Skip to content
PROTOCOL RESEARCH

Curve LlamaLend risk

Curve is winding V1 down: no new V1 markets, a deprecation sweep in mid-2026 and a DAO risk-team mandate that names "the sunset of V1", while lenders in the CRV market have been unable to withdraw in full since October 2025 and a market oracle was exploited in March 2026.
AvoidResearched Aug 6, 2026

Scorecard

watch: Audits & contractsStateMind audited LlamaLend pre-launch (April 2024) and MixBytes covered the factory, Vault, Controller and AMM through May 2024; the core has held. The contract that lost money in March 2026 was a Curve-written oracle helper outside that reviewed surface.
weak: Admin controlV1 markets deploy permissionlessly with any oracle, factory implementations are DAO-upgradeable, an Emergency DAO multisig holds shared admin over LlamaLend loan discounts and vault deposit caps with no timelock, and the independent risk reviewer left in July 2026.
weak: OracleExploited 2 March 2026: a Curve oracle read a spot ERC-4626 exchange rate that any user can move in one transaction, and 27 borrower positions were liquidated. The market's oracle cannot be swapped, so it is being deprecated rather than fixed.
weak: Liquidity & exitCRV-long lenders have been roughly 70% backed and unable to fully withdraw since 10 October 2025; three small markets show more borrowed than supplied; and the whole V1 book is on a declared sunset path.
watch: Yield (real vs emission)Interest is real but thin at 0.7% to 3.3%, and in the largest crvUSD vault the CRV emission leg (1.16%) is bigger than the interest leg (0.74%).
weak: Holder concentrationTwo markets hold about 65% of the book, and the largest is collateralized by the stablecoin of a protocol that borrows from LlamaLend itself and was exploited in June 2025.
weak: Track recordTwo money-losing events on the lending product in eleven months, one still unresolved, plus two registrar-level DNS hijacks three years apart that cost users roughly $570K each.
🟢 strong🟡 watch / caveat🔴 weak / fund-loss risk
Verdict is a gate (worst flaw wins), not an average. Our read, not financial advice.

Live facts

auto-sourced now
TVL$62.2M
30d↑1%
Audits0
Last hacknone

DeFiLlama + our exploits feed. Cross-check the dated report against today.

TL;DR

LlamaLend is Curve's lending arm, a separate product from the Curve DEX. Its engine is LLAMMA, which liquidates a borrower gradually across price bands instead of at a single threshold. The engine is audited and has not been broken. The losses have come from the layer above it: who is allowed to create a market, and which oracle that market is pointed at. On 2 March 2026 an attacker manipulated the oracle of the sDOLA market and hard-liquidated 27 borrower positions. Since the 10 October 2025 crash, lenders in the CRV-long market have been roughly 70% backed and cannot fully withdraw.

Two things a reader needs before anything else. First, this V1 book is being wound down. No new V1 markets are being deployed, a sweep to deprecate legacy V1 markets ran through May and June 2026, and when the DAO advertised for a new risk team on 8 July 2026 the mandate explicitly included supporting "the sunset of V1". Second, the risk reviewer who would normally vet markets, LlamaRisk, is gone: its services engagement ran April 2025 to April 2026, it was the named market curator for the new v2 at launch in June 2026, and the DAO opened a call for replacements on 8 July 2026 with no appointment confirmed. TVL is $61.20M, down about 66% from an ATH of $182.70M in June 2025.

Two disambiguations. DeFiLlama also lists an unrelated protocol called Llamalend (slug llamalend, roughly $17k, category NFT Lending). Nothing here refers to it. Separately, we publish a report on Curve's DEX and crvUSD at /risk/curve-dex, refreshed 5 August 2026. That report owns the stableswap AMM, the crvUSD peg and the composition of crvUSD supply. This one owns the lending markets, and where they touch, the DEX report is the authority on crvUSD itself.

