DSS 2026 / Day 1 / Pearl Hall
Every Safe Is Blind to Its Own Backdoor: Scanning 2M+ Multisigs for Shadows
Elliot Friedman
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:00Hello everyone. I am Elliot Friedman,
0:03the founder of Solidity Labs.
0:05We're an on-chain engineering team where
0:07we embed with protocol teams to build
0:10and secure smart contracts.
0:13I'm also a seal member and help with the
0:15framework around custody. And today,
0:19we're going to be talking about scanning
0:2312 million multisigs for shadow owners
0:26and modules because Gnosis Safes are the
0:30uh
0:31largest and most widely adopted uh
0:34wallet with the the largest uh AUM.
0:38And so let's dive in.
0:41We're going to start here asking a
0:42question.
0:43What's the difference between these two
0:47Gnosis Safes?
0:49So on the left we can see one has no
0:51pending transactions. On the right we
0:53can see there's some pending
0:54transactions that have been
0:56removed.
0:58Does anyone notice a difference?
1:03So if you said there is no difference,
1:05um you got the question right, but one
1:08of those Safes has a shadow module
1:11that can steal all of its funds at a
1:13moment's notice and the other was
1:15completely fine.
1:18So we have a paradox here.
1:20How can an owner both be and not be at
1:23the same time?
1:25And we have the same question for a
1:27module.
1:28If we made
1:30two calls to a Safe, we could get
1:33contradicting answers
1:35because
1:37every major interface relies on linked
1:39list enumeration
1:42and
1:44if an item is not in the linked list, it
1:46means it's not um
1:49it's not an owner or module.
1:51But as we can see here with this fun
1:53little diagram,
1:55all uh owners or modules in Safe must be
1:59in a
2:00linked list
2:01where we start off at the sentinel. The
2:03sentinel points to the first address,
2:05the first address points to the second
2:07address and so on. And then the last
2:09address will point back to the sentinel
2:12address.
2:13So it's a linked list, it goes in a
2:14circle.
2:16But as we can see here, if someone is
2:19able to write to this mapping outside of
2:22this circularly linked list,
2:24now this shadow.eth will be a
2:28owner.
2:30And it will not be returned in the get
2:33owners function and so no interfaces
2:35will be able to see it.
2:36We have the same issue with modules
2:38which are
2:39arguably more powerful than owners
2:42because they do not require quorum
2:45in order to
2:46actually have a transaction go through.
2:48And so this brings us to the core
2:50problem with shadow owners and shadow
2:52modules, which is that they are
2:55invisible almost everywhere across the
2:58ecosystem
3:00because
3:02no one bothered with the tooling to
3:04detect them.
3:05Um
3:07and there are some websites about this,
3:08but
3:09yes, basically what happens here is we
3:11have a second owner that exists on
3:14chain. It's invisible to the safe UI
3:17and
3:18there is no uh infrastructure checks for
3:23for shadows.
3:26Um
3:27but we're going to get into uh a site
3:29that we've built out that helps solve
3:31this problem.
3:33So here we have a
3:36gnosis safe that has a shadow owner.
3:38We can see here in the owners array it
3:40has the owner, the sentinel that points
3:42to the owner that points to the
3:43sentinel. So it's a nice circularly
3:45linked list. We go down to check
3:47addresses, we paste the address in here
3:50and then it says, you know, this is a
3:52shadow owner.
3:54This address is a shadow owner of this
3:56safe.
3:59And
4:02how does it get planted? And why are we
4:04investigating this today?
4:06Because this is a known issue that Safe
4:08can have these shadow owners and shadow
4:10modules.
4:12We were building a Safe-based wallet
4:15and noticed that during the setup
4:17delegate call window, an attacker could
4:19inject arbitrary
4:21uh actions because
4:23the module setup function
4:26calls does a delegate call out to the
4:28target address that's passed.
4:32And so
4:34we said, well, wait a second. If you can
4:36do an arbitrary delegate call during
4:38initialization, then an attacker could
4:41quite easily
4:43get people to set up Safes using some
4:45malicious way
4:47and then install a shadow module
4:50or a series of shadow modules
4:52in that Safe.
