DSS 2026 / Day 2 / Pearl Hall
Onchain Security Markets in a World of Offensive AI
John Ennis
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:01Good.
0:02Tired.
0:04And before I start um on this next one,
0:08I'll press my road to see if my slides
0:09work. Yes, I do.
0:11Um can I get a show of hands here?
0:13Anyone who has a safe multisig just so I
0:15know
0:16kind of the rods. Okay, so we're looking
0:19at
0:19maybe half, maybe half.
0:21Um
0:22if anyone isn't aware of Safe is a
0:24multisig um and
0:27I work I've been working with Safe for
0:28about maybe six years. I originally
0:30joined Gnosis and we when I joined
0:33Gnosis, they were actually originally
0:35working on the
0:36prediction markets and
0:38I think there's still we we you're
0:40probably familiar now with Polymarket.
0:42Um they brought prediction markets to
0:43the main to the mainstream, I would say.
0:46But there's still a bit more innovation
0:47that could be done with regard to
0:48prediction markets, specifically in the
0:50security space, and I want to talk to
0:51you about kind of our insights that
0:53we've gotten from this and kind of our
0:55journey on where we're kind of releasing
0:56Safe Safe Net and uh putting together
0:59some on-chain security markets to tackle
1:01the
1:02the rising threat of offensive AI and
1:04the fact that yeah,
1:06existing threat actors are just more
1:07scalable now with AI than before and
1:11yeah, we need to level up as a result.
1:14This used to be 2.26 million so far in
1:16hacks, but by I mean by by Binance uh
1:19increased this significantly as well.
1:22Um this is so far we're looking at 8
1:24billion worth of losses. Everyone knows
1:26what we're facing here, uh but really
1:28we're looking at deep fakes as as a
1:31rising trend. There is autonomous
1:34agentic exploits happening, constantly
1:35reviewing patterns around your codebase
1:38and also assisted social engineering and
1:40most of all like social engineering is
1:42kind of the number one cause that we're
1:43seeing that is contributing to these
1:45numbers.
1:48Someone had a good talk yesterday around
1:50how you can
1:52kind of triage your agents to look at
1:54different areas of the codebase so they
1:55don't default to one when they're
1:56finding bugs.
1:58While you're doing that and taking those
2:00insights, can also kind of create other
2:02agent roles itself to simply recursively
2:05check what bugs that other agent your
2:08other agents are finding and
2:11chain attacks together, chain exploits
2:13and continuously run. And if you can't
2:15find an exploit in someone's protocol,
2:17be it on the front end, back end,
2:18contracts, or PC level, come back to the
2:21bug finder agents and the agents again
2:24recursively continue to find new bugs
2:26and eventually
2:27you'll get a successful exploit.
2:30Previously this was done by just
2:31security analysts alone, but now with AI
2:34it's just more scalable. You can just
2:36get hundreds of agents working on this.
2:39So the threat actors now just have yeah
2:41more scalable tools than previously. And
2:44what happens as a result of that is
2:47the common vulnerabilities and exposures
2:50available
2:51found in all stack, be it open source or
2:54closed source, just continuously
2:56increases. So we're looking at since
2:58like AI like LLMs kind of began to reach
3:01mainstream adoption, we just see
3:04doubling and doubling doubling in terms
3:05of common vulnerabilities. And
3:08we're already 72,000 in Q3
3:11in. So I'd say we'll hit 100,000 this
3:14year. And again that's doubling from
3:16last year. So as you can see this is
3:18just becoming
3:20you know just
3:22essentially exponentially
3:24increasing. And again it's
3:26agents have made these existing threat
3:29actors more scalable.
3:32So yeah we're kind of in a bad situation
3:34at the moment.
3:35We're basically like trying to kind of
3:37stack up some some defenses quite
3:40quickly.
3:42David had a good talk on that just just
3:44before around
3:46yeah how how they're doing that. But
3:48what since the since LLMs are
3:50essentially using crowd sourced
3:52knowledge
3:54and attack attackers are using AI to
3:56leverage the crowd to attack,
3:58we can leverage the crowd also to
3:59defend. And that's the kind of insight
4:01that we wanted to in into our next step
4:03product that we're building.