Checklist

Audits & contracts. The roster is real and the LlamaLend-specific work is better documented than a casual read suggests. StateMind audited the LlamaLend lending platform before launch, publishing in April 2024, finding 1 Critical, 1 High and 4 Medium issues, all reported fixed; Curve itself circulated that report as "Llamalend audit which was done before launch". MixBytes ran a longer engagement on Curve Lending from 6 February to 31 May 2024 with re-audits through 29 July 2024, covering more than thirty contracts including OneWayLendingFactory.vy, TwoWayLendingFactory.vy, Vault.vy, Controller.vy and AMM.vy, and reporting 1 Critical, 2 High, 5 Medium and 7 Low findings. ChainSecurity appears in Curve's audit history and published on Curve Stablecoin; its specific coverage of the LlamaLend factory implementations comes from a governance thread rather than a report I could open.

The core has earned the credit: no exploit has broken the LLAMMA AMM or the Controller. The caveat is sharper than usual, though. MixBytes' review focused in part on inflation attacks and oracle manipulation, and the protocol still lost money in March 2026 to an ERC-4626 oracle manipulation, because the vulnerable piece was a per-market configuration choice sitting outside the audited core. The contract that lost the money, CryptoFromPoolVaultWAgg at 0x88822ee5..., is Curve-authored Vyper 0.3.10, verified and non-upgradeable, and its own source comments say it is "only suitable for vaults which cannot be affected by donation attack (like sFRAX)". It was pointed at sDOLA anyway. Audited core, unaudited judgement at the edge.

Admin control. V1 markets are created permissionlessly through the OneWayLendingFactory. The create_from_pool path binds a market to a Curve pool's manipulation-resistant EMA oracle, which is the safe route. The plain create path lets the deployer supply an arbitrary oracle contract, which is the route that produced the March 2026 loss. Curve's own resources acknowledge that deployer-designated custom oracles "introduce additional trust assumptions" and need vetting precisely because deployment is permissionless.

Above that, factory implementations are upgradeable by the admin, which is the Curve DAO. Ownership actions need 30% quorum and 51% support, and the vote is followed by a further execution timelock, so the real proposal-to-execution distance is roughly a week of voting plus that delay, not a week in total. Sources disagree on the delay itself: Curve's own readthedocs describes a one-week timelock on ownership execution, while our curve-dex refresh recorded three days. Either way it is slower and more legible than a small multisig, and it deserves credit for that.

What undercuts it, and what the reader should weigh hardest, is that this path can be bypassed. Proposal #1252, which passed on 12 November 2025, gave an Emergency DAO multisig shared admin over crvUSD and LlamaLend risk parameters: debt ceilings, AMM fees, monetary policy, loan discounts and vault deposit caps. Ordinary changes run the vote plus timelock. Emergency ones do not. Those powers are bounded to risk parameters and are mostly risk-reducing rather than fund-seizing, which is the fair counterweight, but loan discounts and vault deposit caps are exactly the knobs that decide whether you can get out and on what terms.

Two more. The veCRV electorate is concentrated, with Convex holding a reported majority, established in our curve-dex refresh of 5 August 2026 and not re-derived here, and it is the same electorate that upgrades the lending factory. And the DAO does not merely set parameters, it directs credit: on 24 October 2025 it voted to mint 5M crvUSD through a new factory contract straight into the sreUSD lending market, explicitly to support a partner protocol. LlamaLend v2 fixes much of this by design, since v2 markets need DAO approval and launch with borrow caps at zero, and v2 is now live on both Optimism and Ethereum. But v2 is a separate DeFiLlama entry at about $3.02M, and the $61.20M is still sitting in the V1 book that is being retired.

Oracle. This is the dimension that has actually fired. At 03:00:11 UTC on 2 March 2026, block 24566937, an attacker flash-loaned roughly 10M USDC and 15,986 WETH from Morpho Blue and attacked the sDOLA/crvUSD market. The root cause is not in dispute: the oracle read sDOLA.convertToAssets() as a spot feed, and that value moves the instant anyone stakes, unstakes or donates, so it can be moved inside a single transaction. Curve's own remediation language names the path directly, saying LlamaLend oracles "should not have the potential for instantaneous price jumps for any reason, including the path involved here by donating to an ERC-4626 vault", and that smoothing should be standardized across all LlamaLend oracles.

