DSS 2026 / Day 1 / Pearl Hall
Why Blockchain Security Has Failed: Single Points of Failure
Kirk Baird
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:00Okay. So, why has blockchain security
0:04failed? If those of you were in Peter
0:06Tschirhart's ski's talk this morning,
0:08we can all agree blockchain security is
0:10currently failing. We're not doing
0:11enough. There are way too many hacks.
0:13There are billion dollar hacks. There
0:14are 100 million dollar hacks every week.
0:17And it's it's becoming a serious problem
0:19for our industry that we need to fix.
0:21And so the point of this talk is to
0:25explain to you why this problem is
0:28occurring and how they are fixing it.
0:30And some of the inspiration for my talks
0:33has come from what banks actually do in
0:35practice, which is a lot of risk
0:37management. I know we all kind of hate
0:39that word and it sounds gross, but it
0:40has a lot of practical effects and we
0:43can definitely learn from it.
0:44So, the problem is arising because
0:47defenders only need to break one
0:49component of your system. They get to
0:51see every single component. They can map
0:53it all out. They find the weakest point,
0:55which is often a single point of
0:56failure, and then they break that one
0:58single point and they have your
0:59protocol. So, the job is a lot harder
1:03for the defenders than it is for the
1:04attackers because we have to secure
1:06every single component, whereas
1:08attackers get it easy and bug bounty
1:09hunters, if you're a bug bounty hunter,
1:11get it easy. You just need to find one
1:13hole and you can get your way in.
1:16So, what are these holes?
1:18The most obvious hole and the biggest
1:20hole that pops up is a single point of
1:22failure. So, a single point of failure
1:24is anywhere in this whole system. Uh,
1:27I'm I'm talking abstract here. This
1:29doesn't need to be a literal piece of
1:30code. This could be one trusted signer.
1:33This could be um your smart contracts.
1:36This could be your choice of blockchain
1:38if you pick a centralized or one with
1:40weak security. It could be any part of
1:43your smart contract. It could be the
1:44oracles. Like, are you getting your
1:46price feeds? So, you don't need to think
1:48of this as like, oh, this code function
1:50is a single point of failure. You need
1:52to think abstractly. You need to work
1:54out what is your system doing and then
1:58what can one of these things do.
2:00This what I want you all doing. I want
2:02you all mapping out the whole system. I
2:03want you all there like, this connects
2:05this, connects to this. This is the
2:07entire abstract system, and once we map
2:10it through, threat model, risk analysis,
2:13work out what the problems are.
2:14And when we find a single point of
2:17failure or a high risk component,
2:19there's three things we can do. The
2:22absolute best we can do is diversify,
2:24redundancies. Think of this as like a
2:26multi-sig. You don't just have one
2:28person signing. You got three. You got
2:29multiple people signing, and you got to
2:31hit two or more of them to break the
2:34system.
2:35Then we can do bound. So bound is
2:38limiting it with something like a rate
2:40limit or something along those lines
2:42that you can slow the flow. You can you
2:44can lower the impact. Maybe rather than
2:47having all your funds in one wallet, you
2:49split it into a couple, so okay, if they
2:51break this one, they now only get $1
2:54billion rather than $6 billion.
2:57Um, and then the last one is harden,
2:58which is add everything else you can
3:00just to make it as hard as you can for
3:02them. Like cause the attacker as much
3:04pain as you can do.
3:06Okay.
3:07So we're going to run through a few
3:09different parts of the stack.
3:11This one, when I wrote this was
3:13absolutely the hot topic because
3:14everyone was talking about how AI is
3:16breaking cryptography, how AI is the
3:19most powerful thing in mathematics.
3:21Um, and one of my fondest topics that
3:23I've spent some time on is ECDSA is can
3:26we use
3:28AI to actually break the algebra
3:30underlying elliptic curves? Right now
3:32the answer is no, but if we could, this
3:35is a single point of failure. This is
3:37something that would give you access to
3:39pretty much every single wallet on
3:40Ethereum. It would also break most of
3:43the internet. It would break TLS. It
3:44would break a lot of other things.
3:46Um, but it is a single point of failure.
3:49Uh, and there's different ways to break
3:50it. And you don't actually just have to
3:52break
3:53the cryptography itself. You can do dumb
3:56things like using a very weak private
3:58key generation. And if you only make a
4:0132-bit rather than a 250 bit
4:04private key, then a hacker can brute
4:07force it. 32 bits is about 4 billion and
4:09a computer can do that in a couple of
4:11minutes.
4:13So,
4:15what you really need to do is diversify.
