DSS 2026 / Day 1 / Pearl Hall
How to Model Cryptoeconomic Attacks and Adversaries in DeFi
Abhimanyu Nag
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:02Hey everyone.
0:03Full house. Nice to see that.
0:06Uh
0:07So, I'm Abhimanyu. Thank you so much for
0:10coming to my talk. I'm talking about
0:12cryptoeconomic attacks and adversaries
0:14in DeFi.
0:16This is going to be a more foundational
0:18talk than some of the others.
0:21This isn't, you know, solely, strictly
0:24based on the MEV
0:26literature that you all have seen or are
0:28used to. And this is coming more from
0:30the econ CS side from anything else.
0:32I'm also my tax fellow at the University
0:34of Alberta.
0:35Would love to talk about it later.
0:39So,
0:40the main primitive that we're going to
0:42be talking about here is going to be
0:46auctions and how DeFi auctions have been
0:49set up and how you can manipulate them.
0:53There are a lot of examples of auctions
0:55that you've seen in DeFi till now.
0:57There's liquidation auctions,
1:00trading auctions, which I mean as
1:01intent-based auctions like cow swap, so
1:03on and so forth. And there's also block
1:05space auctions, which is
1:07the normal
1:09block space
1:11allocation and payments that are
1:13required by block producers.
1:15So, fundamentally, an auction specifies
1:18who can bid,
1:20who receives what, and who pays. And the
1:23main adversarial attack that I'm going
1:25to be looking at for auctions today is
1:26going to be sibyls where how can one
1:30economic actor or a bidder or an agent
1:32in a system is able to manipulate an
1:34auction by spawning
1:37multiple other identities.
1:39So,
1:41fundamentally, an auction has an
1:43allocation rule and a payment rule.
1:45And as as a very nice finite example, we
1:49have two bidders, Alice and Bob.
1:52So, Alice values a good at VA.
1:55And so she bids BA
1:58to be able to get a chance
2:00gain that good. Similarly, you have Bob
2:03who values that same good at VB bids VB
2:06I guess. And the auction
2:09specifies that exact function
2:12which would basically be a probability
2:14of
2:16whether or not Alice or Bob will win the
2:18auction for that good and how much
2:20they'll have to pay for it.
2:22There's a very simple example of a
2:23single item auction which is going to be
2:25specifically what I deal with today
2:28which is going to be the second price
2:29auction. So there's bids which are 10 7
2:33and 4. I bid for a single good at 10 7
2:36and 4. You see that the highest bidder
2:39who bids 10 they end up paying the
2:40second highest bid which is seven.
2:44There are different DeFi auctions as I
2:46talked about them. I'll just reiterate
2:48the months. You've seen them for Maker
2:50DAO which are collateral auctions which
2:52is the significant example of a Dutch
2:55auction where the collateral price
2:57declines with time and purchases they
2:59settle instantly. You have intent based
3:01auctions on cow protocol where solvers
3:03compete to propose settlements and then
3:06on the more fundamental blocks place
3:08level you see flash bots right and their
3:10MV boost.
3:11There are some adversarial questions
3:13that will probably detract us from what
3:15we I want to talk about today and I'm I
3:17have limited time apparently so I'll go
3:19over something more specific.
3:22So an attack can exploit either code or
3:24economics. That's the way I see it. Code
3:27attacks are
3:28all of you are probably way better at
3:30that than I am.
3:31And so today I'm going to be
3:32specifically looking at economic
3:34attacks.
3:35Some of the examples of economic attacks
3:37right now are as I said sibyls
3:40which is one person spawning multiple
3:42agents. At the same time there's
3:43collusion attacks where it's all of the
3:46agents in the system collude with each
3:48other. There's a really subtle
3:48difference with sibyls.
3:50It's basically that these these are
3:51separate identities right.
3:53So the very basic model of an attacker
3:56is that exact action which maximizes
3:59their utility
4:01given the information constraints that
4:02have, and you also minimize the cost
4:06that you need to actually conduct the
4:08attack.
4:11Right now, some specific assumptions
4:14[snorts] that we're going to be making
4:15for auctions is going to be that it's
4:18truthful, where if you remember from our
4:20example, there's VA and BA.
4:24Truthfulness implies that VA is going to
4:25be BA no matter what because that's the
4:27highest utility
4:29for the agent.