4:55[snorts]
4:56And right, you just don't don't write
4:57into the
4:59into the circularly linked list, but
5:00just add into the mapping. This is now
5:03an owner.
5:04And we wanted to know, is this
5:06theoretical or has this attack already
5:08been planted in the wild?
5:11Um because if it's theoretical, then
5:14hey, you know,
5:16it doesn't really matter. But we haven't
5:18really seen anyone do a large-scale scan
5:21of all
5:23all the Safes that contain most of the
5:25funds that Safes hold across multiple
5:27chains and ecosystems. And so we wanted
5:29to do that.
5:31Before we get there though, let's talk
5:32about the CPIMP uh
5:34attack pattern. CPIMP stands for
5:37clandestine proxy in the middle of
5:39proxy.
5:40And this was a hidden proxy inserted
5:43uh during proxy initialization
5:46and it made the proxy be owned by an
5:49attacker.
5:51And
5:52the same thing is kind of true with
5:55shadow owners and shadow modules because
5:57you get created by this hidden entry in
5:59the owner's mapping.
6:00And both of these attacks, they leave
6:03surface logic and getters untouched, so
6:06the smart contracts work exactly as
6:08intended.
6:09But both of these things are invisible
6:12to auditors, Etherscan, Tenderly,
6:16and basically, you know, the in
6:18infrastructure that people rely on
6:20um
6:21on a on a regular basis to ensure their
6:25their safes are are secure.
6:28And
6:28as a as a category, as a
6:31crypto really we we under
6:34underweight um
6:37under index
6:39threat models that are persistent
6:42where an attacker gains access and then
6:44doesn't use it immediately because we're
6:46so used to
6:48instant smash and grab exploits and same
6:50block chains
6:52uh in DeFi, but
6:54I think that this category is
6:57very very concerning and we should be
6:59spending more time on on finding
7:03um and hunting for threat actors that
7:05are,
7:06you know, exploiting these kind of
7:08methods.
7:10So can we even prove that a shadow
7:13exists?
7:14If we have the address of a shadow,
7:17shadow owner, shadow module,
7:19we can prove that
7:22um
7:24we can prove if it's a shadow or not. We
7:25can falsify that question. But if we are
7:28instead
7:30given um
7:34if we are instead given a
7:37arbitrary storage right inside of a safe
7:40that doesn't do anything that we think
7:42it should be doing or we don't know what
7:44it should be doing,
7:45uh it's actually impossible for us. Even
7:47if we used and we had all the computing
7:49power on Earth,
7:51we wouldn't be able to guess whether or
7:53not this
7:55safe was a shadow owner or just some
7:57arbitrary storage right.
8:00And that's because you'd have to brute
8:02force all the Ethereum addresses in
8:03existence and then check them
8:05to to if they're either a
8:07or
8:08an owner in the safe.
8:14So how do we detect how did we go about
8:16detecting shadows? So
8:19we started off by reading all the public
8:21state that we could from the safe
8:24and
8:25with that state we reconstructed the
8:27expected storage try.
8:29We fetched the current storage hash with
8:33eth get proof
8:34and then we checked to see if the
8:38current storage hash from the RBC
8:41matched the expected
8:44um that we computed ourselves.
8:46And a mismatch is going to be imply that
8:49there's storage that that we don't
8:52uh that we can't account for which means
8:54yeah we're probably dealing with a
8:56shadow.
8:57So we started off with a script that
8:59didn't have much state and we quickly
9:02realized that
9:03oh actually um
9:06we're going to need quite a bit more of
9:07a complex system for this.
9:10So we started off with
9:12um we had a lot of false positives
9:14because of approve hashes and signed
9:16messages
9:17because the safe could write these
9:20into storage and everything would
9:25check out um in terms of this being a
9:27legitimate storage right but we had no
9:29way to account for this. So we had to
9:32build indexing into our service and
9:34crawl through the logs
9:35to find all of the approve hashes and
9:38and sign messages.