4:19Don't have one single protocol you're
4:22relying on. Don't rely just on the
4:24elliptic curve discrete logarithm
4:25problem. Don't rely just on ECDSA. If
4:28you can change them up and have two
4:30protocols in parallel,
4:32and notice there's an and there, not an
4:34or, you have to make sure both of them
4:36are secure.
4:37There's a few protocols out there like
4:39near that support SECP curves, they
4:42support ED Ed curves, and they support a
4:45few other hashing mechanisms, but
4:47they're put together in ors. You can
4:49have private keys on each of them, which
4:51actually increases the single points of
4:53failures because if you break Ed, it
4:56happens to have a flaw in it, that's one
4:58path in. If you break SECP, that's a
5:00different path in. And either of these
5:02paths can actually
5:04act as a full break. So, we need the
5:06and, not the or.
5:09But I do understand that no one actually
5:12is implementing this right now, although
5:14I want them all to, but if you're doing
5:16these on chain, it's pretty gas
5:17expensive. Cryptography is one of the
5:19more gas expensive parts. And so, if you
5:22can't diversify, if you really can't,
5:25and I mean you like you think about it
5:26and you cannot, then the option is to
5:29harden, which is do as much as you can
5:31to make sure you're securing the single
5:33point of failure. Put everything you can
5:34in there.
5:36Okay, consensus clients.
5:39Consensus clients is one that I think
5:41Ethereum actually kind of does well.
5:43I've put this as as a pro to see what
5:46can happen when we're doing things
5:47right.
5:49If you have under 1/3 of validators, if
5:52one of the consensus clients is buggy,
5:55it has bad
5:57fork choice rules, it doesn't do balance
5:59checks on user transactions, or whatever
6:01it happens, the other 2/3 plus of the
6:04network will reject it.
6:06They're going to say, no, you're wrong,
6:08ignore it.
6:09Uh if you have one of your consensus
6:11client that gets into that one-third to
6:13two-thirds,
6:15the network still continues on, uh but
6:18what happens is you just lose finality.
6:20And then really bad case is is what
6:23happens if one client is over two-thirds
6:26of the network, then if we have a bug uh
6:28and that bug is exploited on chain by
6:30some malicious users, we end up
6:32finalizing and we're in a world that
6:34hurt all the exchanges around the world
6:35are like finalized. Yeah, that's good
6:37block. We'll start paying out the money.
6:39And then it it's very hard to unroll it
6:42when you get to that point. We we can't
6:43just do the DAO hack where we just go uh
6:45send the money back to us. We want it
6:47all. No.
6:49And this this did happen. Someone was
6:51attacking Nethermind, which only had 8%
6:54of the network at the time. And okay,
6:56the Nethermind nodes went down and that
6:58was a bit annoying for some people, but
7:00the rest of the Ethereum consensus
7:02managed con uh the Ethereum execution
7:04and consensus managed to keep ticking
7:06on. We still have finality and the
7:07network was safe.
7:09Now, look look I I I am
7:13kind of drumming up how good this is
7:15because if those of you who do know the
7:17stats uh know that Geth had at the
7:20similar time almost 80% of the network.
7:23So if we did have a bug in the Geth
7:25execution client, this would cause fine
7:28um finality bugs and this would have
7:30been bad for everyone.
7:33But yeah, that this is this is what the
7:34defense is. We we just need to tidy it
7:36up a little bit more. Make sure that
7:37each of the declines gets down below
7:39that one-third threshold and then we
7:41have this perfect situation where none
7:43of the clients are acting as a single
7:45point of failure.
7:46Um that being said, they they all rely
7:49on the same hash functions. They all
7:50rely on the same elliptic curve
7:52algorithms. So there's plenty of other
7:53single points of failures in there if
7:55you want to get into them, but the core
7:57node and the software itself uh is not.
8:00It it it'd be a spec problem at this
8:02stage rather than an implementation
8:03problem.
8:06Cool. It also allows us to do other fun
8:09things as well. So, there's benefits of
8:10having multiple clients and multiple
8:12protocols, and that's things like
8:14differential fuzzings. So, you get other
8:16security benefits that are auxiliary.
8:19Um and you can also do adversarial spec
8:21tests
8:22where um
8:25if every every past bug from any of
8:28these clients can be fed into all of the
8:30clients, and that way you you get their
8:33compounding security benefits.
8:36Okay, next up.
8:38This is a pretty interesting one. So, on
8:42uh Ethereum or Solidity specifically,
8:45any of the EVM chains, we will see that
8:47there is essentially two compilers. Um
8:51but as a smart
8:53uh smart contract protocol, if you're
8:55writing a Viper, you only get a choice
8:56of one compile. This becomes a very
8:59interesting single point of failure
9:01because if there are any bugs in your
9:02compiler, then there's absolutely
9:04nothing you can do uh and your protocol
9:07is going to be vulnerable to those bugs.