4:04So how exactly do we want to apply this?
4:06So like if they're using, you know, the
4:08crowdsourced knowledge, how can we
4:09defend a crowdsourced knowledge?
4:11Um and where we want to to actually
4:14address that is in specifically
4:16transaction security.
4:18Because the biggest thing causing these
4:20hacks and these exploits is that you
4:23don't really know what you're it's hard
4:25to actually match what you're signing
4:27with what you intend to do.
4:29And that's happened across the board for
4:31all the recent hacks that involve
4:33billions or millions and
4:36that is still a very it's I think it's
4:37what I think it is one of the biggest
4:39problems. It's a multifaceted um
4:41security space with a cat and mouse
4:42game, but I think this is one of the
4:43biggest challenges and one of the
4:45biggest problems we're all trying to
4:46solve.
4:47And the kind of scope today or the kind
4:49of gaps we see is
4:51we tell people, oh, verify what you're
4:52trying to sign with a third-party tool
4:54or use our company
4:56um to do this and we'll we'll review all
4:59your transactions and we'll tell you
5:01whether it's good or bad. And that's
5:03fine, but there's gaps in that as well.
5:06And the gaps that we see is that
5:09one, advisory warnings are structurally
5:11useless. So if you if you give someone a
5:13warning to say what you're signing is it
5:17might be malicious,
5:18people can just click sign and do it
5:20anyway. Structurally they can just
5:21bypass it or if your UI is compromised,
5:25that that warning's not coming through.
5:27You are going to essentially you you
5:30have no last line of defense. You
5:31misinterpret the UI, you sign, and
5:34you're you're wrecked.
5:35Um
5:36the second as well is that
5:38again, self-custody is irrevocable. I
5:40think we all know this as well, but
5:42again, the attacks are getting more and
5:43more sophisticated. So it's getting even
5:45more like the cost like the cost of
5:48attack is getting lower and yeah,
5:50obviously the honeypots are getting
5:51bigger.
5:52And
5:54we've we've noticed this as well.
5:55There's no real economic accountability
5:58really for these third-party tools or
6:01third-party companies that are um
6:03checking transactions. Like ultimately
6:05it's the that's using the wallet is
6:07going is the one that gets wrecked if
6:10someone didn't catch
6:12a threat. So if you have a company
6:14that's checking your transactions and
6:16says, oh, we didn't catch that, but
6:18you're the one that's facing the
6:19millions worth of hack millions worth of
6:22dollars worth of hacks, it's that's
6:24challenging, right? Like that's kind of
6:26in a way that's kind of unfair, but
6:28that's kind of how it exists right now.
6:30Of course, there's legal recourse, but
6:31really there's not as much skin in the
6:33game, and we found that ourselves as
6:35well.
6:38So what we have built now with what
6:40we're what we've released now actually
6:42with Safe Ness is what we want to do is
6:45multi-party security checks. So we want
6:47to leverage the crowd and not just
6:48select one single actor that is kind of
6:51doing the security checks for our
6:53wallets.
6:54We want to leverage multiple parties. We
6:56want them to be accountable. We want
6:57them to put up bonds as a result of each
7:00transaction that they want to
7:01participate in.
7:02And we want to enforce it on chain,
7:04which means it's not just a UI warning
7:06that you can just bypass, that it needs
7:08to be a transaction needs to be attested
7:10on chain in order for you to execute it
7:12as well.
7:13And we want to address specific risks
7:16that these multiple parties we call them
7:18sentinels. We have four sentinels right
7:20now. And we want the sentinels to be
7:22addressing specific risks.
7:24Address poisoning, unauthorized approval
7:27to your contracts or to your tokens,
7:29constantly checking malicious contracts
7:31and whether they are compromised at the
7:33time of the transaction,
7:35hidden ownership changes, so changing
7:36the actual ownership underneath without
7:38you realizing. This can be used through
7:40like unsupported delegate calls or or
7:42nasty like hidden delegate calls.
