On-chain investigationEthereum · Arbitrum · Base24 October and 10 November 2026

The bug that cost Curve $38,200 drained Balancer's fee burner 17 days earlier. We found no public report of it.

A flaw in how two DEXs sell their collected fees let a stranger sell them and keep the cash. Curve was not the first to pay for it.

At a glance
On 17 November, Curve said an attacker had taken $38,200 of its fees through a flaw in its fee burner, the program that sells collected fees on CoW Swap. Balancer's fee burner on Ethereum had the same flaw, and it was drained first. On 24 October, 26 forged orders in 17 minutes took 0.64 ETH, about $1,512. Each order carried a copy of one real Balancer order as its signature, the trick later used on Curve. We found no statement from Balancer. Of 11,021 burner sales we checked on four chains, 164 paid an attacker, all in these two drains. No transfer links the two attackers.
$1,512
Taken from Balancer's burner on 24 October
17 days
Before the same flaw cost Curve $38,200
26
Forged Balancer orders, filled in 17 minutes
164
Forged sales out of 11,021 we checked
Key findingsEach line has its own share link
  1. 17 days

    Before Curve lost $38,200, the same flaw had already drained Balancer's fee burner on Ethereum, on 24 October 2026.

  2. $1,512

    Taken from Balancer in 26 forged orders over 17 minutes, paid in ETH to a wallet that had never been used.

  3. 1 order

    Copied as the signature on every forged order, on both burners. The burner checked the copy and never the order it was filling.

  4. 164

    Forged sales out of 11,021 Curve and Balancer burner sales on Ethereum, Arbitrum, Base and Optimism. All were in the two drains.

  5. 0

    Transfers that link the two attackers. Balancer's ETH went to an instant exchange, and Curve's USDC went to Hyperliquid.

  6. 4 chains

    Where Balancer lists live burners with the same code that we could not check: Gnosis, Avalanche, HyperEVM and Plasma.

01 · What happenedBalancer's fee burner was drained 17 days before Curve's

On 17 November, Curve Finance said an attacker had taken $38,200 of its fees. It was not users' money. It was the share of trading fees that Curve keeps for the people who govern it. Curve also asked teams that had copied its code to check their own.

Curve runs one of the oldest DEXs, exchanges that run on the blockchain with no company in the middle. Balancer runs another. Both take a small fee on each trade and send part of it to their DAO, the group of token holders who vote on how the project is run. Both also use the same kind of program to sell those fees once they pile up. Each calls it a fee burner.

Bitquery reads every transaction on Ethereum and on many other chains. So we did what Curve asked. We found every fee burner built the way Curve's was, then checked who each of their 11,021 sales had paid.

One other burner had been drained. Balancer's fee burner on Ethereum lost about $1,500 in ETH on 24 October. The flaw was the same, and so was the trick. That was 17 days before the attack on Curve. We found no public report of it, from Balancer or anyone else.

The same trick drained two burners, 17 days apartthe Balancer and Curve drains side by side
Balancer v3CurveDate (UTC)24 October 202610 November 2026Forged orders26138Time for most of them17 minutes17 minutesTaken0.64 ETH ($1,512)$38,206 in USDCPaid toA new wallet, 0x33b8…1c54A new wallet, 0x356e…64b8Signature reusedOne real order, every timeOne real order, every timePublic statementNone17 November
Dollar values at the time of each drain. Curve's figures match its own report.

02 · The fee burnerWhat a fee burner does

A DEX collects its fees in whatever tokens people trade. After a week, the pile can hold more than a hundred tokens, many worth a few dollars each. A DAO wants one token it can use. Curve wants crvUSD, its own stablecoin, a token built to stay at $1. Balancer wants USDC.

The fee burner sells the pile. Curve's burner and Balancer's both sell through CoW Swap. On CoW Swap, you post an order, and outside traders called solvers compete to fill it at the best price. The solver pays the network fee, called gas, and the order's owner pays nothing up front.

A person signs a CoW order with the key in their wallet. A program has no key. So CoW asks the program a yes-or-no question instead: is this order yours? The rule for that question is called ERC-1271. If the program answers yes, CoW treats the order as signed.

