TL;DR
On November 18, 2026, K3 Capital's sBOLD vault on Ethereum, an ERC-4626 wrapper around Liquity V2's Stability Pools, was emptied of about $3.43 million across roughly 34 transactions by Blockaid's count, and 97 transactions in eight minutes by Bitquery's, which groups the senders into six actors. The first alerts blamed self-liquidated Liquity troves and collateral the vault left out of its share price. The record contradicts both. The flaw we see in the transaction logs is narrower: swap() checked that BOLD came in and never checked that the collateral went out. Attackers "sold" the vault 4 WETH it kept, redeemed shares at a price that counted the WETH twice, then pulled the WETH back through the approval swap() left behind. A copycat MEV bot pulled out 56% of the money and paid most of that to block builders. A vault that checks only what a swap paid in is trusting the caller to finish the swap.
What Happened to K3 Capital's sBOLD Vault?
sBOLD is a savings vault for BOLD, Liquity V2's dollar stablecoin. Depositors put BOLD in, the vault spreads it across Liquity's three Stability Pools (WETH, wstETH, rETH), and the pools pay interest plus discounted collateral from liquidated loans. Liquity's documentation identifies sBOLD as a third-party vault developed and maintained by K3 Capital.
The first drain landed at 12:25:23 UTC. By 12:33:35 the vault was empty: Bitquery puts the loss at 3,428,799 BOLD, about $3.42 million, everything the vault held. Our archive-node reads agree: totalAssets() was 3,428,804 BOLD in block 26,441,928, the last block before the drain.
The alarms trailed the drain. Blockaid posted at 12:50 UTC and called the exploit ongoing. K3 Capital answered at 13:08 UTC that it was "investigating malicious activity involving sBOLD" and that "Contracts have been paused". On-chain, the pause transaction landed in block 26,442,128 at 13:05:11 UTC, 40 minutes after the first drain and more than half an hour after the money was gone.
K3 Capital had not published a postmortem as of November 19.
What Is an ERC-4626 Share Price, and Where Does sBOLD's Come From?
An ERC-4626 vault mints shares in proportion to totalAssets(). Push totalAssets() up while you hold most of the shares, and you redeem for more than you put in. Every other holder pays the difference.
sBOLD's totalAssets() has two parts. One is the BOLD it holds in the Stability Pools, plus yield. The other is collateral it holds, priced by oracle and haircut for slippage and fees. That collateral normally arrives from liquidations. A public swap() function turns that collateral back into BOLD through a swap adapter, which on sBOLD is 1inch's AggregationRouterV6, called with calldata the caller writes.
One parameter matters here. maxCollInBold, which we read on-chain at 7,500 BOLD, makes deposits revert once the vault's collateral is worth more than that.
Did the Attackers Self-Liquidate, or Was Collateral Left Uncounted?
Bitquery's trove data and the vault's verified source rule out both, and the reconciliation is the most useful part of this story.
ExVul's alert wrote that sBOLD's totalAssets() "excludes unswapped Stability Pool collateral below its 7,500 BOLD accounting threshold", fed by self-liquidated troves, and that the path matches ChainSecurity's audit. In the ChainSecurity sBOLD report, finding CS-SBOLD-008, rated High, walks through exactly that drain: self-liquidate, mint shares while collateral below maxCollInBold is ignored, call swap(), redeem, repeat. The same report marks it "Code Corrected": "In Version 3, the collateral value is included in the valuation of the sBOLD vault."
The deployed vault runs that fix. Its verified source on Sourcify (exact match) computes totalAssets() as pool BOLD plus collInBoldNet, the collateral's oracle value after the slippage haircut, with no threshold carve-out. Bitquery, which did not read the source, said so plainly: it could not tell "whether the fix in ChainSecurity's report shipped, or whether this drain used a path the fix missed."
