DSS 2026 / Day 1 / Amber Hall
The Danger Hidden in the Fog
Jelle "0xFlint" Goethals
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:03Hello, hello everyone, and welcome to my
0:05small talk on the dangerous hidden in
0:08the fog.
0:10N degree integrations.
0:12Now, what do I mean by this? If this
0:15light moves.
0:21[laughter]
0:23One.
0:25Okay.
0:28Ah, okay.
0:30So, what I mean by this? When we do our
0:32work security auditors, we always get a
0:35bounded scope.
0:37A piece of code that we have to verify
0:39if it is working correctly. This can be
0:41a completely new protocol or an update
0:43on an existing protocol, but it's always
0:46small and locally bounded. So, we make
0:49sure that the code does what it is
0:51supposed to, that there are no
0:52unintended consequences, and that every
0:55integration is working correctly because
0:58nearly all protocols in DeFi are
1:01integrated with external integrations.
1:04However,
1:06due to our limitation,
1:09however, every integration has its own
1:12integrations.
1:14You can integrate with Aave, but Aave
1:16can be dependent on spare on specific
1:18Oracles, or you can integrate with
1:21Chainlink or with an AMM pool which
1:23depends on the lending market.
1:25And we don't really check these extra
1:27integrations because we don't have the
1:29time.
1:31But even if we go beyond this,
1:33beyond this,
1:36where do I point to get this work?
1:41That's the
1:43One down. Yes, we go technology.
1:46So, but even if we go beyond like this
1:48internet connected mesh of protocols,
1:51underneath we also have a sea of
1:54liquidity. All the assets and asset
1:56pairs and the different protocols, which
1:58also affect the workings of protocols
2:02somewhere else. So even if there is no
2:04direct or indirect uh contact in the
2:07code, if liquidity in one's asset pool
2:10in one protocol gets affected somewhere
2:12else, this can still change a state
2:14which can become a problem for another
2:16protocol.
2:19And when you look at the big picture,
2:22what we actually do as security
2:23researchers is we verify manually or
2:27through through testing a very small
2:30local piece of code and the direct
2:32integrations.
2:33But DeFi is one gigantic connected graph
2:37and we implicitly trust that everything
2:40else is supposed to work or we say well,
2:43I hope it works because I don't have the
2:45time or the ability to check this.
2:47But this is a problem even those more so
2:50today than before
2:52because flaws which are multiple
2:54integrations away can affect the code
2:56that you're verifying. And you can say
2:59well, in my opinion this code is safe. I
3:01haven't found any flaws, but those are
3:03based on dependencies and assumptions
3:05which you can't really check.
3:07And dependencies multiple hops away can
3:10affect you and destroy the the code that
3:12you're verifying.
3:14And we see this more and more this year
3:17we've had like Kelp and Resolve, last
3:19year Stream and a few years ago Euler,
3:22very complex hacks where the
3:25the root cause was multiple hops away
3:27which makes it very difficult to verify.
3:32Now why don't we do this? Why don't I
3:34simply audit everything?
3:36Well one, naturally because I don't have
3:38the time.
3:39The work that I have to do is based on
3:41the amount of lines that I have to uh
3:43verify with some adoption for the
3:46integrations or complexity, but I simply
3:48don't have time.
3:50And secondly, well if it integrates with
3:52Aave or for example Lido, I'm not really
3:54going to check Lido or Aave itself. I'm
3:57simply going to assume that it works
3:59because it's a known and trusted
4:00protocol. If it's new I might spend some
4:02time, but it's also limited.
4:05And testing by itself doesn't really
4:07solve this, because this also is
4:10dependent on assumptions. Formal
4:12verification also at some point you have
4:13a specification which says, look, the
4:15these inputs can be go this minimum and
4:17this maximum and not beyond.
4:19So that doesn't solve really the
4:20problem.
4:22It's a difficult problem.
4:23And like every presentation today, we
4:26always have AI,
4:28because I think a reason why hacks like
4:30these are becoming more prevalent is
4:33that before it was really difficult to
4:35find it, because if one protocol has
4:37five integrations which has five and
4:38five and five, we have so many paths you
4:41can't really explore them manually.
4:44But you can build a harness which checks
4:46every single path, one out of a 10,000,
4:49and finds the one isolated path where
4:51there is that vulnerability or that
4:53dependency which a manual auditor can't
4:55really find.
4:57But as white hats, we can also build
4:59harnesses. So I think um and some
5:01companies already doing this, we also
5:03have to enable this and utilize this to
5:06not only check the code locally uh or
5:08with the first integrations, but somehow
5:11map all the possible integrations and
5:13make sure that to the second degree,
5:16third degree, fourth degree, we can't
5:19find an exploitable path or we can at
5:21least map the dependencies and the
5:23assumptions that we have to have to
5:25make.
5:26So three quick takeaways for all the
5:28security researchers. So when or
5:31developers, so when you're verifying
5:33something, it's important to take into
5:35account that your code is not just an
5:37isolated thing. It is a small node in a
5:40vast network and you can't really see
5:43far beyond.
5:45A flaw many degree degrees away can
5:47still destroy you.
5:49And lastly, AI can help us in the sense
5:51that it is lifting the fog. The black
5:54hats are doing it, so must we.
5:56Thank 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: 316222b553ee3c24b63b9b8b1227301a2b89cb409e3434b825b7dc6c5799805a