The exact sequence is described differently by different sources and I am not going to pretend otherwise. Curve's post-mortem describes a two-stage attack: a large LLAMMA exchange first pushed positions into soft liquidation, after which the sDOLA rate was moved up about 13.79%, from 1.189 to 1.353, through the donation path. An independent writeup by DarkNavy instead describes a mass redemption that drove the rate down. Both accounts agree on the manipulable spot feed, on 27 liquidation calls against borrower positions, and on an attacker take of roughly $240,000 (6.74 WETH plus 227,326 DOLA). The post-mortem puts the debt involved near $10.9M and borrower equity seized near $822K. That $822K is the number that matters to a user, and it is the one most coverage omits.

Curve stated that lenders in that market were not affected and that borrowers were. The market cannot be repaired because it predates the introduction of oracle proxies to LlamaLend, so its oracle cannot be swapped; it is being deprecated instead, removed from the UI with borrow rates raised to push users out. Oracle proxies now exist for newer markets, which is a genuine fix going forward, and it does not help anyone who was in that market.

Liquidity & exit. The 10 October 2025 crash left the CRV-long market underbacked by about $700,000. Lenders there hold deposits worth roughly 70% of stated value and cannot fully withdraw. That has been true for about nine months. In April 2026 Egorov proposed a market-based fix rather than a bailout: tokenize the impaired claims and let trapped lenders sell them into dedicated crvUSD and cvcrvUSD pools at a discount, priced around 71% solvency with a 1% fee, giving buyers an option-like bet on CRV. Partial coverage begins near a CRV price of $0.957 and full modeled recovery near $1.242. Curve was implementing this as of 1 May 2026. It is an honest mechanism and it is worth saying plainly that it does not eliminate the loss or guarantee recovery; it converts a frozen position into a sellable one at a discount.

Live data adds two warnings. Of the 31 markets in DeFiLlama's lending feed, three currently show more borrowed than supplied: sDOLA on Ethereum ($55.8k supplied against $62.4k borrowed), ARB on Arbitrum ($31.4k against $49.7k) and wFRAX on Fraxtal ($15.7k against $161.6k). Those are tiny in dollar terms and the reading may partly be a feed artifact, but borrowed exceeding supplied is the on-chain shape of a market you cannot exit. The larger warning is structural: this is a book Curve has said it intends to sunset, and a deprecating market is one where liquidity leaves first and the last lender out gets the worst fill.

Yield: real vs emission. Lender yield is borrower interest, which is real revenue rather than a token subsidy, and that is the right structure. It is also thin. Base APY on the crvUSD vaults runs about 0.7% to 3.3%. CRV gauge emissions sit on top, and in the single largest vault, holding about $23.05M, the emission leg (1.16%) is larger than the interest leg (0.74%). Pools carrying any reward APY represent about 64% of pool TVL. There is a second, subtler drag: the 5M crvUSD of DAO-minted supply sitting in the sreUSD market exists specifically to push borrowing rates down, which means the largest market's lenders are earning a rate the DAO deliberately suppressed.

Holder concentration. The book is top-heavy. Two markets, sreUSD ($21.71M supplied) and sfrxUSD ($14.27M), account for about 65% of the $55.35M of supplied value, both running at 88% to 90% utilization with LTV parameters of 0.970 and 0.980. High LTV is defensible for stable-to-stable pairs and it leaves almost no cushion if either wrapper slips.

The larger concern is what sreUSD is. It is the yield-bearing stablecoin of Resupply, a protocol built on top of LlamaLend that borrows against Curve lending vault shares, and that was itself exploited in June 2025 for roughly $9.5M to $9.8M depending on the source, through an ERC-4626 donation attack that inflated a share price. So the biggest market on Curve's lender is collateralized by a token issued by one of its own borrowers, substantially funded by crvUSD the DAO printed for that purpose, and the failure mode that broke that borrower is the same class of bug that broke Curve's own sDOLA oracle nine months later. That is a loop, not a diversified book. On the governance side, veCRV concentration in Convex applies here too, because it is the same electorate that upgrades the lending factory.

Track record. Be precise about what belongs to this product. The July 2023 Vyper reentrancy exploit hit Curve pools. LlamaLend did not exist then; it soft-launched in March 2024 and DeFiLlama first listed it on 19 March 2024. That incident is not part of the lending record and I am not importing it.