The per-round arithmetic points the same way. In the first round of the first drain transaction, sent at 12:25:23 UTC by 0x2324 running a copy of the author's contract (more on that below), the attacker deposited 9,000,000 BOLD for 8,209,404 sBOLD, roughly 72% of the shares then outstanding. It paid 7,159.16 BOLD into swap() and redeemed for 9,010,367.62 BOLD, a 10,367.62 BOLD surplus. If the vault had ignored the attacker's WETH, the round would have added only the 7,159 BOLD, and a 72% holder would recover about 5,180 of the 7,159 it paid. That is a loss. The observed surplus needs about 14,318 BOLD of new value, which is the BOLD payment counted once and the WETH counted again on top.
And across all three Liquity markets during the drain, Bitquery counted troves opened and closed 95 times each, 301 redemptions, and not one liquidation.
The Root Cause: swap() Verified Its Output, Never Its Input
swap() hands the adapter an allowance for the collateral, makes the call, and checks only the BOLD side. The function, which Dedaub's report reproduces verbatim in its note on BOLD pricing, approves _inAmount of the source token to the adapter, calls it with the caller's _swapData, then compares the vault's BOLD balance before and after against _minOut. ChainSecurity's system overview describes the same design: the adapter "is called with user provided data, and it is checked that the call results in the desired increase in BOLD balance."
So the vault decides which collateral it is selling, and the caller's calldata decides which token the router actually pulls. Nothing checks that the two match, that the collateral balance went down, or that the allowance is gone afterwards.
The event log of the first drain, together with our call trace of the same transaction, shows each round:
- The attacker transfers 4 WETH to sBOLD (log 45).
- sBOLD approves the 1inch router for 4 WETH (log 48) and calls the router's
swap()with the attacker's calldata. That calldata names as its source token a contract the attack contract had deployed minutes earlier, for an amount of 1 wei, with sBOLD as the BOLD receiver. The router pulls 1 wei of the dummy token, the attacker's executor supplies BOLD, and 7,159.16 BOLD reaches the vault. - sBOLD emits
Swap: 4 WETH in, 7,159.16 BOLD out, minimum 7,159.16 BOLD (log 51). The attacker paid the minimum to the wei. No WETH has left the vault. - The attacker redeems its shares (log 90).
- It sends sBOLD 1 wei of wstETH and calls
swap()on the wstETH pool, which approves the router for 1 wei (log 94). This time the attacker's calldata names WETH as the source token, amount 4 WETH, with the attack contract as receiver. - The router calls
transferFromon WETH and moves 4 WETH from sBOLD to the attacker (log 95), spending the allowance the first call left standing. The vault logs a 1-wei wstETH sale for 3,000 wei of BOLD (log 98).
The vault booked a WETH sale it never made. For the length of one redemption it held the 7,159 BOLD of proceeds and the 4 WETH those proceeds were meant to replace, and priced its shares on both. Then the second "swap", a 1-wei decoy, carried the real WETH out.
The 7,500 BOLD deposit gate explains the sizing. At an oracle value of about 7,195 BOLD, 4 WETH sits just under the cap, so deposits and redemptions keep working while it is in the vault. Each round could fake at most one cap's worth of collateral, which is why the first transaction ran eight small rounds instead of one large one.
The Attack, Step by Step
Figures are from that first drain transaction. Bitquery says the author's own drain at 12:25:47 and the others ran the same steps at different sizes.
| Step | Action | On-chain effect |
|---|---|---|
| 1 | Flash-borrow 10,000 wstETH from Morpho | Capital for one transaction |
| 2 | Open a Liquity V2 trove on the wstETH | 9,010,000 BOLD minted to the attacker contract |
| 3 | Deposit 9,000,000 BOLD into sBOLD | 8,209,404 sBOLD minted, about 72% of supply |
| 4 | Send 4 WETH straight to the vault | Collateral worth about 7,159 BOLD after haircut enters totalAssets() |
| 5 | Call swap() on WETH; calldata pulls 1 wei of a dummy token and pays 7,159.16 BOLD | Vault logs 4 WETH sold, still holds it, share price counts both |
| 6 | Redeem all 8,209,404 sBOLD | 9,010,367.62 BOLD out, 10,367.62 more than deposited |
| 7 | Call swap() on 1 wei of wstETH; calldata names WETH as the source | Router spends the stale WETH allowance, 4 WETH returns to the attacker |
| 8 | Repeat steps 3 to 7, eight rounds in all | Each round nets about 3,200 BOLD from other depositors |
| 9 | Close the trove, repay Morpho, redeem leftover BOLD with Liquity | 21,480 BOLD burned for WETH, wstETH and rETH |
Each round calls swap() twice, once to pay in and once to take back:
Who Actually Took the Money?