9:40And then we also found another after we
9:43we fixed that um set of false positives
9:45we found another set of uh false
9:47positive which were pass keys
9:50which basically we had these account
9:52extraction wallets on base
9:54where we configured um
9:59web off um
10:01pass keys basically or the owners of the
10:04safe configure these web off pass keys
10:07and uh it just made these three
10:10deterministic storage rights to these
10:12different addresses and then there was
10:13different modules and different versions
10:15of the pass keys that would
10:17write to different addresses and so we
10:18had our different slots and so we had to
10:20go and update our um
10:24our eliminator service
10:26and reviewer service to understand this
10:29behavior so it could um
10:32clear all of these these false
10:33positives.
10:35So this is our funnel on Base. We had
10:3860,000
10:39um storage root mismatches flagged once
10:43we had the logic to you know get rid of
10:47the approve hash
10:49and pass keys
10:51um
10:53that that ended up eliminating uh
10:55you know most of
10:57of the of the safes that we found and
10:59flagged and then we were we ended up
11:02with
11:0312 unexpected uh
11:06safes
11:08with with potential shadow
11:11uh owners or modules
11:13but these were on you know addresses
11:16that didn't have enough funds
11:18um
11:19to to warrant concern and most of them
11:23were were unused anyway.
11:25So these are our results.
11:27We ended up scanning 12.1
11:30million safes across Base, Ethereum,
11:34Optimism, and Arbitrum
11:36and
11:38you know the majority of them ended up
11:40reconciling against the expected clean
11:43root which is great
11:45great news. We don't need to have a seal
11:46war room.
11:48Yeah it we we didn't really find
11:51um evidence of a large-scale shadow
11:54owner or shadow module
11:56um
11:57you know campaign.
12:00And here we have just a UI, a couple
12:04screen grabs here. You can see the
12:06different networks.
12:08In the future we might add more.
12:10It'll show you how many safes we've gone
12:12through,
12:13what what block we've indexed to, the
12:16different versions of the safe that we
12:17covered.
12:19[snorts]
12:20Here we have an unexplained storage
12:21mismatch. We have this route that we
12:23expected. This is the actual route.
12:26And there's a little drop down
12:27explaining, you know, what this could
12:29be.
12:31Here we have uh
12:33a safe that cleared. So hey, we can see,
12:36you know, we have our linked list here.
12:38The sentinel goes to, you know, these
12:40two modules and then it closes and
12:43there's one owner.
12:44And so everything kind of
12:46matched up there.
12:48Here's another one that's clean.
12:51Um
12:52again, that's great news, you know, this
12:54is how the majority of them came back.
12:57Here we have a
12:59shadow safe where we have an owner.
13:02Goes to the, you know, sentinel to owner
13:04to sentinel
13:06uh modules. It says there's no modules,
13:08but we checked this address and it said,
13:10hey, this is a shadow.
13:14So here we can see the
13:17shadow scan URL.
13:19And you're able to
13:22um look through all of the safes of
13:25version 1.
13:2641 and 1.50
13:29uh on this site. You can get a real-time
13:31view of scan progress and flag safe.
13:35And you can also see, you know, a
13:37recomputed
13:39uh storage route locally that gets
13:40checked against RPC. So you don't need
13:43to just trust the back end. You you
13:45actually get that uh in browser.
13:48So we're going to end with uh
13:50a little bit of a tip here. If you see a
13:54safe that has been created
13:56that you're using and it has an
13:58unverified delegate call,
14:01uh if you see that, you should treat
14:03that as a potential uh full compromise
14:06[snorts] and immediately contact Seal
14:09911 and if funds are in that safe that
14:13has a potential compromise or a
14:14potential shadow owner or module, um
14:17immediately migrate those funds to a new
14:20safe.
14:23So yes, stay safe, uh stay in touch. Um
14:27yeah, this was this was a uh multi-month
14:30effort to get all of this
14:32together and so I want to just thank the
14:35other engineer on the project, Pratik,
14:36who really
14:37uh
14:38stepped in and helped out a lot. So and
14:41you can follow us more um on Twitter at
14:43Solidity Labs
14:45and you can check the site out here.
14:48Thanks so 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: 046e44d927488eab302f4213a7d279778d4de10b48e920467653bfe1d9b877bb