03 · The flawHow the Balancer hack worked

When CoW asks that question, it sends two things. The first is the order it is about to fill, in a short coded form called a hash. The second is the signature, extra data that the order's owner attaches. For these burners, the signature holds a copy of an order.

The burner checked the copy. Is this one of my orders, selling my tokens and paying my DAO? If so, it said yes. It never checked that the copy was the order that CoW was about to fill.

So anyone could post an order in the burner's name. Sell this token, and pay the money to me. Then they could attach a copy of a real burner order as the signature. CoW would ask, the burner would say yes, and a solver would fill the order.

The burner checked the copy, never the orderhow one forged Balancer order passed
The order CoW filledSell siUSD from the burnerPay the ETH to 0x33b8…1c54The signature attached to itBalancer's own order: sell 1.30 DAIPay the USDC to BalancerThe burner's checkIs the attached order mine?Yes, so the answer is validIt never compared the two ordersWhat happenedCoW settled the forged order.The ETH went to the attackerA smart contract cannot sign with a key, so CoW asks it a yes-or-no question (ERC-1271).CoW passes the burner two things: the order it is about to fill, as a hash, and the signature.The burner checked the signature and ignored the hash. Any order could borrow a real one's signature.
The first forged Balancer order, as its transaction records it. Every other forged order on both burners worked the same way.

Balancer's version of the check sits in a function called verify. In the code Balancer published for this burner, the hash arrives with no name, so the code can never read it. CoW's own template for programs like this one does compare the two orders. Curve's report shows that its own version says in a comment that it ignores the hash.

04 · BalancerBalancer lost 26 small balances in 17 minutes

At 02:13 UTC on 24 October, the first forged order filled. It sold the burner's siUSD for ETH and paid the ETH to a wallet that had never been used. Ten minutes later, the rest came quickly. By 02:30, the thief had sold 26 tokens.

Every order carried the same signature. It was a copy of a real Balancer order to sell 1.30 DAI for USDC and pay Balancer's fee wallet. The record of each transaction shows it. The orders that filled sold other tokens and paid ETH to the new wallet.

26 tokens, worth $1,512 togetherwhat each forged Balancer order sold, in dollars
$0$100$200$300siUSD$237PRL$192DVstETH$286PXT$164rstETH$145Bold$132ALCX$130msY$93sUSDat$30RLP$18PNKSTR$19phUSD$11USAT$9RPL$6RZR$6sUSDf$5WAO$5PROS$61R0R$5USDaf$5tBTC$2T$2ONE$2eBTC$1USDT<$1DAI$1
Listed in the order they filled, from 02:13 to 02:30 UTC on 24 October. The early orders took the biggest balances.

The thief took the burner's leftovers. These were small balances that had not yet sold. The biggest, in DVstETH, was worth $286. Half were worth less than $8. Together they came to about $1,500.

That is small for a DeFi hack. It is still about 12 days of the fees these two burners normally sent Balancer, based on their median day in September and October.

The new wallet never needed ETH to start. Solvers paid the gas for each order. The thief paid nothing until it was time to move the money.

05 · The moneyBalancer's ETH went to an instant exchange

About 80 minutes after the last order, the wallet split the ETH in two. It sent 0.16 ETH to one new wallet and 0.48 ETH to another. At 10:47 UTC the same day, both sent everything to one address. Bitquery's labels mark that address as a deposit wallet of CCE.cash, an instant exchange that swaps one coin for another without an account.

There, our trail ends. An instant exchange pays out from its own funds, so the coins that leave it are not the coins that came in.

06 · Balancer's burnerThe burner went quiet the next day

On a day when they sold anything, Balancer's two Ethereum burners made between one and five sales. On 24 October, they made 27. Twenty-six were forged. On 25 October, each made one ordinary sale. Then they stopped. Through 17 November, neither has sold anything since.

26 forged sales on 24 October, and none after 25 Octobersales per day by Balancer's two Ethereum burners
0102030No sales since 25 Oct24 Oct: 26 forged1 Sep1 Oct1 NovSales that paid BalancerForged sales
A sale counts when a burner sent a token to CoW. Data from the Ethereum dataset.