4:31At the same time, civil proofness where
4:32as I said, one actor, they cannot gain
4:34more utility by dividing into several
4:37bits.
4:38And the main point I want to drive is
4:39that if an auction is considered safe
4:41for one specific bid, it might not be
4:43this it might not be safe for multiple
4:45bids at the same time.
4:47So complete civil attack, very very
4:49simple example. Think of a lottery,
4:51right?
4:52You have a prize which is say worth 120
4:54and there is a lottery here where your
4:57probability of getting the item is
4:59proportional to the number of players in
5:01the game.
5:02Say
5:03if it's 120, then the probability of you
5:06getting and if sorry, if there are say n
5:08bidders, your probability of winning is
5:10one by n, right? So let's have a
5:13baseline of three identities.
5:15I have an attacker and I have two honest
5:18people, right, who are bidding alongside
5:19me.
5:21I my way of attacking would be to divide
5:24my one bid into two or maybe, you know,
5:26I don't have to divide that one value. I
5:28can have a value greater than that so on
5:29and so forth. The main thing that you
5:31have to think about is that when there's
5:32one attacker with two identities, it's
5:34my probability of winning the good is
5:36one by three, which is $40 in this case.
5:39Similarly, if my once I deviate and I
5:42divide my bid into two, my actual
5:45probability of winning is one half now,
5:47which is $60. So you see that difference
5:50of $20 there, right?
5:52This will give you some more intuition
5:54where you might think that in order to
5:56make the system or the auction
5:58completely secure, you're going to have
5:59to impose a certain cost which has to be
6:02at least greater than $20 in this case.
6:05So now that you think about it, when you
6:07take it across agents, across people,
6:09attack search is an optimization
6:10problem.
6:11And what we are actually going to be
6:13optimizing is that the utility you get
6:15from being a civil agent versus a
6:17truthful agent, you're basically trying
6:19to maximize that exact portion, which is
6:21delta here as you see it on the screen.
6:25There's several ways to calculate it,
6:26which are grades, normal optimization
6:29techniques that you see in Python.
6:31There's also some constraints here with
6:33gas, deposits, timings. You get the
6:35gist.
6:36The main point right now, as of today,
6:39the second price auction is the only
6:41auction that's proven to be civil proof
6:43for now in the single item case. There's
6:45a lot of literature in the multi item
6:47case as well that's been done by the
6:49wonderful people at Cow Protocol, right?
6:51You should definitely check them out.
6:53We have some more results for
6:56civil proof auctions. Maybe I'll talk
6:58about it later once we publish the
6:59paper.
7:01So one really interesting aspect that I
7:02want to talk about is AI agents as
7:05attackers.
7:07So
7:08AI agents are the way I see it on a pure
7:11economic level, they're not going to be
7:12very different than human beings about
7:13thinking,
7:15about how to manipulate a protocol.
7:18They're going to do the same exact thing
7:19that we do. They're going to read the
7:20rules. They're going to form a
7:21hypothesis about how to actually attack
7:23it. They're going to evaluate how much
7:25it's going to cost them to actually
7:27conduct the attack and they're going to
7:29basically revise their constraints from
7:32feedback, market feedback, people
7:33feedback, so on and so forth.
7:37The next thing that I want to talk about
7:39is a very, very simple
7:41infrastructure
7:42for how you as an agent or like you
7:45having an agent, please don't do this.
7:47Your AWS is is going to hate you. Your
7:49bills are going to be huge. I'm not
7:51going to lie.
7:52But the main thing that I want to talk
7:53about is how
7:56what a basic infrastructure might look
7:58like, right?
8:00So I have been an AWS max here all my
8:02life. I'm sorry if I offend someone here
8:04with that, but I would basically take an
8:08S3 container where I'm going to query
8:11information or data from Dune or
8:14on-chain data. Hopefully you have a node
8:15setup of your own.
8:17You're going to take the on-chain data,
8:18you're going to take all the code, the
8:20hash, the seeds,
8:22and you're going to give it to your
8:23agent. Your agent after pulling it is
8:25going to form a hypothesis, right? It's
8:27going to take the information at a
8:29microscopic level. It's going to skim
8:31through it, figure everything out. It's
8:33going to give you a structured point of
8:35attack.
