DSS 2026 / Day 2 / Pearl Hall
Security Starts Before the Signature
Steffen Kux
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:12So
0:13this was this is just in Uniswap
0:15transaction. I did
0:18change one ether in another currency
0:22and what I'm supposed to sign is
0:26this data.
0:27And even if I use my
0:30wallet, my hardware wallet, it doesn't
0:33make it much better, ne? I I have a big
0:36display on my hardware wallet, but when
0:38I look at it, it's exactly the same
0:40information.
0:41So what can we really know when we have
0:46such information to make
0:49a educated education that oder educated
0:52decision that this is a transaction I
0:54want really to sign.
0:57A real problem.
1:01Let's go back one year.
1:03This time this was the biggest hack,
1:06but this changes quickly. We had some
1:09bigger hacks now,
1:11but last year in February there was a
1:13hack. It's
1:15a little bit
1:16not everything directly correlates to
1:19the topic we will discuss today, but at
1:21least we see the problem.
1:23So in the Babil hack the problem was
1:25there was a manipulated user interface.
1:29And when designers signed
1:33a very normal transaction moving
1:35something from the cold storage to the
1:37hot storage oder to the hot wallet,
1:41they signed something which was not the
1:43transaction they really
1:46intended to sign.
1:48And the result was
1:51there were 1.5 billion dollars in tokens
1:55lost because there was a transaction
1:58which did not do transaction as
2:01intended, but changed the logic of the
2:03smart contract oder auf der auf der
2:05Multisig.
2:07Now we can say okay, this is one special
2:09problem. There was a very direct attack
2:13to do this in in this application, but
2:17this is a problem we are confronted
2:19everywhere. So we have transactions
2:24and
2:25when we want to be sure what this
2:27transaction is doing, we need somehow to
2:31simulate it. So in the former version of
2:34the
2:35Gnosis Safe you had the possibility to
2:39simulate your transaction, ne? There was
2:41a connection to Tendely. You could
2:44simulate it. It was sent to their
2:45servers. They did the simulation,
2:47returned it to you.
2:49But actually this is also only a shift
2:52of trust, ne? When I ask someone else on
2:55an OBC again a centralized service, I
2:58have to trust that the simulation
3:00results is ja really what I'm intended
3:04to do.
3:05So and it's
3:07actually also a privacy issue because I
3:09ja
3:10reveal data before I do my transaction
3:13to another party. So
3:16I don't say that we cannot trust those
3:19companies, but we also know that even
3:22big companies sometimes are compromised.
3:24Na, even if I trust this company in
3:27person, but if the data is going there
3:29and if it happens that they are
3:31compromised, I would leak my data.
3:36So signing is not enough, ne? Today
3:39mostly we created transaction.
3:42For instance, I'm using a DeFi tool or
3:46exchange. Let's say we are using
3:47Uniswap. I want to change one Ether to
3:51USDC. Then I create a transaction. I see
3:54it on my screen and then in my wallet I
3:57have to sign something and I sign it and
4:00hope that the transaction is correct.
4:03What we need instead of doing it in this
4:06simple way, we need first to create a
4:09transaction of course, then we need to
4:11verify that every information which is
4:14used to create a transaction is really
4:17correct. We have to simulate it and
4:19analyze what are the
4:21results and then I have to explain the
4:23results so that I really understand
4:25these results belong to this transaction
4:28and then I really know what I'm going to
4:30sign and then I can do an educated
4:33decision what I really sign.
4:38So our topic from Colibri
4:41we are building a stateless client and
4:43we say don't trust any information you
4:45get, always verify before you use it.
4:49And now we want to use this this topic
4:53also to
4:55to validate transaction before we send
4:57them.
5:01When I prepare a transaction normally I
5:04need some data, I need some balances, I
5:06need some information from the contract,
5:08some state information.
5:10So if I don't validate this information
5:13I'm even if I have
5:15verified everything else I'm still in a
5:18trust system. So York other colleague
5:20from my from me today mentioned this in
5:23his presentation is that even if we have
5:26an ordered the system with unordered
5:28data it's an unordered system.
