K3 Capital’s sBOLD vault, a savings product built on Liquity’s BOLD stablecoin, was emptied of about $3.42 million in eight minutes on Wednesday 18 November, and most of that money did not end up with the attacker. According to an on-chain investigation published by Bitquery the same day, Ethereum block builders received about $1.78 million of it, or 52%, in payments from bots racing to copy the exploit. The wallet that first put the attack on chain and a second wallet it funded took about $0.52 million.
CoinDesk, which reviewed the same data in a report published at 05:39 UTC on 19 November, adds a caution that matters for everything below. The $1.78 million is what builders were paid, not what they earned: builders pass part of their income to the validators who publish their blocks, and the analysis does not establish how much any builder kept.
K3 Capital has paused the contracts and said it is investigating. By midday UTC on 19 November it had not published a loss figure, a cause or a plan for depositors.
Where the $3.42 million went, by Bitquery’s count
Bitquery read every transaction that minted sBOLD shares on 18 November. It counts 97 drain transactions, sent by six wallets across 37 blocks, between 12:25:23 and 12:33:35 UTC. The vault lost 3,428,799 BOLD, which was everything it held. Blockaid, the security firm that raised the alarm at 12:50 UTC, had put the loss at about $3.43 million “so far” across about 34 transactions. Both figures describe the same incident.
One group took more than the five others combined. A bot that sent its transactions from 13 wallets through a single contract made 57 of the 97 drains and pulled out about $1.91 million, 56% of the total. It paid 930 ETH to the builders of the blocks it landed in, about nine ETH in every ten it took, and moved the rest to one payout wallet. Another wallet, the first to drain the vault, paid builders a further 56 ETH. That makes 986 ETH, which Bitquery values at $1.78 million with ether at $1,801, its price at 12:30 UTC on 18 November.
The table shows where Bitquery located the money at 16:14 UTC on 18 November, at block 26,443,076. Wallets are named by the first characters of their address; none is attributed to a person.
| Where the money was at 16:14 UTC | Value | Share of the loss |
|---|---|---|
| Block builders (986 ETH received) | $1.78 million | 52% |
| Copycat bot’s payout wallet | $231,000 | 7% |
| THORChain vaults (sent by the attacker, no memo) | $292,000 | 9% |
| Maya Protocol, with memos asking for bitcoin | $232,000 | 7% |
| Wallet 0xbabe and a fresh wallet it sent 100 ETH to | $216,000 | 6% |
| Wallet 0xf601 | $174,000 | 5% |
| Wallet 0x2324, the first to drain | $75,000 | 2% |
| Attacker’s two wallets (0xdc0e, 0x8fb8) | Under $1,000 | Below 1% |
| Fees and costs (Liquity fees, swaps, gas) | $428,000 | 13% |
Bitquery sums the first two lines to 59% and writes that builders and the bot “kept” that share. On the builders’ side, its method counts payments to the address that collected each block’s fees. It does not follow what a builder then paid on to a validator, which is the gap CoinDesk points to.

Block builders, bribes and copycat bots: who does what
On Ethereum, the validator chosen to propose a block rarely assembles it. Specialised firms called block builders collect transactions, decide their order and bid for the right to have their block published; the winning builder pays the validator. Our guide to block builders and sandwich attacks describes that market in detail.
Order is worth money. When several parties want the same profit, the first transaction in the block gets it and the others fail. A searcher that needs to be first can pay the builder directly inside its own transaction, on top of the normal gas fee. Traders call that payment a tip or a bribe, and the profit available from ordering transactions is known as maximal extractable value, or MEV.
Some bots do nothing but watch for profitable transactions or contracts and reproduce them with their own address swapped in. The trade name is generalised front-runner. Such a bot does not need to understand what it is copying. It simulates the transaction, sees that it ends with more ether than it started with, and bids for the top of the block.
That is the pattern Bitquery describes here. Once the attack was public, the vault’s remaining BOLD became a prize that several wallets could claim with the same code, and the bidding rose as the balance fell. The bot paid 12.5 ETH for its first drain, and the price reached 44 ETH for one block near the end. Titan built most of the blocks concerned, Bitquery writes, and BuilderNet, Ultrasound and Quasar built the rest.
Bitquery is explicit about what this does and does not show. Paying a builder for a place in a block is normal practice in Ethereum MEV, it writes, adding: “We have no sign that the builders knew what the bot was doing.” Neither Bitquery nor CoinDesk reports a statement from Titan or any other builder, and neither reports a return of funds.
How the sBOLD exploit worked: eight rounds in one transaction
BOLD is the dollar stablecoin of Liquity V2, minted by borrowers who lock ETH, staking means depositing a proof of stake coin with a protocol that stakes it for you and issues a transferable token (such as Lido stETH or rETH) representing the deposit plus rewards, which you can use or sell while the coins stay staked.">wstETH or rETH in a loan called a trove (our guide to how CDP stablecoins such as BOLD work covers the model). Liquity runs three Stability Pools, one per collateral, where BOLD holders earn interest and absorb liquidated collateral. sBOLD sits on top: depositors hand BOLD to the vault, receive shares, and the vault spreads the BOLD across the three pools. When the pools hand it collateral, a function called swap(), which anyone can call, sells that collateral back into BOLD.
The weak point, in Bitquery’s reading, was the price of a share. The vault ignored collateral it held below a set amount when it valued itself. Someone who placed a little WETH in the vault could therefore deposit BOLD at a share price that left that WETH out, call swap() so the WETH became BOLD that did count, and redeem at the higher price.
