers, two AI engineers.
6:06Uh and the most of work was not about
6:08like picking the the best model. Of
6:11course, the most work was about
6:13properly
6:15uh gathering curated documentation and
6:18context.
6:21Uh so
6:22uh as a part of this experiment uh it
6:25was measured then when you have a proper
6:27context
6:29you have a real leverage over the model
6:32quality.
6:33So both uh detection rate increases and
6:36false positive rate decreases
6:39uh with you gathering context about your
6:41protocol, your dependencies and so on.
6:45Uh
6:47three lessons we learned from this
6:49pilot.
6:50The first one as I said context is the
6:52lever. So
6:54pay most attention about like uh what
6:57your design intent.
6:59Write it down because
7:02design intent is something that
7:06uh not agent, not security auditor, not
7:09a white hat can understand pretty well
7:12from
7:13out of nowhere
7:15unless you written it out.
7:17Uh and of course don't rely on
7:20on any particular engine. Use bunch of
7:23tools. They might be overlapping but if
7:26you have
7:28the whole family of tools um
7:31working together you get the most of the
7:33coverage.
7:36Uh you can learn about this uh
7:39particular pilot uh on Nazar Mine blog.
7:42Numbers,
7:44particular results, everything shared.
7:46So uh
7:49it
7:50it's available for like async
7:53investigation.
7:54Uh for our internal track we built Lidar
7:57Lens. Uh
7:59it's the
8:01over like it's the engine that works
8:04over the clock.
8:06It has 39
8:08uh repositories or projects at scope.
8:12And it has like just four
8:15separate AI models underneath
8:19that allows us to
8:21sleep better than we slept before.
8:27Okay. So
8:30the follow-up about Lidar Lens will be
8:33provided by Sergey.
8:35Please stay tuned. It will be in an hour
8:38or so.
8:40What you should start doing on Monday?
8:43Like prepare your context pack.
8:46If you don't have a team to spin up this
8:49particular like never-ending audit, then
8:52try to gather external team
8:55and
8:57of course agents propose, people
9:00approve.
9:01Approach works here.
9:04Audits is a snapshot. Security must be
9:07evergreen.
9:09That's pretty it. Thank you.
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: d042187b25c471b464f1217bcc01f1ed7cc3f20b36212f05cc77b8310e05ffc3