9:10Now, uh
9:12this did happen, and this caused a $70
9:15million hack. It happened because of the
9:18non-reentrant check was being optimized
9:22out by the Viper compiler, and so that
9:24meant anyone who is relying on this
9:27non-reentrant check is therefore
9:28vulnerable to the bug. Uh there were a
9:30bunch of other protocols, not just
9:32Curve, who are vulnerable to it, but
9:33Curve is by far and away the biggest uh
9:36protocol using Viper as a language, and
9:38that's why Curve got hacked by this.
9:41Um so, how do we fix it?
9:43Again, uh this is a tough one because
9:46people don't normally have multiple
9:48compilers. So, it makes a diversifier
9:50problem a lot challenging uh because
9:53Viper doesn't necessarily run through
9:55the Solidity compiler. It's got its own
9:56compiler. So, this is where we need to
9:58get into the harden and the bounds as
10:00much as we can. So, pin and track your
10:03compiler versions. Make sure you have
10:05pinned uh your version control. Uh then
10:08testing the actual deployed bytecode is
10:11way to check up on your compilers. So if
10:13what they had done is compiled the
10:15code uh put it into testnet and then run
10:18their tests on the testnet they would
10:19have found this bug. Um
10:21but they didn't because everyone just
10:23assumed that the compiler was valid and
10:25they hadn't done uh thorough enough risk
10:27management.
10:29Uh
10:30yeah bound. Don't put everything under
10:32one version. Uh some potential ways to
10:35get around this diversified problem
10:37would be to have multiple versions of
10:39your code uh run it under multiple
10:41different versions of the compiler and
10:43see if the behavior is different across
10:45these.
10:46And so it's it's not quite as powerful
10:48as diversify but it is testing different
10:50things.
10:53Okay admins admin keys and upgrading oh
10:57a little bit of a touchy topic at the
10:58moment I think.
11:00I don't know if you guys been following
11:01the news. No you'll you'll have So we
11:03know where this is leading but
11:06what happens uh
11:07as an upgrade key? So smart contracts we
11:12often claim that these are decentralized
11:15they're neutral but if all of these come
11:18under one proxy key
11:20uh that is if you have any proxy in your
11:24smart contract and there is a key that
11:25can update the proxy that is a single
11:27point of failure. They can take complete
11:29control over the smart contract all of
11:31the funds in it and everything.
11:33And it is one of the biggest single
11:35points of failure that happens in pretty
11:37much all of the protocols that we review
11:39because everyone wants to have this
11:42upgradeability function.
11:43Uh Ethereum deliberately didn't add
11:47upgradeability because they wanted it to
11:49be immutable. They wanted the code out
11:51there to not be able to change. Uh but
11:53that's not really that practical for a
11:56long-term commercial engagement where
11:58you build your product over the time you
12:00want to add more features.
12:02And so how do we actually deal with
12:04these admin keys?
12:07Well we could uh
12:11just just be hacked.
12:14That's one way to do it. But no, it is
12:16it's a single point of failure.
12:18And you guys may not know this one, but
12:21they
12:26Yeah. Now they they didn't when they
12:28upgraded, they didn't upgrade properly
12:31securing the initialization function. So
12:32someone hit the initialization function
12:34with a new trusted route which was their
12:37own and therefore they had control over
12:39the entire protocol.
12:40Single point of failure.
12:43This one again is a bound. If you want
12:46to have the capability to upgrade, it's
12:48it's really doesn't make sense to have
12:50diversifiers redundant. So
12:54these are all things that we recommend
12:55in every review and I think every
12:57security engineer in the in this room is
12:59going to nod along and they're not going
13:00to be surprised, but things like
13:02timeouts,
13:03staged rollouts. Make sure when you do
13:06the upgrade process as well that you are
13:08actually doing like a mainnet fork and
13:10you are testing the code on the forked
13:12code.
13:13It's amazing how many people will come
13:14to us with migration scripts and we're
13:16like
13:17yeah, but but do you have actual tests
13:19to run this on the mainnet fork or are
13:21you just kind of hoping it works after
13:22you've written it? And they're like oh
13:24no no no it'll work. And then of course
13:27it it doesn't work.
13:30But that's what we can do.
13:31So keys, signing, and bridges.
13:35So a five of nine threshold is great.
13:38This is exactly what I've been talking
13:39about. Diversify. Don't have a single
13:41point of failure. Hey, if if you want to
13:43get the keys to our bridge, you need to
13:45have five keys out of nine.