7:45And just generally trying to manipulate
7:48again what you're trying to do with what
7:50they want you to do as well.
7:53And we have this charter, which is
7:55essentially a dynamic rule book, let's
7:57say a court of law. Since there's always
7:59new laws coming through or there's
8:00always new attack patterns, so we'll
8:02always be updating this kind of charter,
8:03this court of law of what to you need to
8:06address or the sentinels need to
8:08address.
8:10So just to walk you through how this
8:12would work,
8:13um
8:14we have
8:16you propose a transaction and we'll walk
8:18through the happy case. We propose a
8:19transaction. There's sentinels. Right
8:21now we have four, but really we want
8:23more sentinels and we actively want
8:25anybody that's operating in this space
8:28to potentially evaluate running a
8:29sentinel.
8:30Um the sentinels check against the
8:32charter. Are there specific risks or
8:35specific rules that this transaction's
8:36breaking?
8:37If if it's unanimously secure, they all
8:40say thumbs up, it's good. The validators
8:42are more of an admin role. They'll
8:44attest the transaction on chain. They do
8:46this thing called a frost signature,
8:48which is just again they all participate
8:50in kind of generate a cryptographic
8:51signature. And once that is attested,
8:54it's broadcasted cross cross chain and
8:56then the guard can verify that yes, this
8:59user can actually execute this
9:00transaction because the validators have
9:02created the frost attestation and the
9:04sentinels say this is secure, we back
9:07this transaction.
9:09Um that's if you enable the guard. You
9:11don't have to enable the guard. You can
9:12just simply use warnings instead, but if
9:15you do enable the guard, you have to go
9:16through a delay and many people in the
9:18previous talks talked about having
9:20delays associate with critical
9:21operations or delays associate with
9:23critical like large-scale transactions
9:25that that's true and a lot of hacks
9:27could have been prevented or not
9:29prevented, let's say, but it could have
9:31added another layer of defense if there
9:33was delays incorporated as well.
9:36For the insecure case, if one sentinel
9:38says this is insecure and this is the
9:40reason based on the charter why this is
9:42insecure, that will be reported to the
9:44user. You will The validators will not
9:46attest and you have to go through delay
9:49if you have the guard enabled.
9:53Right now the validators who are
9:54performing the administrative role, this
9:56we have like Nosis Safe, we have
9:58Greenfield, we have core contributors
10:00running validators, they're generating
10:02frost signatures,
10:03and the sentinels they're putting up
10:05bonds per each transaction
10:07um that they want to participate in and
10:09we have the of Blocksec, OpenCover,
10:11Candide, and TenGard. Again, we want
10:13more validators and we want more
10:15sentinels to make this truly kind of
10:17permissionless system
10:19and build up because it's a permission
10:20system right now, but this is what we're
10:21working with right now. And yeah, happy
10:24to chat to anyone after if you're
10:25interested in looking into this deeper.
10:30Again, this is live right now in Safe
10:32Wallet. You have to subscribe to Safe
10:33Pro, but
10:35users simply, you know,
10:39either pay pay a subscription or they
10:40pay on a per transaction basis.
10:43The sentinels get that transaction, they
10:45get the fee as well. If the sentinels
10:46participate, they get they get a portion
10:48of the fee.
10:49And again, they're they have bonded
10:52verdicts. I don't have enough time, but
10:53if the sentinels disagree and one of the
10:55sentinels is wrong after the fact, they
10:57lose their bond and that bond is
10:59distributed to the other sentinels. So
11:01the game theory works that the sentinels
11:03want to find things that the other
11:06sentinels disagree with,
11:09which is actually hacks like because
11:11like malicious transactions tend to be
11:13the 0.00%
11:16of transactions. So you want the game
11:18theory to work that sentinels find
11:20specific things and get the bonds of the
11:23other sentinels as well and of course
11:26participate so they're getting the fees
11:28on a on a constant basis as well.
11:31Again, the validators do they perform
11:33the frost signature, they create the
11:35attestation in the guard, and then yes,
11:37the user
11:38can execute a transaction if they have
11:41the guard enabled.