That may not be a reaction to the drain. Balancer was winding down. A year before, in November 2025, a rounding flaw in its older v2 pools had cost it far more. That flaw sat in other contracts. Most of its pools moved to withdrawals only on 30 October, after a DAO vote, and Balancer said its bug bounty ended that day. Balancer's public list of its programs still marks the drained burner as active.

07 · CurveCurve lost $38,200 to the same trick

Curve's burner sells once a week, in a one-day window that opens at midnight UTC each Tuesday. On 10 November, the first forged order filled seven minutes after the window opened. Within 17 minutes, 133 had filled. The last five came over the next 11 hours. Each paid USDC to one new wallet, $38,206 in all.

Both attackers sold the biggest balances firstevery forged order, by minute and dollar value
$1$10$100$1,0000 min5 min10 min15 min20 minBalancer, 24 Oct (26 orders)Curve, 10 Nov (133 of 138 orders)Dollar value per order, log scale. Each drain starts at minute 0.
Minute 0 is each drain's first order. Balancer's first order came ten minutes before the rest. Curve's last five orders, over the next 11 hours, are left out.

The pattern matches Balancer's. Curve says the attacker reused one real order as the signature for all 138. The orders took small balances that had not sold, and the median order was worth $254. In both drains, the orders ran roughly from the biggest balance to the smallest. Curve's regular sales that week went through as normal and raised 97,318 crvUSD for its DAO.

Curve's Emergency DAO, a small group that can act fast in a crisis, turned the burner off that afternoon. A DAO vote to replace it closes on 20 November. The new burner checks that the two orders match.

08 · Two trailsThe two attackers left different trails

Curve's attacker funded a wallet the day after the Balancer drain. On 25 October, 43 hours after the first Balancer order, a wallet received 0.064 ETH from Railgun, a privacy tool that hides who sent coins. Two minutes later, it sent 0.03 ETH to the wallet that would collect Curve's USDC. That ETH paid for the wallet's own transactions later.

Balancer's ETH went to an instant exchange, Curve's USDC to Hyperliquidwhere the money from both drains went
Balancer, 24 OctoberBalancer burner26 token balancesCoW settlementsolvers fill orders0x33b8…1c540.64 ETH0x1751…10af0.16 ETH0xd8a6…a6f80.48 ETHCCE.cash wallet10:47 UTC, same dayNo transfer links the two attackers' walletsCurve, 10 NovemberRailgun25 Oct, 20:58 UTC0x15bb…41980.064 ETH0x356e…64b8gas: 0.03 ETHCurve burner$38,206 USDCArbitrum bridgeUSDC to HyperliquidHyperliquid65 XMR1Tornado Cash0.3 ETH, 15 Nov
The Curve trail after the attacker's wallet comes from Curve's report. Short addresses stand for the full ones in the record below.

After the attack, Curve's report says, the money went to Arbitrum and then to Hyperliquid, an exchange that runs on its own chain. There it bought 65 XMR1, a token named after the privacy coin Monero. We have followed XMR1 before, in the Safe wallet drain. A last 0.3 ETH went into Tornado Cash, a mixer.

Are the two attackers the same person? We cannot say. They used the same method, and both sold the biggest balances first. But no on-chain transfer links their wallets. Balancer's thief was paid in ETH and used an instant exchange. Curve's was paid in USDC and used Hyperliquid.

One more thing shows up in the Curve attacker's wallet: 2,324 USDC from an address that looks like CoW's. Those are fake tokens, sent by bots that copy real transfers, a trick called address poisoning. They are not part of the theft.

09 · Who elseWe checked 11,021 burner sales on four chains

To find other burners, we listed every program that answers CoW's yes-or-no question about its own orders. We looked on Ethereum, Arbitrum, Base and Optimism. Two teams came up, Curve and Balancer. We checked each list against Balancer's public list of its programs and against Curve's own transactions.