13:48And you have to have many signers.
13:51But the threshold only actually exists
13:54if every key is separate, is a separate
13:58single point.
13:59It does not count if you put five of the
14:02nine keys on a single device that is
14:05held by one company because then it
14:07becomes a single point of failure again.
14:09And so this is what I really want to get
14:11across to you that you have to think
14:13abstractly. You can't just look at the
14:14protocol code and be great, this is a
14:16multi-sig. It's got five of nine. You
14:18actually have to analyze the
14:19architecture, see what they're doing,
14:21not just in the code, but the broader
14:23system, run some risk analysis and be
14:25like, hey,
14:26no, you've you've actually got five of
14:28your keys on the same device. You've
14:29made a single point of failure again.
14:30That's not great.
14:33Bybit is also another very interesting
14:35example of this. They had a single point
14:38of failure that is a little bit
14:40different, too. They had a multi-sig.
14:41They were doing everything right. They
14:43were like, okay, we have this multi-sig.
14:45We all have to sit down. We have to
14:47review the code and we have to make sure
14:48it's good before we transfer anything
14:50from our cold wallet. But what they
14:52didn't consider as a single point of
14:53failure was the front end of the safe
14:56software itself. So it showed all of
14:59their signers when all of them sat
15:00there, they all looked at the same
15:02front-end UI and it said, this is a
15:04valid transaction doing exactly what you
15:06want it to do. But what the front end
15:08was doing is switching the back-end
15:10transaction so it was different to the
15:12one that was being displayed. All of the
15:14signers didn't check their device that
15:16that matches the front end. And so when
15:18they hit sign, the 1.6 1.4 billion
15:21dollars went went over to the hackers.
15:25And even more recently, what happened
15:27again? So diving a little bit into this
15:30Binance one, it was a single point of
15:32failure. They didn't have just a single
15:34private key sitting on machine. No, they
15:37had some MCP set up in the background
15:40where they had a bunch of different
15:41servers.
15:42And what the single point of failure
15:45here was that it was one path to approve
15:49this MCP to sign everything.
15:51So one of the DevOps had their laptops
15:54breached and then they went in
15:57the attacker is using that proposed a
15:59malicious transaction to the MCP and the
16:02MCP went, great,
16:03this is good. This is a genuine real
16:05transaction.
16:07And so that interface there was the
16:09single point of failure because it is
16:11one point that allows the attacker to do
16:13whatever they want. And unfortunately,
16:15the pains of EIP 7702 mean that you can
16:18exploit bugs like this extremely
16:20quickly.
16:22Uh so one transaction allowed them to
16:24drain the entire contract.
16:26So what do we do? In in some of these
16:29cases they'd they'd already diversified.
16:30They did have independent signers, but
16:32have independent stacks, have
16:34independent ways of reviewing this code.
16:36Uh add second paths as well. Uh what you
16:39need to do is add limits, delays,
16:42monitorings. So once the transaction is
16:45proposed, have someone else come in
16:46there and prove it as well. And harden
16:49your wallets, harden your
16:50infrastructure.
16:51Everything on those lines.
16:53Okay. Lastly, we are running out of time
16:55here. So I'm going to breeze through
16:57this one.
16:58But oracles are absolutely another
17:00single point of failure. Even if you are
17:03using an oracle like Chainlink, that
17:05doesn't mean which has a distributed set
17:07of nodes, that doesn't mean that it is
17:09not a single point of failure. If
17:11someone can adjust the config, an admin
17:13can come in and switch out your oracle
17:15to something malicious.
17:17These things have happened. Uh oracles
17:19been manipulated, configs have been
17:21changed, hacks have occurred. Um
17:24oracles all need to use a median, not a
17:26mean. And infrastructure needs to make
17:28sure that we have multiple front ends so
17:31we don't end up like Bybit. And bound
17:34and harden them too.
17:36So what I want to leave you with uh is
17:39what is a single point of failure? Do
17:42not get too caught up in the specifics
17:44of the code. Keep it abstract, think at
17:46a high level, brainstorm, understand
17:48your system. And when you understand the
17:49system, you can tell where the risks
17:52are. And then you can build a threat
17:54model and then you can understand how to
17:56solve these problems.
18:00Beautiful.
18:02And considering the state of our
18:04industry, look, just make yourself as
18:06hard to hack as possible and they're
18:07going to go somewhere else because
18:09they're easier targets.
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: 6a4f3cfa609c79a9277f0aa137f170b9f2e3b1ab94d1011b71ec0fedbb87ea79