11:43And here you can just see right now this
11:45here just creating a safe transaction.
11:47Sorry, I don't know why I'm looking at
11:47this. I know these slides.
11:50You can see the you can see someone
11:52create a safe transaction. The sentinels
11:54found no issue
11:56with that transaction, so you can simply
11:59go ahead. That's the happy case execute.
12:01If the sentinels did find something,
12:03this would be reported. You would have
12:05one sentinel reporting something or
12:06many, and you can click on the explorer
12:09and here the explorer will tell you
12:12who said, who said this is actually
12:15insecure or not secure,
12:17and what's the rationale of why they
12:19said insecure. If So here you can see
12:23block six said secure. Candid and 10
12:25cards denied it and said insecure.
12:27And here this is on all on chain, so it
12:30has to be quite a rigorous explanation.
12:33But we will summarize this in the UI,
12:35but ultimately it fails because you have
12:39a wrapped ETH.
12:41You've created an unlimited approval to
12:43the wrapped ETH contract with that
12:44account. We don't want unlimited
12:46approvals in this day and age with
12:49with specific like, you know, basic user
12:51operations. So yeah, this is labeled
12:53insecure as per the charter.
12:57So why why why do we think this is
12:59different from the existing kind of
13:00state of play right now? It's the fact
13:02that we have multiple parties competing.
13:04Like right now in like the state like we
13:07have multiple attack
13:10like
13:12attack participants or let's say
13:13malicious actors all competing.
13:16Whereas it doesn't really have that
13:17same, you know, game theory for the
13:20people who are trying to defend. It's
13:22all kind of like you're on your own
13:25mentality. So we believe that like that
13:27the we want to leverage the crowd here
13:29and leverage the experts and the best is
13:31to actually have multi-party competition
13:33as a result of this.
13:35Again, no just warnings that you can
13:37bypass on chain enforcement. So if the
13:39UI is hacked, you can still actually,
13:42you know,
13:43have critical transactions blocked on an
13:46on chain level.
13:47We have a dynamic charter. Like the law,
13:49there's always new laws or new attacks
13:51arising. We'll also have We have a
13:52charter as well that's always being
13:54updated with new attacks that are being
13:56created because this is a cat and mouse
13:57game at the end of the day. There's no
13:59one solution. Not This This won't fix
14:01everything, but it will be another layer
14:03of defense.
14:05And again, making making people
14:06accountable actually, making people
14:08accountable for self for not finding
14:10things or finding things and reporting
14:12them wrong.
14:14And yeah, we believe this is in the
14:16direction of creating more
14:17accountability for uh yeah, defending
14:20against these threats.
14:23In the future, we want to create a more
14:24open
14:25sentinel market, open validator market.
14:28Um, and again, we want to make them
14:31specific with transaction security. We
14:33want to support multiple accounts, so
14:34this is not just a safe solution. I'd be
14:36happy to talk to any wallet providers or
14:38any security companies that would love
14:40for their for our charter to be updated
14:42to their um
14:45uh structure. And yeah, we can also
14:47create insurance products on top if the
14:49sentinels can guarantee that they will
14:51check these things as well.
14:54That's me done. Um, I would just say you
14:57can scan this QR to learn more. If you
14:59want to reach out to us directly about
15:02running a wallet, running a sentinel,
15:04running a validator, you can get this QR
15:06code. And also, we're in Mumbai Cricket
15:08Club uh on
15:10Wednesday.
15:11Um, and yeah, we have a we have an event
15:13there with the Ethereum Foundation um if
15:15you'd like to attend. So scan that QR
15:17code if you'd like to attend the event
15:18as well.
15:19And yeah, I think I'm all out of time.
15:22So um yeah,
15:23I don't chat to me if you have any
15:25questions after. I can't I don't think
15:27I'll I don't want to run into anyone
15:28else's time.
15:30Thank 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: 8ccb8156078d08c1ea936ec3bb24b6ed7009bb5ec8e81b94b3963423d834e174