DSS 2026 / Day 1 / Amber Hall
The State of DeFi Security - 2026
Peter Kacherginsky
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:00All right.
0:01Welcome, everyone. Namaste, India. It's
0:04so good to have every single one of you.
0:07Welcome to the state of DeFi security.
0:09It's an honor to give the talk now
0:11fourth year in a row where we look back
0:13at the industry and learn.
0:18So in case we haven't met, my name is
0:19Peter Kacherginsky. I'm the editor of
0:21the Blockchain Threat Intelligence
0:23newsletter for the past eight years,
0:25and also a security advisor to some of
0:27the projects out there.
0:30So before we begin, a quick
0:31announcement. In light of everything
0:33that was happening in the past two days,
0:36I decided to put the Blockchain Threat
0:38Intelligence newsletter out out there as
0:41a public good, freely accessible, so you
0:43can subscribe, learn, and hopefully get
0:46the lessons to not get hacked.
0:51So let's begin our story. Usually when
0:53we gather here, we jump into the
0:55statistics and the threat models and all
0:58the case studies, but I want to do
1:00something different today. I want to
1:02tell you a story.
1:04It's a Western.
1:06It has a nasty two gangs operating
1:09there. There are lots of robbed banks,
1:12and some of you are in it as well.
1:22Maybe the slides.
1:25There we go.
1:27So to get to get an idea of how we got
1:29to these pretty tough times, we have to
1:31start at the beginning, the drought.
1:34Our story begins in the winter of
1:362025-2026,
1:38just as we all coming down from the
1:40hangover of the bull market of the past
1:42year.
1:43Bitcoin just hit 60k,
1:45and all of the projects out there are
1:47starting to really tighten their belts.
1:49The budgets are getting cut, especially
1:51security budgets. Audit jobs and
1:54contests and other security work that
1:56used to be a plenty in the past year
1:58really going down.
2:00Suddenly all of the security researchers
2:02out there who used to be really busy and
2:04could make a decent living find
2:06themselves unemployed without a job on a
2:08curb with no future prospects.
2:13Some of the smarter of you switched over
2:15to bounty hunting and that worked for a
2:17little while until
2:18bounty programs, just like the rest of
2:20the world, were overwhelmed by the AI
2:23slop and started introducing
2:25various types of friction, especially
2:28the submission fees.
2:31And just to give you a context of the
2:33world, both inside crypto and outside
2:35crypto, what was happening with the AI
2:38craze invading everyone,
2:40managers and founders realized their
2:43ultimate goal of having employees who
2:46don't sleep, don't eat, don't require
2:49any asks and just work, work, work all
2:52day.
2:53So AI is going to write the code, test
2:56the code, audit the code
2:59and that's how the layoffs begin all
3:03around the world.
3:06And just around that same time when the
3:09whole AI seems to be taking over the
3:11world, the frontier models introduce
3:13excellent models like Opus 46 and GPT
3:1753.
3:18Where a year ago these models could not
3:19even tell a difference between
3:21re-entrancy and price oracles, suddenly
3:24you could get ready-made foundry POCs
3:26that you could just run on the main net
3:28and hack whatever you want to.
3:32And it is in this context where you have
3:34a whole bunch of security researchers
3:36who are well skilled
3:38and now very well equipped out of the
3:41job
3:42in a down market,
3:43what do they do?
3:46Well, a lot of them they left the market
3:48altogether, they moved on to other jobs,
3:50but some of them they burned the whole
3:53town down.
3:54We went from three or four hacks a week
3:57to three or four a day.
4:00The last time we met here in in
4:02Argentina, we had just 24 hacks a month.
4:06By April, as we called bloody April,
4:09we hit 67.
4:11That's more than double.
4:14And how is everyone getting hacked?
4:16Well, guess what? AI is really good at
4:19finding bugs like signature
4:21verification, which wasn't even
4:23registering in the past years.
4:25Seeming inability by projects to
4:28properly verify the message integrity.
