PropAMMs: the good and the bad, and how to keep trustless infrastructure competitive

Transcript

Hi, I'm Ari. I'm co-founder of FTL Labs, which has developed a lot of DeFi infrastructure, including uh Gyroscope, and I've also done a lot of DeFi research. And I'm joined here today by uh Tammy who's an independent blockchain researcher who previously worked at uh centralized exchanges. And we're going to talk today about prop amm the uh the good and the bad. The landscape of onchain liquidity has been changing with the addition introduction of things called prop amms.

Uh and they have two competing headlines today. uh one of them is about efficiency where they've been bringing lower tra trading spreads to to DeFi and actually out competing passive liquidity and the other is about spoofing uh where they've been uh providing attractive quotes and simulation and then worse execution in the middle of the transaction sometimes as much as 50% uh worse. Now before we dive into these mechanics about how all of that works, let's walk through what is a prop amm. So a propm what it means is proprietary automated market maker. It works through programmable and active liquidity with

an important new liquidity innovation. And that liquidity innovation is the ability to do zero latency pricing uh using information that's already available at the time of execution. And how this uh usually works is that before a swap uh the the operator of the prop amm is able to import new information on chain. On Salana this takes the form of uh being able to do cheap parameter rights that'll land before uh the execution of the swap. And on EVM, there's something known as uh application controlled execution, which will allow you to guarantee that your pricing information update lands at the

front of the block before the execution of swaps later in the block. And you can also import information that just already exists on chain and use that to inform what happens in in the prop amm. And then during the swap execution, the first step is that uh this new type of contract, it reads the inputs that it wants to read. It then decides dynamically what the terms it's going to offer are and kind of updates uh what those terms are. That includes like what prices they're going to offer, what fees, and potentially like screening of who they're going to actually offer these prices to. And only after that is done does it then either fill or decide to revert. And this compares to passive

pools where you really don't have those earlier steps. There's just like a defined uh a defined logic about how it prices and then it either fills or or reverts. It's also different than how uh normal market making works on like say centralized exchanges because in those cases the market maker has to dynamically try to uh reverse its prior quotes before they're filled. And there's no guarantee they're going to be successful. There's no guarantee that they can run something uh ahead of an incoming swap, but they try to.

So what is so interesting about prop amm? So one thing is that programmable liquidity can be used defensively. It can remove stale quotes that the LPS would otherwise be leaking value to arbitrageers. And this means you can have more efficient LPS that can quote tighter spreads and make liquidity more efficient. So that's the efficiency story. But it also gives a superpower that can be abused for spoofing. Uh the prop amm can quote a good price in order to lock in order flow and then once it gets that flow it can execute at a bad price and pocket the difference. And in this talk we want to understand the economics of prop amms both the good and

the bad why these things happen and how they happen what happens in practice. Uh we characterize a what we call a split route extraction which is how a spoofing pool can maximize what it can extract from from traders. And then we talk a little bit about how to mitigate these downsides. We start with a simple model for prop amms that's based on uh what's known as like the gloss glen Mgrim marketmaking model. Uh what you need to know about this model is that it just explains how market makers decide their bid ask spreads in order to cover the losses that they will end up incurring because of better informed traders who know more

information than they do uh while learning about what the assets true value is from incoming trades. And on top of this model uh which is a well-established model on top of that prop amms add uh a gamma parameter which screens against uh arbitrage flow and a new uh choice uh variable m where they can decide post routing uh to do a quote reversal. And at a high level what's happening in this model you don't need to know too much of the details for this talk but you can read more in in the paper when we release it. Uh at a high level you have trades incoming and a proportion alpha of them are informed or arbitrage

flow basically and the remaining proportion is uninformed or mainly retail flow and then you go through a screening process based on gamma where the prop amm can uh avoid stale stale fills on a proportion of the uh the incoming swaps and then the propm decides what spread it's going to offer. Uh, and this spread then covers the residual adverse selection, the chance that they still have an informed trader that that got through. And then after that, they decide a hold up amount M, which is what they're trying to extract by doing a quote reversal basically. And then the all these trade stuff all it happens in