Mostly block builders. The wallet that wrote the exploit, with a second wallet it funded, took about $0.52 million, roughly 15% of the loss.
Bitquery's bytecode comparison shows the wallet starting 0xdc0e, the author, deployed the attack contract at 12:14:11, eleven minutes before any drain, and never called that first deployment. It redeployed and ran its own first drain at 12:25:47. The wallet ExVul named as the attacker, 0x2324, ran a byte-for-byte copy with its own address swapped in and drained first, at 12:25:23. Bitquery labels 0x2324 a MEV bot with 978 prior transactions and reads it as a generalized frontrunner that copied the code once it went on chain, though it cannot see how. That put the copy 24 seconds ahead of the author.
Then a second copycat bot arrived at 12:26:47. Operating from 13 zero-prefixed wallets, it sent 57 of the 97 drains, and its transactions drew $1.91 million out of the vault, 56% of the loss. The bot bid for block position the normal MEV way, paying builders inside its own transactions, about nine ETH in every ten it took: 12.5 ETH for its first drain, rising to 44 ETH for one block near the end. Builders collected 986 ETH in all, about $1.78 million or 52% of the loss: 930 ETH from the bot's drains and 56 ETH from 0x2324's. The bot itself kept about $231,000, 7% of the loss, in its payout wallet. Bitquery adds those two to get the 59% in its headline. Its full split of the loss: builders 52%, the bot's wallet 7%, the author's THORChain and Maya deposits about 15%, three other early wallets 13%, and fees and costs 13%.
The author's two wallets took about $0.52 million and, within three hours, sent 162 ETH to THORChain vaults with no memo and 129 ETH through Maya Protocol toward five Bitcoin addresses.
Each round needed only a flash loan and 4 WETH. The first transaction's rounds paid about 3,200 BOLD each, and later drains were bigger: Bitquery found the bot's transactions drew about twice as much near the end as at the start. It puts about $428,000 of the loss into Liquity fees, swap costs and gas.
Why This Matters Beyond sBOLD
A swap check that only watches the output is a price check. It confirms the vault was paid enough for whatever the caller says it bought. With a general router and caller-written calldata, the caller decides whether the input moves at all. The Summer.fi donation-inflation drain was another vault whose share price trusted a balance an attacker could move.
An unconsumed approval is a standing withdrawal right. An approval with no reset after the call belongs to whoever controls the next call through the same spender. We wrote up the same failure in the jaredfromsubway honeypot, where allowances the bot granted were never consumed or revoked, and one coordinator transaction swept them. sBOLD's allowance lived for one swap() call, long enough for the next call to spend it.
ChainSecurity's fix closed the path it described. The High finding was about collateral the share price ignored, and the fix made the share price count it. The exploit used the opposite failure: collateral the share price counted after it had been paid for. Dedaub's report reproduces _execute(), and ChainSecurity's overview describes its BOLD-only check. ChainSecurity's finding list fixed reentrancy through the adapter, and Dedaub's summary lists no critical, high or medium issues. We found no finding in either published report for an adapter call that pays BOLD without taking the collateral. A vault that prices its shares off an attacker-reachable call needs the invariant written down and tested: no call may raise totalAssets() by more than the value that left the caller.
What Operators Should Do
- Check both sides of every swap. Record the source-token balance before the adapter call and require it to drop by the amount you approved. Treat a
Swapevent whose collateral did not move as a failed swap. - Zero the allowance after the call. Use
forceApprove(adapter, 0)(or approve-call-reset) around any external call that takes caller-supplied data. - Do not let a permissionless caller write router calldata. Restrict
swap()to keepers, or route through an adapter you own that builds the route itself and enforcessrcToken,amountandreceiver. - Fuzz the share-price invariant. Drive
swap()with adversarial calldata inside a same-transaction deposit and redeem, and asserttotalAssets()never rises by more than BOLD received minus collateral value removed. - Alert on share-supply shocks and make the pause fast. A single deposit minting 2.6 times the existing supply and redeeming in the same transaction is an attack signature. Give a guardian role or an automated responder the right to call
pause(), so the response does not wait on a multisig quorum.