4:30They don't know how to check for
4:31nonsense, what should be in it, what
4:33shouldn't be, what can be replaced. And
4:35then we have devastating consensus
4:37failure bugs. Those are the type of bugs
4:39where it attackers figure out a way to
4:41print infinite amount of money from
4:43bridges and chains. And we had that
4:45particular bug class in the past, maybe
4:47one a year. This year we had 10.
4:51And even more were responsibly disclosed
4:53and patched before they were getting
4:55exploited.
4:56Such as the Zcash earlier this year and
4:59the Ripple infinite minting bug just a
5:01month ago.
5:02And of course, the dreaded price oracle
5:04manipulation. We've worked for years to
5:07bring down the risk of this very well
5:09understood and known attack class, but
5:11driven by all these tiny little old
5:13hacks from projects on mostly Binance
5:15Smart Chain,
5:16this shot right back up to the top.
5:20And who is getting hacked? So again,
5:22looking back in November of last year,
5:24those are mostly projects in the sweet
5:26spot between 100k and a million dollars,
5:28making it worthwhile for attackers to
5:30spend time understanding the protocol,
5:32weaponizing their bugs, and so on.
5:35By April and September and it continues,
5:39most of the hacks were under 100k.
5:42Writing exploits, looking for bugs,
5:44weaponizing them,
5:46and executing them became so cheap that
5:49whatever little project out there had
5:51with some TVL is now a worthwhile target
5:54for attackers.
5:58What was unusual
5:59back in April is that it's been a few
6:02months since we heard from our North
6:05friends.
6:06Were they taking a vacation?
6:08Nah. They were just busy casing their
6:11next targets.
6:12Their first target was the Drift
6:13protocol, which unprecedented was
6:16targeted first in person at a conference
6:19just like this one, where they were
6:20approached by someone to build trust and
6:23then later on build a relationship to
6:24send malware. So be very careful who you
6:27interact with at DSS and DEFCON in the
6:29coming days.
6:30And then of course, we have a classic
6:32trader-trader tactic of compromising a
6:35an engineer laptop at layer zero to
6:37eventually deploy a pretty sophisticated
6:39exploit against layer zero
6:44server nodes.
6:47We'll come back to these guys more.
6:49But what I want to bring up is that
6:50around April time
6:53Frontier Models, Frontier Labs realized
6:55the security sells.
6:57And how are they going to monetize it?
7:00Well, through scarcity and guard rails.
7:03To protect us all from ourselves, the
7:05latest models like Mythos were now a
7:08gated product where you have to be on a
7:10special list and you have to know
7:11somebody in Anthropic and then you have
7:13to sign up for an extra expensive API
7:15plan to get access to a highly
7:17token-hungry
7:18model.
7:20For the rest of us, we were forced into
7:22this guarded garden to make sure that
7:25God forbid we do something security
7:27related.
7:29What was interesting on top of that as
7:30we learned from the Hugging Face
7:32incident is that those guard rails not
7:35only apply to
7:37potentially offensive
7:40security work, but also a completely
7:42legitimate work. When Hugging Face
7:44engineers tried to use the very same
7:46model that was attacking them to defend
7:48themselves, they were told, nah, you're
7:50doing something suspicious.
7:52Lesson learned.
7:57By July of this year, both the guard
7:59rails and also just going through all of
8:01the low-hanging fruit contracts that
8:03were out there was having some effect on
8:05the number of hacks. So we briefly
8:07dipped to around 51 hacks.
8:12But then what happened? The Chinese
8:14models arrived. Models like GLM 52 and
8:18DeepSeek V4 and all the Quens 36s and
8:2138s and Next and so on
8:23suddenly had very similar, not the same,
8:27but good enough cyber capabilities to
8:30bring all the attackers back into the
8:32game.
8:35Most importantly,
8:37the models did not have guard rails. And
8:39whatever guard rails were put in there
8:41were quickly removed through
8:43obliteration and uncensored
8:45versions of those same models.
8:47Even more interesting, a lot of the
8:48security researchers started publishing
8:50specially post-train models like
8:52Cantina's Apex that are specifically
8:54designed for doing security work.
8:58In the end result,
9:01by September of this year, we not only
9:03matched the numbers of Bloody April, we
9:05surpassed it.