the context of a trade being routed uh of total size like bigger than is going through the the propm and the total the total routing has a tolerance uh slippage tolerance uh tow that and and only a share of it is being routed through to the prop AMM. Um and the profitability of the prop amm in this model then comes down to two main parts. One is how well does it screen against arbitrage flow and then the second is the post routing discretion can create another uh spoofing profit and we get several results uh that help us understand what's happening in prop amm uh the first one is straightforward

that screening lowers the spreads uh that prop amm end up quoting and in ter and that's because it lowers the costs and the chart here uh it shows shows dyn kind of visually what's happening there. Um, and you you have the how the spread behaves on the y- axis at different levels of the gamma parameter uh for avoiding toxic flow on the the x-axis. And the important thing you need to know is just that it's concave and decreasing in of course how well it uh it filters the toxic flow. Perhaps the more consequential result though is its effect on passive pools. So prop AMMs uh because of this can undercut passive pools while earning positive markouts

and because they can quote tighter uh they can attract the uninformed flow actually from the passive pools which makes the passive pools less competitive uh the propm is to characterize what we call split route extraction uh which is how a spoofing pool can maximize what it can extract uh from traders. Uh now let's walk through a particular example of this because it can be a little unintuitive why they can actually extract so much.

So let's consider the example where $10,000 is being is being quoted uh through a routing and this is set with say 1% uh tolerance for slippage and so the trader is guaranteed to get out at least $9,900 and the 98% of this goes through uh an honest pool uh with an honest leg. It quotes $9,800 and fills at about $9,800. And then only 2% goes through an opportunistic pool which quotes $200 but then actually fills that $100 which is a 50% shortfall. Very big.

And then the result of the total routing is that you end the trader ends up with 9,900 which is uh meets their slippage tolerance for the global trade. The global check passes and it's able to settle. Uh and here it's important to note that the problem isn't that the trader is setting bad slippage tolerance. Actually, it's rather that because of this uh new way that the propm can amplify uh it can actually extract the whole slippage tolerance of the whole route even though it only gets a small leg.

And uh what's actually happening in the model is that this the spoof quotes uh they they allow the the the quote to change by the factor mid route. and you they can actually maximize this this particular equation. Although it's worth noting that in practice prop AMMs don't know all of the parameters but they can guess pretty well and they do guess pretty well in practice. The result is also similar to what happens in sandwich attacks but it's much more direct in this case. You don't need to be the uh the searcher uh arranging transactions and like front running and back running.

It's just directly the pool operator quoting as the counterparty. And the third main result helps us understand the different types of prop amms based on how profitable they are after doing split route extraction. And there are two main types that uh that kind of show up. One is a pool that's purely extracting by spoofing quotes. They're basically being a loss leader uh and they quote at a net negative margin. And the other is a prop amm that actually does competitive pricing um but with a positive and and that has a positive base margin. And uh the chart here then lets us differentiate like what is happening in these cases. Um and

it shows the shape of profitability on the the y-axis and uh based on the amount of the total trade that's being routed through the prop amm on the uh the x- axis. And what's happening is that there's a kink here. And this is because the holdup revenue of the split route extraction uh saturates when the to the total global slippage tolerance is taken. And then after that for the uh the spoofer, the pure spoofer, they make less after that because they can't really take out anymore after the global slippage tolerance is taken and they lose uh on the rest. But for the other class of uh of prop amms, they're still providing competitive quotes after that and it actually has to settle at that

competitive quote and they can make money but they make more money from spoofing. >> Hello. >> So I guess every everyone now is the expert on the model. So we now comes to the excited part. So how um how where's the note? how the glo how the aggregator actually enforce a global check or in reality. So to do that we did two exercise. One we conducted a control experiment against sex aggregators on Ethereum and on Salana. So under different trade setting we want to understand whether the aggregator

execute or revert swap if one pool in a split trade delivers substantially less than its quote. And then we analyze some real life transaction on base and on Salana to see how does the opportunities pool capture values by exhausting global slippage tolerance on split lag in real life. for the control text. So we created a test trait um um and then we executed it into two different lags. One is a honest lag um it executed a price that is the same as its quote and then another one

is the mock lag. It executed at a price which is worse than it. We run the test trace through 32 uh different trade scenario. Um it is constructing from um this uh the list of control variable we show on this slide. So we have a different global slipage torrent setting um different input shares for mock legs and then four different target output shortfall. We repeated this task for eight aggregator router configuration.