The transaction Blockaid flagged, recorded at 12:25:47 UTC in block 26,441,931, shows one such run. Its sender, the wallet 0xdc0e, took a flash loan of 10,000 wstETH from Morpho, opened a Liquity trove with it and minted about 9.01 million BOLD. Then, by Bitquery’s account, it repeated one round eight times: deposit the BOLD into sBOLD, send about 4 WETH straight to the vault, call swap(), redeem for about 10,600 BOLD more than went in. It closed the trove, repaid the loan and left with roughly $23,000. Etherscan lists 231 token transfers in that single transaction.
A profit of $23,000 per transaction explains the 97 transactions. It also explains why copying was easy: the whole recipe sat in one contract.
Blockaid says a trove was self-liquidated, Bitquery counts none
The two firms do not describe the entry point the same way. Blockaid’s alert on X said the attackers “flash-loaned wstETH, self-liquidated a Liquity V2 trove into the stability pools, minted cheap sBOLD shares, swapped the collateral” and redeemed other depositors’ BOLD. Bitquery says the security firm ExVul gave the same account.
Bitquery disputes that step. “No loan was liquidated,” its report says. It counts 95 troves opened and 95 closed on Liquity’s three markets during the drain, with 301 redemptions and no liquidation, and says the collateral in each round was WETH the attacker sent to the vault itself. Blockaid has not published a longer analysis, and K3 Capital has not said which description is right.
ChainSecurity rated a gap of this kind high in 2025
A weakness of this kind was on the public record. ChainSecurity’s audit of sBOLD, dated 19 May 2025, lists a high-severity finding titled “Collateral Value Is Not Accounted Below maxCollInBold Threshold” (CS-SBOLD-008). The auditors wrote that leaving collateral out of the share price up to that threshold, 7,500 BOLD in their example, let new depositors mint shares at a discount and redeem them at full value, and that repeated self-liquidations could use it to drain the vault.
The report marks the finding “Code Corrected”. It reviewed three versions of the code: in the second, collateral below the threshold was still left out of total assets as unrealised gains; in the third, the whole collateral amount is counted. Its summary of that third version says: “The possibility of value extraction through self-liquidations has been mitigated.”
Whether that finding is the bug used on 18 November is not established. ExVul wrote that the drain matches the path in the audit, Bitquery reports, but nobody outside K3 Capital has shown which version runs on chain. Bitquery says it did not read the deployed source code, so it cannot say whether the fix shipped or whether the drain used a path the fix missed. CoinDesk writes that it remains unclear whether the attack used a related mechanism or another accounting weakness. The audit carries the usual warning that such reviews “are time-boxed and cannot uncover all vulnerabilities”, a limit explained in our guide on how to read a smart contract audit report.

What K3 Capital and Liquity have said
The firm’s public account of the incident is two sentences long. “We are investigating malicious activity involving sBOLD. Contracts have been paused,” it posted on X at 13:08 UTC on 18 November, 35 minutes after the vault was empty by Bitquery’s timeline.
Liquity followed at 13:59 UTC. It described sBOLD as “a vault built and run independently” by K3 Capital, told users not to deposit until further notice and wrote: “Liquity V2 is NOT affected.” It said it was supporting K3. Bitquery’s data agrees on that point: troves opened, closed and paid redemptions normally during the drain.
Depositors were not the only people exposed. sBOLD also trades on Pendle as fixed-yield and variable-yield tokens, and Bitquery notes that holders of those tokens lost value too, without measuring how much.
Can money paid to builders come back?
There are precedents for funds returning from the MEV supply chain after a hack, and they differ from this case in one respect: the party that returned the money was holding it.
On 20 January 2026, an exploit of the Makina protocol’s DUSD/USDC smart contract holding reserves of two or more tokens, supplied by users, that a decentralized exchange uses to execute trades automatically at a price set by a formula.">pool on Curve involved about 1,299 ETH, of which an MEV builder received 1,023 ETH. Two days later, Makina said about 920 ETH had been returned to a recovery multisig, net of a 10% bounty under the SEAL Whitehat Safe Harbor programme, AMBCrypto reported on 22 January. A further 276 ETH or so had gone to an address identified as a Rocket Pool validator, which Makina was still trying to reach.
During the Curve pool exploits of 30 July 2023, the operator of an MEV bot known as c0ffeebabe.eth front-ran an attacker on the CRV-ETH pool and sent 2,879 ETH, about $5.4 million at the time, to Curve’s deployer address, The Block reported the next day.
In the sBOLD drain the money is spread more thinly. The 986 ETH arrived as payments for block space from two senders, across several builders, and part of a builder’s receipts normally goes on to validators. No builder has announced a refund and K3 Capital has not said it has asked for one. As for the attacker’s own share, Bitquery traced 162 ETH to three THORChain vaults, sent without the memo that normally says what to swap into, and a net 129 ETH into Maya Protocol with memos asking for bitcoin at five addresses. By 15:30 UTC on 18 November, two of those addresses had received 3.10 BTC between them.
The open questions
The document that would settle most of this is K3 Capital’s post-mortem. It should say which code version was deployed and which of Blockaid and Bitquery describes the entry point correctly. Whether any builder, or the operator of the copycat bot, returns ether is a separate matter that no party has addressed. The final loss, once Pendle positions are counted, is not known either. K3 Capital has announced neither compensation for depositors nor a date for reopening the vault.
Bitquery’s snapshot stops at 16:14 UTC on 18 November, when the wallets were still moving money. At that point the bot’s payout wallet held 149.9 ETH, wallet 0xf601 and its contract about 94.7 ETH, and a previously unused wallet 100 ETH received from 0xbabe.