9:07If you want to connect the dotted lines,
9:09unless something changes in the near
9:11future of how we do security,
9:13we should see around 95 plus hacks by
9:17March of next year.
9:21What's even more interesting
9:23are the type of hacks that are started
9:25happening.
9:26We started with the low hanging fruits,
9:28all those price oracles and just
9:30contracts don't know how to implement
9:31burn functions over and over again.
9:33There's the same bot that I keep
9:34watching. There's just
9:36pillaging the whole Binance chain.
9:39Once those were done, attackers started
9:41moving down the stack to more and more
9:43complicated targets. They started with
9:45bridges. They continued to whole chains.
9:49We had the Liquid Network hack, $320
9:51million.
9:52They later moved on to wallets where it
9:55seems
9:56every day we hear about a weak PRNG or
9:58some other cryptographic issue of how
10:00wallets are getting generating their
10:01seed.
10:03And they continue going down this phrase
10:04to down this the stack until they're
10:07hitting signing infrastructure and so
10:09on.
10:12So I think folks, as we gather here
10:15today, it's time for a reckoning of all
10:18the things that broke
10:20this year.
10:23On the one hand,
10:24we have the $8.7 billion number
10:29that was stolen so far this year.
10:32And the vast majority of that is North
10:35Koreans just pillaging our ecosystem,
10:39exploiting weak operational security,
10:43security programs that cannot withstand
10:45a single laptop compromise.
10:49On the other hand, we have 574 hacks
10:52this year so far. That's almost double
10:54to the same time period last year.
10:57And that's I call them the the swarm of
11:00AI hackers who are just exploiting
11:03anything that was not audited, did not
11:05have the right code coverage,
11:07and more interestingly, did not have the
11:09right incentives for security
11:10researchers to do the right thing before
11:12them.
11:14And when I say incentives,
11:16this is what bug hunting looks like
11:18today.
11:20A good number of you switched over to
11:23the dark side
11:25where it is now more preferred action to
11:28negotiate the bounty after the hack,
11:31where you name your own
11:34bounty percent, like 10%, 20%, whatever,
11:37arbitrary,
11:38instead of doing the right thing and
11:39falling the programs, paying the
11:41submission fees, maybe waiting a year,
11:44negotiating back and forth, and then
11:46eventually maybe getting something or
11:48more likely just downgraded to
11:49informational.
11:53And how do we get there?
11:55Well, between bounty programs shutting
11:57down, uh platforms rather shutting down
11:59altogether like Code Arena, rest in
12:01peace,
12:02to
12:03remaining bounty programs simply
12:06becoming hostile to ethical security
12:08researchers.
12:10If you are a researcher today, you're
12:12doing this mental gymnastics where
12:15you're thinking, well,
12:17I could submit this critical that I
12:18discovered,
12:20but do I want to risk the submission fee
12:22and do I want to invest the next year
12:24dealing with all this back and forth,
12:27becoming a lawyer more so than a
12:28security professional?
12:31And I think a lot of them are just
12:32choosing not to submit.
12:34And I get it.
12:36A year ago,
12:38you you were overwhelmed
12:40by the AI slop. There were too many
12:42submissions, and I personally know so
12:44many of you triagers out there who just
12:46burnt out.
12:47But that was a year ago.
12:49You had a year
12:51to learn and use the same tools that are
12:53overwhelming you to improve the triage
12:56process because that is your only job to
12:58broker between ethical security
13:00researchers and vulnerable projects that
13:03also need that intelligence, and you're
13:05failing at this mission.
13:09We're experiencing a brain drain.
13:12The rooms I'm seeing are smaller and
13:14smaller. I'm seeing senior and junior
13:17talent leaving the industry for good.
13:20On the senior side, we heard Sam Zizane
13:22announcing that he's stepping back from
13:23Seal
13:25like Open Zeppelin getting acquired by
13:27S&P, the S&P 500 people. The trail of
13:31bids, the rumors they're negotiating
13:32with OpenAI, and I personally know so
13:35many of you
13:36who left the industry altogether to join
13:39the Anthropic and OpenAI's of this of of
13:41the other industries.