5:32So how are we doing this?
5:35So Colibri a stateless client is
5:39a layer in between the RPC and my
5:42application. That means I read data from
5:44whatever source I want, let's say an RPC
5:47or any other source and I also get the
5:51proofs
5:52so that I then I can verify that this
5:55information is correct and we are using
5:56two type of proofs.
5:59So the first is the execution layer
6:00proof, so I proof that let's say a
6:02transaction is part of a block.
6:12a belongs to a this block hash. And then
6:15I know this block contains this
6:17information and now I need also to prove
6:20that this information belongs to the
6:23canonical chain. And here I need the
6:25consensus layer proofs. I need to prove
6:27that the validators, the signatures of
6:29the validators are valid signatures. And
6:32this I can do starting from a checkpoint
6:34in the past, verify all the transitions
6:37of the validator committee up to the
6:39block where I'm now so that I know I
6:42have the correct signature. Or I can it
6:44more efficiently have a zero knowledge
6:46proof which proves this whole chain of
6:51of
6:52of transitions of the validator set. And
6:54with these two proofs I can produce
6:58verified state. So I don't need to trust
7:01the information provider anymore. I just
7:03get data, I get proofs and then I have
7:06verified state.
7:08So and based on this verified state I
7:11can use an EVM. And of course also this
7:14EVM should not be somewhere in the
7:16cloud. I should have it in my
7:18application. So Colibri also brings a
7:21very small EVM in included so that we
7:24now can the verified information,
7:26execute the transaction in the EVM and
7:29then have all the information which are
7:32the results of my transaction.
7:35What calls are done,
7:37what is transferred,
7:39how what are the state changes, which
7:42events are
7:44given and so on.
7:46So we we do it directly in the
7:48application.
7:51But here we have another problem. So
7:52when we list this to the user,
7:55that is the same when I look into a the
7:58report from Tenderly. So by the way, I
8:00really love Tenderly. So as developer
8:03it's a great tool to analyze when I do
8:06an transaction with complex smart
8:09contract interactions
10:10Und hier kommt ein AI Model
11:45So, ja.
11:57We celebrity has been doing some years
12:01before or explain us what what is a
12:03elephant. We are not interested in our
12:06wallet to to get this information. We
12:08just want to have a model which is very
12:11quick, very small and very smart in one
12:14topic. And here we are using domain
12:18small language models. These are small
12:21LLMs which are trained with information
12:25only relating to this topic.
12:28And then we get very
12:31good models which we can use and our
12:34models are
12:36about 2 GB and we also try to make them
12:41even smaller and they give answers as
12:44good as these frontier models because
12:47they are trained with them and can
12:50ja produce this good this good
12:53information.
12:55So what
12:57is our
12:58and what are the limits
13:00we can verify all information we get
13:02from the
13:05from the blockchain. We
13:07we can see if anything in this
13:10transaction is something which might not
13:13fit to our intent
13:15but of course
13:17we cannot solve everything. If the state
13:19is changing before we included there
13:22might be a different things.
13:25Ja and we also need to understand what
13:29my intent is. Now this is also something
13:32we are researching right now not to make
13:35a clear definition to see this is the
13:38intent so that it is easy for the people
13:40to understand that the intent is really
13:43fitted.
13:44So and with this we have a full workflow
13:47from the data making the data trustless
13:50with Colibri the the core module then
13:53simulated in the local EVM analyze it
13:56and extract all the information now for
13:59instance to give the model the right
14:02information we need also the code
14:04the which is really affected from this
14:06call.
14:08And in this point it's not feasible to
14:11give him all the code we find for
14:13instance if you do a Uniswap transaction
14:16the whole Uniswap
15:01Thank you very much.
15:24Thank you very much.
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: a6264ea903553f5f7da5bea2f2a9f5cbc44db74259bb878db9deb64241b58a4c