What does belong: on 13 June 2024, founder Michael Egorov's oversized CRV positions were liquidated for roughly $140M and produced about $10M of bad debt, which he confirmed affected only the CRV market on Curve Lend, and which he repaid after selling 30M CRV. Credit where due, that one was made whole. Sixteen months later the same market broke again, and this time the shortfall has not been repaid. Add the March 2026 oracle exploit.

Separately, and distinct from any contract failure, Curve has been hit twice at the registrar level by what looks like the same failure mode. On 9 August 2022 attackers compromised the registrar iwantmyname, repointed the curve.fi DNS to a cloned site and took roughly $570,000. On 12 May 2025, again through iwantmyname, the curve.fi DNS was hijacked a second time and users lost roughly another $570,000 to signature phishing; a week earlier, on 5 May, Curve's X account was briefly taken over to post phishing links with no funds lost. The 2025 incident is what pushed the project off curve.fi and onto curve.finance. Reported loss figures for the two incidents drift between about $530,000 and $570,000 in secondary coverage. Contracts holding while the domain is hijacked is still users losing money, and it happened twice with the same registrar three years apart.

Worst case

You lend into a market whose collateral is exotic or whose oracle reads a manipulable exchange rate. Price moves, or someone moves the oracle inside one transaction, and the soft liquidation does not clear the position. The shortfall lands on you, not the borrower. Your deposit is marked at a fraction of its stated value, withdrawals are blocked because borrowed exceeds supplied, and your exit is to sell a discounted claim to a speculator betting on the collateral recovering. This is not hypothetical. It is what happened to CRV-long lenders in October 2025 and it is unresolved today.

A second path, and the one most specific to right now: you deposit into a book that is being retired. Incentives and attention move to v2, borrowers refinance elsewhere, the remaining lenders are the ones holding the least liquid markets, and the exit gets worse the longer you wait. A third: the DAO, whose electorate is concentrated in Convex, upgrades a factory implementation, or the Emergency DAO multisig changes a loan discount or a vault deposit cap with no timelock, and your risk changes after you deposited.

Worth stating clearly, because the mechanism is easy to misread. Soft liquidation protects the borrower against a cliff-edge liquidation and against the cascade that happens when many positions share one threshold. It buys time and can hand collateral back if price bounces. It does not make anyone whole. Every band crossed sells collateral low and rebuys higher, so a borrower who is soft-liquidated and fully recovers still ends up with less than they started with, by design. It needs arbitrageurs to trade through the bands, so a fast gap-down leaves the conversion incomplete, which is exactly how October 2025 generated bad debt. Hard liquidation still exists below the range. And none of it is lender protection: soft liquidation is a borrower-side feature that, when it fails, moves the tail risk onto the lender.

Bottom line

Avoid, for showcase money. The engine is good, the isolation model is real, and a sophisticated lender sitting in a conservative market is in a materially better position than this verdict suggests. The Ethereum WETH market runs 57% utilization and wstETH 49%, though at $0.74M and $1.94M supplied those are small books, not a deep haven.

What decides it is that the protocol does not underwrite market quality, and the people and processes that would have are leaving. Whoever deploys a V1 market picks the oracle. The outside reviewer who checked that work is gone and unreplaced. An emergency multisig can move loan discounts and deposit caps without a timelock. Borrowers were liquidated by an oracle bug five months ago, lenders from nine months ago still cannot get out, TVL is off two thirds from its peak, and Curve has said it is sunsetting this version. "Just pick a safe market" is not an instruction we can hand a reader who does not audit Vyper, and it is a worse instruction when the whole book is scheduled for retirement.

The path back is legible. If Curve appoints a credible replacement risk manager, resolves the CRV-long shortfall, and the DAO-gated v2 model becomes where the TVL actually sits, this is a different report about a different product, and it should be re-run against the v2 slug rather than this one.