13:44And then on the junior side,
13:47all those little contests, all those
13:48little audit jobs and and bounties, that
13:53was our school. That was the pipeline
13:55that would bring the new junior talent
13:58into the industry to become the next
14:00senior security researchers.
14:03And that pipeline is now drained.
14:07As the market will turn around as it
14:09always does, who will audit all those
14:12contracts?
14:13AI. AI is going to write the test the
14:17code, audit the code. Do I need to spell
14:19it out exactly how this will end?
14:23Just ask Balancer that
14:26had the last day and closed the closed
14:28their doors for for good today.
14:30Ask dozens of other DeFi projects that
14:33already shut down this year.
14:38And then on top of it all, we have North
14:40Korea to deal with, who continue
14:42operating unimpeded.
14:45The government is not going to help you.
14:47They are a nuclear state. Law
14:49enforcement is not going to help you.
14:53They don't have jurisdiction and every
14:55sanction imaginable is already in there.
14:59In the meantime, they're setting new
15:01record year after year after year.
15:04This year alone, we just had that I had
15:06to update this slide two days ago. They
15:08just set a new record with $6.3 billion
15:10of a Binance hack.
15:13And they will continue doing so unless
15:15something changes.
15:19Worse, we as an industry are not doing
15:22nearly enough to stop them.
15:25On the one hand, we have something like
15:27Arbitrum Console who drew the line in
15:29the sand and said, we are decentralized,
15:32but that does not mean that you can
15:34abuse our chain and you can use it to
15:36transfer stolen funds.
15:39And on the other hand, we have projects
15:40like Thorchain who are like monks
15:43repeating this mantra of credible
15:45neutrality right up to the point they
15:47themselves get hacked and then they will
15:49shut down the chain.
15:51Or we have someone like Circle who are
15:53hiding behind court orders which take
15:55weeks to obtain when transactions take
15:57seconds or minutes to go through through
16:00their
16:01infrastructure.
16:07Defenders on the other hand,
16:09defense is so difficult. It's a
16:11political game even more so than
16:13technical, and they're still stuck
16:15operating at human speed just as we
16:18already transitioned into the machine
16:20world. They still don't have access to
16:22the latest models, and whatever models
16:23they have, they have guard rails all
16:25around them.
16:28The last but not least, I hope you all
16:30realize that the crypto cryptocurrency
16:33security is kind of a canary in a coal
16:35mine.
16:37What we see happening in our industry
16:40usually predates by six months or years
16:43what happens in traditional security.
16:45So when we had our bloody April, it was
16:48only the matter of time until the Black
16:49October was going to come, where
16:51everything is getting hacked, from DNS
16:53servers to certificate authorities, you
16:55name it.
16:57And before we pat ourselves on the back
16:58saying that we're so far ahead and we
17:00had a time to prepare for it, realize
17:03that all of those decentralized
17:05projects, they actually sit on very much
17:07centralized infrastructure. And now all
17:10of that Black October mess is coming
17:12right back to us. Did you notice the
17:14sudden increase in front-end
17:16compromises? Yeah.
17:19That's what's happening.
17:20So if you are a chain operator or
17:22bridge or or CeFi or DeFi and you have
17:25some kind of back-end infrastructure,
17:27pay attention to that, please.
17:32So just to give you an idea of what
17:33trajectory we are in right now,
17:35it doesn't look so good.
17:37Uh I had the first item written before
17:39the Binance hack where I'm waiting for
17:41DPRK's favorite time of the year to
17:44break things, which is winter holidays
17:46when people are distracted. So that
17:48already happened. That's that's a that's
17:50a thing.
17:52Um we're going to see multiple 100
17:54million plus chain chain killer events.
17:5895 plus compromises a month,
18:01overwhelming. I have to track every
18:03single one of them every week, so it's
18:05it's brutal.
18:07DeFi projects will more and more so will
18:10be compromised through their centralized
18:12dependencies.