So five chosen from Ethereum and then one chosen from Salana. So the result is that the tested aggregator pass allow the MO light to consume the global slippage allowance as soon as the total output met the minimum. And I think this is the um we have some very interesting finding. We obtain this observation while we doing our analysis on the split trade on base and on salana. Um we find out in general there are two common method to spoof aggregator. The general idea are very similar is to quote a price and execute

it at a worse price. So they can they uses dynamic fee pool to set lower fee at quote and higher fee at execution or they use pools that can uh use oracle to update price. So in order to so they can better uh execute sorry they can call a better price and then execute at a worse price for trader. We also detected that there are two methods to detect value extraction opportunity. Um the key is to be able to differentiate aggregator quote simulation and execution.

So um one approach is to check whether the transaction uh to check the transaction origin. some code simulation uses the zero address, the now address um as the transaction origin. So the rate provider can check that address and return one rate during the code simulation and then later return a different rate um when the real transaction arrive with the nonzero transaction or origin. And then there's um another approach of use gas price as a signal. um normally zero not no sometime no gas fee for quote

simulation um and higher gas fee for real execution transaction. We find out between January and October in 2026 we have seen um value extraction opportunity is driven by DAX volume especially for prop amm there's a high number of DAX hooks like unis swap v4 balancer auto vtier to spoof aggregator and then we have um observed that the split trait extraction happen more often on base than on Salana.

Um I would like to show you two split tray extraction that I found on base. So this is a trade happened in August 2026. Trader bought CBBTC with over 100k USDC. um base settler routed the trace through seven swap lags and we have detected the swap lag use uh via unis swap v4 hook executed the CBBDC at over 85,000 USDC per CBBTC which is 809 basis point above all the other uh combin the value weighted average price of the other combined l in

the same trade and at the end the trade is executed at a overpriced um overall price at 26.3 basis point above the Binance spot price at the same minute. So um while other lacks um the is the the combined value weighted um average price is only 1.9 basis point above the Binance spot price. By inspecting the trade data, we have noticed that this unis swap v4 hook enforce a 7.5% high fee at execution.

So um there's no it is almost quite impossible for the aggregator who have accepted such high fee at the first place. And this is another example I found on base in January. So this is a trade of buying um BNKR with USDC. The USDC to rep uh swap was one of the intermediary routes. So over 12,000 of USDC was used to buy the red ease. The aggregator split its trade through four pools and we have seen that detected that the tessera executed the their portion at 23,000 USDC per which is over 77,000 basis point above the combined value

weighted average price of the other. So these two transaction are good example for the split trade um execution mechanism we described earlier in the model. So uh we we've seen there's a bunch of issues um and there are actually ways to address these issues. Um, for the most part, uh, I'll go through them briefly, but we can talk more about it in in the Q&A after. But one very easy way to address the the first issue is to make sure that routers are actually enforcing

per leg caps and not just global slippage tolerance because this then makes it impossible to do the sort of amplification that uh, prop amms are doing today. You can also do other things like audits and probation basically of like specific pools. But this is actually quite hard to to rely on solely because it's really easy to launch a new contract and change parameters about old contracts. You can also do more execution sort of checks uh either in simulation or on chain. There's another issue I want to mention just very briefly, which is that uh external LPs, if they're allowed in prop ammms, uh they're actually delegating the pricing risk to the operator. It's not really like the true non-custodial

sort of uh sort of area. Um and these can go wrong. Uh we haven't seen that many like prop amms yet. is kind of new, so there's not like tons of examples, but we did have one in a gyroscope pool where the operator was uh was was using these pools to spoof, but then abused their powers uh to rug protocol fees out of the the pool balances. And there's various things you can do toward uh making more trust minimized infrastructure, which we can also talk about in the in the Q&A. And now to take stock, we've we've walked through a model for understanding the these sort of prop ammms and what's going on with both efficiency and spoofing. We've characterized the maximal split route extraction that it

happens in practice and discussed a little bit at least about how to how it can be mitigated. Uh and the overall idea is that I mean there's a lot of potential benefits to prop amms. uh they can revolutionize liquidity and they are but they have they have a lot of tradeoffs and they can easily spoof things with the current infrastructure. Uh if you want to hear more about this follow for updates uh on on Twitter our Twitters are here and the paper and code will be released soon and otherwise thank you and uh we have a little bit of time for for some Q&A.