8:37Then in order to solve the deviation, in
8:40order to actually see all the values,
8:42all the bits that it has to do, how much
8:45civil bit do they actually need to
8:47actually manipulate the protocol,
8:48they're going to use
8:50an EC2 worker. That's what I would
8:52presume. That's what I would do
8:53technically, where I would do
8:54computations at
8:56at the millisecond level, I presume.
8:59I I don't think AWS does that. It's
9:01pretty slow.
9:02But on top of that, once those
9:04simulations are complete, you're going
9:05to know the agent at least is going to
9:06know by the end of it exactly what the
9:09deviations are required.
9:10Then there's going to be an independent
9:12replay, which is again like that's big
9:15words for saying back testing where
9:17you're going to take block data, you're
9:19going to see the civil deviations that
9:21would have given you more rewards than a
9:22normal person. Then you can back test
9:25it.
9:26You can see exactly what gain, what
9:29utility you're going to get as an
9:31attacker versus a normal agent. And then
9:33you're going to verify it,
9:35send it back to S3 as logs, right? You
9:38can pull it, you can check it, verify
9:40it, so on.
9:42So I am not going to leave you with just
9:45like you know just that, just a way to
9:47attack the system. I'll also talk about
9:50I'm I just going to want to talk about
9:51how
9:52an sorry, how a protocol designer is
9:55able to figure out a mechanism or learn
9:58a mechanism against an attacker.
10:00For that, I the way I see it, it's going
10:03to be a coupled optimization function.
10:07Big words, but what it actually is is
10:09that you're going to have an inner
10:10problem where you're going to freeze the
10:12mechanism parameters
10:14and you're going to search for Sybil
10:15attacks for any auction. You have an
10:17auction A with payment rule A payment
10:19rule B sorry and you're going to figure
10:21out a way
10:23you're going to put a million Sybil bids
10:26to make as many bidders as possible from
10:28one person and see what your utilities
10:31are in that case
10:35and then you're going to solve the outer
10:36problem where you're going to solve it
10:38for the protocol designer now. The outer
10:40problem is basically I want to increase
10:43I want to maximize my revenue at the
10:45same time minimize
10:46the instances of Sybil's.
10:49What I mean by that is that when you
10:51minimize the inner problem
10:53you're basically figuring out exactly
10:56how
10:57expensive it is for a Sybil attacker to
10:59actually attack the protocol. You're
11:01going to increase that expense sorry
11:03you're going to make that really really
11:04expensive as expensive as you can while
11:06at the same time
11:08improving your revenue maximizing your
11:10revenue at that very instance. What I
11:13mean by that again at the end is that
11:17you're going to end up learning a
11:18mechanism where you know using neural
11:21networks deep learning that's what we
11:22did for our very recent paper
11:25hopefully when when it comes out I'll be
11:27able to talk more about it but what it
11:29actually does is it's able to learn
11:31mechanisms where attackers no matter
11:34what they do do not have any incentive
11:37to attack the auction at least using
11:39Sybil's.
11:41At the end of it I have some final
11:43thoughts I have like three minutes left
11:45as well
11:45that
11:48the full proforma
11:50of actually conducting a Sybil attack or
11:53thinking about attackers in
11:55economic protocols in general is you
11:57have to specify the agents you have to
11:59specify exactly how Sybil bids can be
12:02inputted into the system you have to
12:04compare the utilities between being an
12:05attacker and being a normal agent and
12:08you have to separate when I say evidence
12:10from guarantees what I actually mean is
12:12that separate back-tested variables with
12:15the actual market conditions.
12:17And my final point that I want to drive
12:18with this
12:19small talk, I guess,
12:21not so small now that I think about it,
12:23is that economic security it actually
12:25needs an adversarial model. It doesn't
12:27need like a specific it doesn't need
12:29like a best-case model to be fair.
12:32The worst-case scenarios are always
12:33considered in consensus research. It's
12:36always
12:37considered in even in auction
12:39literature. But I I haven't I'm yet to
12:41see ways in which people are able to
12:44combine those two. And by finally, I
12:47just want to say that I we have figured
12:49out a way to create civil-proof auctions
12:51from the ground up, turn any auction
12:54into a civil-proof one using
12:56wistfulness. So that'd be something I'd
12:59love to talk about.
13:01That's all I had. If you have questions,
13:03please come find me. It was great to
13:04meet you. Thank you 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: 26b0df5c8af51d2468ca988d6d4fe20c26741f58d5a6c572e313c23fc39c1f40