18:14And the last but the one that worries me
18:16so much is the unholy birth of gray
18:19market bounty brokers that already
18:21exists in traditional security,
18:24where vulnerabilities are not reported
18:26to the projects, but in a reported to a
18:28gray market where they're sold to the
18:30highest bidder who are not necessarily
18:33the the projects themselves.
18:36We don't want to see that here.
18:41So is there hope for redemption to all
18:43of this?
18:45I think so.
18:47All the things that I'm going to share
18:48with you, they already exist. They've
18:49been tried this year and the past year
18:51and they work. We just have to choose to
18:54adopt them.
18:58On AI and defense,
19:02if you are a defender and your idea for
19:04what you're going to do this year and
19:06next year is to just throw AI at what
19:08you're already doing, you will lose.
19:11You will fall behind the race.
19:14Instead, I want you to step back and
19:17realize that as a defender, you are the
19:19ones that are setting the terrain for
19:22the attackers and AI and their bots or
19:24whatever to operate it.
19:27So why can't you design a terrain
19:30that is so difficult to operate in, so
19:32difficult to navigate that it drains
19:35even the most well-funded attackers? How
19:38can you burn their tokens to the ground?
19:42And before we even start talking about
19:44AI, I want to reintroduce some basics.
19:48How can you lock down what you already
19:51have? How can you set hard rules with
19:54invariance, setting hard privileges,
19:57allow lists?
19:59If you can't lock it down, cap it.
20:02There is no reason, I've been saying
20:03that for years, for your vault to be
20:05drained in a single transaction.
20:09Set the outflow limits, time locks,
20:11multisigs. Last but not least, add a
20:15human element, which is the most
20:17difficult part for automation to bypass.
20:21Cheapen your mistakes. Realize that you
20:23will mess up and that's okay.
20:25So what can you do
20:27to make your mistakes not existential.
20:31Set your guardian accounts that will
20:32save you at the last moment.
20:34Set automatic circuit breakers if you
20:36are if you are an exchange and one of
20:39your chains was already drained maybe a
20:41good idea to automatically shut down
20:43everything else or drain it.
20:47Last but not least
20:49with the Black October especially
20:51just operate under the assumption that
20:53you are already hacked. Your laptop is
20:56hacked.
20:57Your developer is already hacked.
21:00So how can we build a resilient security
21:03program that can survive not one or two
21:06but more compromises of your machines,
21:08of your keys
21:10and still survive.
21:13Then and only then can we start talking
21:15about AI. And I'm not talking about
21:18saying
21:19Claude hack me do not do not fail.
21:22That's not going to work.
21:24Build a threat model. No one else
21:26understand
21:27the weak points and the most critical
21:29parts of your protocol and your project
21:32than you.
21:33And once you identify those sensitive
21:36points then apply AI to it.
21:39Run models 24/7 against it. When a new
21:42model gets released update your harness,
21:45run it again. Do that before someone
21:48else does it for you.
21:52And speaking of models
21:55a reminder from the hugging face
21:56interface you cannot rely on commercial
21:59models for security work. Even the most
22:02expensive and permissive models that you
22:04can obtain today will still have guard
22:07rails. Worse those guard rails may be
22:10applied to you at a moment's notice just
22:12at the time when you most need it in the
22:14middle of the incident.
22:16You must practice self-reliance. Either
22:18run your own infrastructure or rent a
22:21redundant one. Practice with it, build
22:23the harnesses so you know exactly what
22:25do to analyze the locks, to reverse
22:27engineer contracts.
22:29Have that infrastructure on the ready,
22:31so when the worst hits,
22:34you will be ready for it.
22:39And of course, what do we do about North
22:41Korea?
22:42Well, for that, I invite you to look
22:45back in history at traditional finance
22:48that we so like to make fun of.
22:50The thing is, exactly 10 years ago, they
22:53were in the exact same spot we are right
22:56now.
22:57In 2016, North Korea attempted to steal
23:00$1 billion from the Bank of Bangladesh.
23:03How many of you are familiar with that
23:05incident?
23:06Right?
23:08The hack was unsuccessful by sheer luck.