>> Amazing. We do have a few minutes for questions so we can proceed. Um the first question that's coming up is if trustless infrastructure can still be manipulated through the data it receives, how should protocols protect themselves against that? >> Yeah, that is a great question and which we didn't have enough time to cover everything. But one very basic thing is just making sure that uh you're properly checking the execution. What uh make sure that the that you as a protocol are checking the execution as much as you can. One easy thing to do is just make sure that each uh each leg of a of a routing that you give it an

allowance and if it's exceeded because the propm is trying to extract too much then you revert and that's better for you as a trader if you revert. It's a little annoying uh but hopefully over time the UX could become easier. Perfect. Um, our next question is, how will Glam Amsterdam's gas repricing changes? Um, for example, EIP8037 and EIP 8038 affect the operating costs of prop AMMs on Ethereum. >> That's a good question, one that we haven't thought about too much uh yet.

Um, it will really come down to any changes will come down to uh uh searcher revenue basically. uh because how how the prop ammms work on Ethereum is that they they work through coordinating with a searcher to make sure that any pricing updates happen earlier in the block. That's for one type of propm though where you're actually trying to like proactively update prices uh like in real time. You also have other types of prop ammms that maybe try to do that occasionally but otherwise just try to draw whatever information they can on chain. So every unis swap v4 hook is primarily of this type. Um and in that case like the cost of that doesn't really change I suppose.

>> Great. Um our next question is is the goal of running prop amms to make maximum profit by spoofing? >> Um it's not necessary because like there's also good propm that runs off for like a much tighter price and then most effect um and then have a better execution quality on chain. So I think in some scenario there will be if there's opportunity for some prop am in question if there's opportunity to extract value from civ trade they will go for it if there's no control at the aggregator level um but we also see it's

not only the prop amm is trying to do that but also pools with the ability to change the feier as we discussed in the uh talk earlier. Um also try to extract the value um from the spit trade. >> To add a little bit too, one thing we can all be doing when we're designing our protocols is to try to design the game for this so that it's not profitable to be spoofing and that involves having to to do different things in how we're how we're doing routing. >> Great. Thank you. Um I think we have time for two more questions and the final two questions on the screen. Um, so our next one is what is the biggest unpriced risk or hidden friction point in this model technology that most

people are overlooking? >> Uh, so I mean most people are actually aware of spoofing at this point. They've heard it a lot. Um, I think as far as I know like what we walked through today about the like maximal way of spoofing is the first I've like seen that directly characterized. It's good to know uh and also like lets you decide how to mitigate it. But yeah, most people have heard the headline that like prop amms are spoofing. But maybe what they don't think as much about is to the extent that you can actually be an LP in propm which you you can't be in all in some cases they strictly gate it to proprietary capital but in some they

don't. Um, but you should be realizing I suppose that like uh the operator has a lot of arbitrary power over the pool and it's not quite the same non-custodial setup as like lping in a standard unis swap pool. Um, actually the the operator could be malicious or even just like go down for a while and you could end up losing money because uh the strategy doesn't work very well. these operators are running highly sophisticated strategies and you have to trust that they're going to to keep doing that. That's unless we like also try to make kind of more guard railed infrastructure for prop amms too which in principle can be done but doesn't really exist right now.

>> All right. And our final question for today is should prop AMMs or data aggregators publish data on quotota price versus executed price periodically as a transparency measure. >> This is actually a very interesting idea. But with the current public data on chain, I think it's actually quite good for use as the purpose to detect any price manipulation in this like u um split trade scenario. Maybe they can be have a better tooling to trying to detect like uh events like this. So people will know um to understand which particular pool is been

doing the monia stuff.