Data appendix

  • TVL: $61.20M (DeFiLlama curve-llamalend, last chart point 2026-08-05 19:51 UTC). 30d $62.44M, 90d $70.22M (2026-05-08), 365d $133.94M (2025-08-06). ATH $182.70M on 2025-06-11, so 66.5% below peak and 54.3% below a year ago. Chain split: Ethereum $60.05M, Arbitrum $0.99M, Fraxtal $0.15M, Optimism $0.007M. Total borrowed $42.41M. Source: api.llama.fi/protocol/curve-llamalend.
  • Scope note: this headline is V1 only. DeFiLlama carries a separate curve-llamalend-v2 entry at $3.02M (verified this run) which our watchlist deliberately does not map. Both are children of parent#curve-finance alongside curve-dex at $1,287.8M, so LlamaLend is a sibling of the DEX, not a subset of it. The two TVL figures must never be added or conflated. An unrelated NFT-lending protocol holds the DeFiLlama slug llamalend at about $17k and is not this protocol.
  • Markets: 67 pools in DeFiLlama's /pools feed (Ethereum 38, Arbitrum 16, Fraxtal 7, OP Mainnet 6) and 31 markets in /lendBorrow totalling $55.35M supplied. Largest by supplied value: sreUSD $21.71M (87.9% utilization, LTV 0.970), sfrxUSD $14.27M (90.1%, LTV 0.980), WBTC $5.08M (63.3%, LTV 0.910), CRV $3.58M (82.9%, LTV 0.654), wstETH $1.94M (48.8%), sUSDe $1.82M (88.1%). Top two = 65.0% of supplied. Three markets show borrowed above supplied: sDOLA Ethereum ($55.8k / $62.4k), ARB Arbitrum ($31.4k / $49.7k), wFRAX Fraxtal ($15.7k / $161.6k). Sources: yields.llama.fi/pools, yields.llama.fi/lendBorrow, joined on pool id.
  • V1 sunset: no new V1 markets are being deployed and a sweep to deprecate legacy V1 markets ran during May and June 2026. The Curve DAO's 8 July 2026 call for a replacement risk team lists supporting "the launch of Llamalend V2 and the sunset of V1" in the mandate. V2 launched on Optimism 10 June 2026 with a 250,000 OP grant and three markets (ETH/wstETH, wstETH/USDC, WBTC/USDC) opening at zero borrow caps pending a seven-day DAO vote, and is now also live on Ethereum. Exact V1 shutdown timetable and any migration path for existing lenders: refresh.
  • Audits: StateMind, LlamaLend, published April 2024 before launch, 1 Critical / 1 High / 4 Medium, all reported fixed (statemindio/public-audits; circulated by Curve on X). MixBytes, Curve Lending, 6 Feb to 31 May 2024 with re-audits to 29 July 2024, scope over 30 contracts including OneWayLendingFactory.vy, TwoWayLendingFactory.vy, Vault.vy, Controller.vy, AMM.vy; 1 Critical / 2 High / 5 Medium / 7 Low, with stated focus on inflation attacks, oracle manipulation and factory vulnerabilities (mixbytes/audits_public). ChainSecurity published on Curve Stablecoin and is named in Curve's audit history; its coverage of the new AMM and Controller implementations for the crvUSD and LlamaLend factories comes from gov.curve.finance thread 10378, which returns 403 to automated fetch: refresh. Curve's own audit index (docs.curve.finance/references/audits) returned 404 on repeated attempts this run, so the firm list is assembled from the auditors' own publications rather than from Curve's index.
  • Admin / governance: V1 markets permissionless via OneWayLendingFactory; create_from_pool binds a Curve pool EMA oracle, create accepts a deployer-supplied custom oracle, which Curve's resources flag as an added trust assumption. Factory implementations upgradeable by admin (Curve DAO). Ownership actions require 30% quorum and 51% support plus an execution timelock after the vote; Curve's readthedocs describes a one-week ownership timelock while our curve-dex refresh recorded three days, so exact on-chain delay: refresh. Proposal #1252, passed 12 November 2025, granted an Emergency DAO multisig shared admin over crvUSD and LlamaLend risk parameters including debt ceilings, AMM fees, monetary policy, loan discounts and vault deposit caps, bypassing the vote and timelock. LlamaRisk's services engagement ran April 2025 to April 2026 and it was the named market curator for v2 at its June 2026 launch; the DAO published a two-week call for replacement risk teams on 8 July 2026 and no appointment is confirmed: refresh. veCRV concentration in Convex carried from our curve-dex refresh of 2026-08-05, not re-derived here.