Would Monitoring Have Changed the Outcome?
Not with the pause sBOLD had. Each drain was atomic, so nothing alerts ahead of the first block, and our archive reads show totalAssets() falling from 3,428,804 to 3,327,814 BOLD in block 26,441,929. That first block, about 101,000 BOLD, was gone before any signal existed.
After that, every block counted. Our reads put totalAssets() at about 1,498,000 BOLD at 12:29:35 (block 26,441,950) and under 15,000 BOLD at 12:32:35 (block 26,441,965). A pause in the block after the first drain keeps about 3.33 million BOLD in the vault. A pause four minutes in still keeps about 1.5 million. A pause after 12:32:35 keeps almost nothing.
pause() is onlyOwner, and the owner is a Gnosis Safe that we read on-chain as 2-of-3. No other address can pause. The pause transaction is a Safe execution that landed at 13:05:11 UTC, 40 minutes after the first drain and about 32 minutes after the vault was empty.
The signals were there from the first block. One deposit minted 8,209,404 sBOLD against a supply of about 3.13 million, then burned all of it in the same transaction. The vault emitted a Swap for 4 WETH while its WETH balance did not fall, with proceeds equal to minOut to the wei. And totalAssets() dropped about 3% in one block with no liquidation anywhere in Liquity.
An alert is only as fast as the lever behind it, and sBOLD's lever needed two signatures. The Safe could have enabled a module holding only the right to call pause(), with no change to the vault itself. Tripwire is built for that setup: it checks rules like the three above against contract state on every block and pending transaction, and sends the pause through a pause-only role, so the response lands in the next block instead of after a signing round. The bug itself is a code finding, and an adversarial DeFi security audit of the swap path is where it gets closed for good.
Frequently Asked Questions
How much did K3 Capital's sBOLD vault lose?
About $3.43 million, everything it held. Blockaid's alert said $3.43 million; Bitquery counted 3,428,799 BOLD, about $3.42 million at its BOLD price.
Did the attackers self-liquidate Liquity troves?
No. Bitquery found zero trove liquidations during the drain. Each round's collateral was 4 WETH the attacker sent to the vault.
Was this the bug ChainSecurity flagged in its sBOLD audit?
No. ChainSecurity's High finding CS-SBOLD-008 covered collateral the share price ignored below maxCollInBold, and the deployed source carries the Version 3 fix that counts it. This drain abused swap() accepting BOLD without spending the collateral, plus the allowance it left behind.
Who ended up with the stolen funds?
Mostly block builders, paid by a copycat MEV bot that drew 56% of the loss. Builders received $1.78 million. The exploit's author and its second wallet took about $0.52 million, routed through THORChain and Maya Protocol.
Are Liquity V2's own Stability Pools affected?
No. The flaw sits in K3 Capital's sBOLD swap(), and the BOLD that left the pools left through sBOLD's own redemptions.
Sources / References
- Unchained: K3 Capital Pauses sBOLD Contracts, Blockaid Says Attackers Have Drained $3.43M
- Bitquery: K3 Capital sBOLD Hack: Bot and Builders Kept 59% of $3.42M
- crypto.news: Gonka bridge and sBOLD vault hit in $4.48M Ethereum exploits
- ExVul on X: exploit alert on the K3 Capital sBOLD vault
- ChainSecurity: Code Assessment of the sBOLD Smart Contracts (PDF)
- Dedaub: K3 Capital smart contract audit report
- Sourcify: verified source of the sBOLD vault, 0x50Bd66D59911F5e086Ec87aE43C811e0D059DD11
- Etherscan: first drain transaction 0xaa3c7f00
- Etherscan: event log of the first drain transaction
- Etherscan: sBOLD pause transaction 0x19cf2ead
- EIP-4626: Tokenized Vaults