Then we checked every sale those burners ever made, on Arbitrum and Base as well. A sale is right if it paid the team's own fee wallet in the same transaction. Of 11,021 sales, 164 were not, and all of those were on Ethereum, in the two drains.

164 of 11,021 sales paid an attacker, all in the two drainsevery burner sale we checked, by chain
Paid the right walletPaid an attackerEthereum · Curve6,420 of 6,558 · 138 forgedEthereum · Balancer681 of 707 · 26 forgedEthereum · Balancer460 of 460Ethereum · Balancer (retired)18 of 18Arbitrum · Curve86 of 86Arbitrum · Balancer672 of 672Base · Curve38 of 38Base · Balancer2,482 of 2,482Bar length is the share of each contract's sales. Optimism's three Balancer burners never sold anything.
Data from Bitquery. The full histories are also sold as files, as the Arbitrum dataset and the Base dataset.
Balancer fee burners, by chainWhat we found
EthereumDrained on 24 October. No sales since 25 October
ArbitrumAll 672 sales paid Balancer
BaseAll 2,482 sales paid Balancer
OptimismCollected fees, never sold
Gnosis, Avalanche, HyperEVM, PlasmaListed as active. Not checked

Balancer also lists live burners on Gnosis, Avalanche, HyperEVM and Plasma. They run the same code. We cannot see those chains, so we cannot say whether anyone has tried them.

10 · The limitsWhat the chain cannot tell us

We do not know who the attackers are. We do not know whether they are one person or two.

We did not follow Balancer's money past the instant exchange. The CCE.cash name is Bitquery's label, and we did not confirm it with the service.

We counted only burners built the way Curve's was. A team that copied the code into a new design might not show up in our search.

We do not know why Balancer's burners stopped selling on 25 October.

11 · The methodHow we checked the burners

We started from the programs that CoW's order system asks about their own orders. For each one, we listed each sale the burner made through CoW and checked that the team's fee wallet was paid in the same transaction. For each failed sale, we read the order that was filled and the order attached as its signature.

We read transfers and calls from Bitquery's smart contract data, and names from Bitquery's own labels. Dollar values use Bitquery's price for ETH at the hour of the drain. Our count of Curve's forged orders matches Curve's report.

The full history is also sold as files, as the Ethereum dataset on the Bitquery Data Store. For cases like this one, we also run crypto investigation services.

Limits on these findings

Transfers, calls and sales were read up to 17 November 2026. Anything after that is not covered.

We checked Ethereum, Arbitrum, Base and Optimism. Balancer lists burners on Gnosis, Avalanche, HyperEVM and Plasma that we did not check.

We found burners by one design: a program that CoW asks about its own orders. A copy of the code built another way would not show up.

A sale counts as right when the team's own fee wallet was paid in the same transaction. We read each sale that failed the test by hand.

The Curve trail past the attacker's wallet, to Arbitrum, Hyperliquid and Tornado Cash, comes from Curve's report. We did not trace it ourselves.

The CCE.cash name is Bitquery's own label on one address. We did not confirm it with the service.

We attribute both attacks to nobody. Shared methods and timing do not show that one person carried out both.

We do not know why Balancer's burners stopped selling on 25 October.

The Balancer dollar figure uses an ETH price of $2,380, Bitquery's hourly price at 02:00 UTC on 24 October.

12 · The recordThe addresses behind both drains

0x9958317b80ee5f10457017d54c2484d722059157
0x9789073eeb82d1b77c2ae4f96c22e22c63f5b82fb2f25838f72f6f80127416b4
0x840938c698334aae5cbf414024f114d833d2084499450bf7ca929bb8654ec981
0x33b80a1fe45b3307c0c5665abd2408167d731c54
0x175106f27eb03f240d9bcc18120e21945fa710af
0xd8a6292439c0d5612d084accacf933eb46fea6f8
0x0361897d757d13a4afad64a2e1bc561b96a8c7cf
0x9ff471f9f98f42e5151c7855fd1b5aa906b1af7e
0xc0fc3ddfec95ca45a0d2393f518d3ea1ccf44f8b
0x356e0178b0f07fe7d97c3a61e502ce52bc5564b8
0x15bb2f1f9d697801c70e841b1876bea72cb74198
0xd85eb231ce6f255f36cad475c00c5041dc3ca9f247261f36c37c939f10625087
0xa2bcd1a4efbd04b63cd03f5aff2561106ebcce00
Run it yourself

Check a contract's sales in plain English

Every transfer and call in this story comes from Bitquery's data. The Bitquery MCP server puts that data behind an AI assistant, so you can ask what a contract sold and who got paid without writing a query. The full history is also sold as files on the Bitquery Data Store, as the Ethereum dataset.