  • Oracle exploit: 2 March 2026, 03:00:11 UTC, block 24566937, sDOLA/crvUSD (sDOLA-long2) market. CryptoFromPoolVaultWAgg (0x88822ee517bfe9a1b97bf200b0b6d3f356488ff2, Vyper 0.3.10, verified, not a proxy) used spot convertToAssets() as its feed. Flash loans of about 10M USDC and 15,986.11 WETH from Morpho Blue. 27 liquidation calls against borrower positions; Controller n_loans() fell from 30 to 4 across the block. Attacker take about $240,567 (6.74 WETH plus 227,326 DOLA). Curve's post-mortem reports about $10.9M of debt involved and about $822K of borrower equity seized, and describes the rate being moved up 13.79% (1.189 to 1.353) via an ERC-4626 donation path after a large LLAMMA exchange forced positions into soft liquidation; DarkNavy's independent analysis instead describes a mass redemption driving the rate down. Directional sequence contested, and the debt and equity figures rest on search-surfaced summaries because gov.curve.finance thread 11020 returns 403 to automated fetch: refresh. Market being deprecated because its oracle predates oracle proxies and cannot be swapped.
  • Bad debt: CRV-long market, originated in the 10 October 2025 crash, about $700,000 underbacked, lenders about 70% backed and unable to fully withdraw. Recovery proposal 27 April 2026 (tokenized claims sold into dedicated crvUSD and cvcrvUSD pools, about 1% fee, priced near 71% solvency; partial coverage from CRV about $0.957, full modeled recovery about $1.242), implementation underway as of 1 May 2026, no guarantee of recovery. Current recovery progress as of today: refresh. Prior event: 13 June 2024, founder liquidation of roughly $140M CRV caused about $10M of bad debt attributed to the CRV market on Curve Lend, repaid after a 30M CRV sale; some reporting puts LlamaLend's specific uncollateralized share at over $1M rather than the full $10M, so the split: refresh. Sources: coindesk.com 2026-04-27, gov.curve.finance thread 11062, cointelegraph.com and cryptobriefing 2024-06.
  • Holder concentration: top two markets 65.0% of supplied value. Largest collateral, sreUSD, is Resupply's yield-bearing stablecoin; Resupply is built on LlamaLend, borrows against Curve lending vault shares (cvcrvUSD), and was exploited in June 2025 for roughly $9.5M to $9.8M via an ERC-4626 donation attack on its own contracts, not Curve's. Curve DAO voted 24 October 2025 to mint and supply 5M crvUSD directly into the sreUSD market with an initial 5M borrow cap. Whether that 5M is still outstanding at full size and what share of the market's current $21.71M it represents: refresh. Depositor-level wallet concentration per market: refresh. Sources: news.curve.finance, quillaudits, halborn.com, blocksec.com.
  • Yield: base APY on crvUSD vaults about 0.7% to 3.3%; largest vault ($23.05M) shows apyBase 0.7424% against apyReward 1.1578%; 8 pools carry reward APY, representing $44.65M or 63.6% of pool TVL. Source: yields.llama.fi/pools.
  • Recent news scan (to 2026-08-06): 8 July 2026 DAO call for a replacement risk team, mandate includes the V1 sunset. June 2026 LlamaLend v2 on Optimism, later live on Ethereum. May to June 2026 deprecation sweep of legacy V1 markets. 27 April 2026 bad debt recovery proposal. 2 March 2026 sDOLA oracle exploit. 12 November 2025 Emergency DAO multisig (proposal #1252). 24 October 2025 5M crvUSD mint into sreUSD. 10 October 2025 crash bad debt. Front-end and DNS record, distinct from contract exploits: 9 August 2022 registrar compromise at iwantmyname, roughly $570,000 lost; 5 May 2025 X account takeover, no funds lost; 12 May 2025 second iwantmyname DNS hijack, roughly $570,000 lost, prompting migration from curve.fi to curve.finance. The July 2023 Vyper reentrancy exploit hit Curve pools and predates LlamaLend's March 2024 launch, so it is excluded from this product's record. No separate 2023 front-end loss for Curve was found; a claim to that effect did not survive checking.
  • External cross-reference: no second-opinion Curve LlamaLend report was located this run; this verdict is derived independently. Presence or absence of a Yearn Curation report: refresh.

Maintained monthly. Methodology: DeFi Research Instruction v2.

← all protocols