23:12They made a typo. They made a typo in
23:14the Swift transfer form. Someone in the
23:17Federal Reserve Bank of New York caught
23:18it. Great.
23:20There were a few more attempts following
23:22the hack.
23:23So what happened next?
23:25The Swift network sat down all of the
23:28member banks
23:29and instituted a stringent security
23:32controls that all member banks must
23:35apply.
23:36Secure enclaves,
23:3824/7 hotline and incident response,
23:42shared threat intelligence.
23:45And what happened within 4 years? There
23:47were a few more attempts, but by 2020,
23:50North Korea had no more successful hits
23:53against traditional banking
23:54infrastructure where they actually stole
23:56some assets.
23:59And by then, North Korea moved on to a
24:01much
24:02easier target, which is us.
24:06So can we have our very own Swift
24:09moment?
24:11Can we finally decide to self-regulate
24:15before the regulation is forced upon us?
24:19If you saw the FBI notice this morning
24:23telling various projects,
24:25such as Store Chain, such as node
24:28to do something about funds.
24:31That was your first warning, that
24:33something is coming.
24:35And I have a feeling that this industry
24:36and people in this room much prefer to
24:39self-regulate instead of someone telling
24:40them what to do.
24:42So can we adopt an industry-wide
24:45security controls? Think seal
24:48certificates,
24:50but with teeth, with incentives.
24:53Apply security controls, secure
24:56yourself,
24:57and enable us to invest in you, to
24:59integrate with you, to provide liquidity
25:01and insurance, or stay away in your own
25:04walled garden.
25:07Can we agree on the security first
25:10neutrality?
25:12I get it, we want to be decentralized,
25:15but that is not an excuse for letting
25:17stolen funds and abuse of your platform.
25:21All you have to do is to publish
25:24an intervention policy, define the exact
25:27times and boundaries when you will apply
25:29it,
25:30be accountable to it and set realistic
25:32timelines measured in minutes to hours,
25:35not days or weeks.
25:38Last but not least, shared intelligence
25:40and coordination is critical.
25:43Again, I'll keep on using seal because
25:45this is this is our best hope right now.
25:48We have seal 911 and seal intelligence,
25:50but these are perpetually underfunded,
25:53overwhelmed projects. How can we support
25:56this and more
25:58in our ecosystem?
26:02We must bring security researchers back.
26:06And it's not hard. It's not some magic
26:08trickery. We just need to align the
26:10incentives and make it worth their time,
26:13not just financially, but even morally,
26:15not treat them like they're doing
26:16something criminal by telling us of the
26:18vulnerabilities we're trying to fix.
26:24Last but not least,
26:26to all of you security researchers,
26:28defenders,
26:30bounty hunters,
26:31who keep again and again and again that
26:34AI is going to replace us.
26:37It is not.
26:40AI is an excellent tool, but that's all
26:43it is, a tool in the hands of security
26:45researchers.
26:47Even if the whole world is trying to
26:49tell you otherwise, and even sometimes
26:51you yourselves, it is not a replacement
26:53for human creativity,
26:55for the grit and desire to learn and
26:59explore.
27:01And if you need any evidence of that,
27:03just look at the talks at this very
27:05conference written by the creative
27:09minds
27:10that are exploring this world together.
27:17Folks,
27:19it's always darkest before the dawn.
27:23And it was a very dark year, the darkest
27:26I've seen in the eight years I've been
27:27following the space.
27:29But I am optimistic
27:32that we will bring the light
27:34to our industry once more.
27:39We have the tools,
27:42the technology, and most importantly,
27:44the people. They're all already here. We
27:46just need to choose to use them,
27:49choose to win.
27:51We can choose to bring security
27:52researchers back.
27:54We can choose to use the same offensive
27:57tech as the best defensive tools we've
28:00ever had.
28:02We can choose to unite and protect this
28:05ecosystem together instead of protecting
28:08our little kingdoms.
28:11So let's light this fire together
28:14because this time around the power to do
28:17so is in our hands.
28:20Thank 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: b8aa4734d8650bed1bc690314b5f31f5527a59cb96b4911ec95145aa9d512661