List every sale a contract made through CoW SwapCheck who was paid in each transactionFollow ETH and USDC from a wallet, one transfer at a timeCheck an address against Bitquery's labels
Explore Bitquery MCP →Ethereum, Arbitrum, Base and Optimism data from Bitquery. Data read on 17 November 2026.
Legal disclaimer

This article is provided for informational and educational purposes only and reflects analysis of publicly available on-chain data as of the dates indicated. It does not constitute legal, financial, compliance, or investment advice, and nothing in it is a recommendation to buy, sell or hold any token or asset, or to use or avoid any exchange, wallet, bridge, protocol or service.

The findings describe transactions observed on Ethereum, Arbitrum, Base and Optimism up to 17 November 2026. Attributions come from on-chain behaviour and Bitquery's own labels; they may be incomplete or incorrect and may be revised as more data becomes available. Curve Finance has published its own report; Balancer has not published one, and any later report may differ from what is described here.

References to Balancer, Curve Finance, CoW Swap, CoW Protocol, CCE.cash, Railgun, Hyperliquid, Tornado Cash or any other named project, company or service describe what the public record shows and what those parties have said in public. They are not statements about any party's intent, security practices, solvency or conduct, and nothing here asserts that any named party acted unlawfully or negligently or knew of the events described. The wallets that took and moved the money are described by their behaviour and attributed to nobody.

Nothing herein should be relied upon as a definitive determination of fact. Readers should conduct their own independent verification before taking any action. The authors and publisher accept no liability for any loss or damage arising from reliance on this material. All trademarks and company names are the property of their respective owners.

Reported by Gaurav Agarwal for Bitquery Research, with AI tools; every figure was checked against the raw data.

FAQ

Was Balancer hacked?

Yes, in one place. On 24 October, an attacker took about $1,500 in ETH from Balancer's fee burner on Ethereum. The burner held only fees owed to Balancer's DAO. It held no user deposits. We found no attack on Balancer pools.

How did the Balancer hack work?

The burner sells fees through CoW Swap and approves its own orders with a yes-or-no check. The check looked at a copy of an order attached as the signature. It never looked at the order being filled. The attacker attached a copy of one real Balancer order to 26 orders that paid the attacker instead.

Is this the November 2025 Balancer exploit?

No. In November 2025, attackers drained Balancer v2 pools through a rounding flaw, a far bigger loss. The October 2026 drain hit a fee burner in Balancer v3 and took about $1,500.

Is it the same bug as the Curve Finance hack?

Yes. Curve's burner had the same flaw and lost $38,200 on 10 November. Curve's report says the attacker reused one real order as the signature for every forged order. The Balancer attacker did the same.

Were user funds affected?

No. Both burners held only fees collected for the DAO. Neither attack touched pools, loans or user deposits.

Is the same person behind both attacks?

We do not know. The two attacks share a method and came 17 days apart. No transfer links the two wallets, and the money left by different routes.

Are other fee burners at risk?

Every Curve and Balancer burner we can see on Ethereum, Arbitrum, Base and Optimism is turned off, has stopped selling or has never sold. Balancer lists live burners with the same code on Gnosis, Avalanche, HyperEVM and Plasma, which we did not check.

Run the next investigation in minutes

Ask any question of the blockchain. Bitquery's AI Investigations agent traces funds, attributes wallets, and writes auditable reports across every major chain. Prefer to do it yourself? Connect the Bitquery MCP server to your AI assistant, pull the raw datasets, or query the chain directly with the Bitquery API.