Klever: Marketplace settlement mints KLV when referral % + royalty % exceed the bid (negative seller share silently skipped)

Summary

When a marketplace order is settled (MarketBuy / BuyItNow, and auction Claim), the buyer's payment is split three ways — referral, royalties, and the seller (market-order owner) remainder:

marketOwnerAmount = CurrentBid − referralAmount − royaltiesAmount

Referral and royalties are paid out unconditionally, but the seller remainder is only paid when positive (computeMarketOwnerAmount returns Ok and pays nothing when the amount is <= 0). When referral% + royalty% exceeds 100% of the bid, marketOwnerAmount goes negative and is silently skipped — so the marketplace pays out more KLV / sale currency than the buyer paid in, minting the difference out of thin air.

The combined ceiling royalty% + referral% <= 100% is checked once, at listing time (Sell). But the two percentages are sourced asymmetrically at settlement:

• referral % is snapshotted into the order at Sell (MarketOrderData.ReferralPercentage);
• royalty % is never snapshotted — it is read live from the asset at buy time (asset.Royalties.MarketPercentage).

So the listing-time invariant is a time-of-check/time-of-use guarantee only. After a valid listing, the asset owner raises the royalty MarketPercentage via AssetTrigger → UpdateRoyalties; at the next buy the live royalty plus the snapshotted referral exceed 100%, and the settlement mints the overflow. The minted funds land in attacker-controlled referral / royalty addresses.

This was actively exploited on mainnet (see Evidence), minting tens of millions of KLV before the emergency guard was deployed.

Affected component

• Repository: klever-io/klever-go (node).
• Settlement / mint site: core/kapp/market/market.go — executeBuyMarket (L575+), computeReferralAmount (L361+), computeRoyaltiesAmount (L490+), computeRoyaltiesFixedDeposit (L443+), computeMarketOwnerAmount (L540+).
• TOCTOU sources: Sell combined check (market.go:908), order snapshot of referral but not royalty (market.go:997), live royalty mutation via core/kapp/kda/trigger.go — handleUpdateRoyaltiesNFTandSFT (L613+, sets asset.Royalties.MarketPercentage at L670).
• Reachable from both Buy (BuyItNow, market.go:204+) and auction Claim (market.go:705, market.go:731).
• Pre-fix: not gated by any fork flag — exploitable on mainnet. The fix is gated behind the new FixMarketBuyOverflow activation-epoch flag.

Root cause

1. Settlement pays referral + royalty unconditionally, seller remainder only if positive

core/kapp/market/market.go — executeBuyMarket (L575+):

referralAmount, _  := tools.ComputePercentageI64(marketOrder.CurrentBid,
                          int64(marketOrder.ReferralPercentage), ...)        // L583: SNAPSHOT referral %
royaltiesAmount, _ := tools.ComputePercentageI64(marketOrder.CurrentBid,
                          int64(asset.Royalties.MarketPercentage), ...)      // L587: LIVE royalty %
marketOwnerAmount := marketOrder.CurrentBid - referralAmount - royaltiesAmount  // L591: can go negative

// ---- FIX (FixMarketBuyOverflow), added by the patch ----
if m.forkController.FixMarketBuyOverflow() && marketOwnerAmount < 0 {         // L593-596
    ctx.Receipts().AddError(ctx.ContractID(), common.ErrFieldInvalidRoyalties, common.ErrInvalidValue.Error())
    return transaction.Transaction_AmountInvalid, common.ErrInvalidValue
}

m.computeReferralAmount(ctx, marketOrder, referralAmount, currencyID)   // pays referral in full
m.computeRoyaltiesFixedDeposit(ctx, marketOrder, asset)                 // pays fixed royalty (KLV)
m.computeRoyaltiesAmount(ctx, marketOrder, asset, currencyID, royaltiesAmount) // pays % royalty in full
m.computeMarketOwnerAmount(ctx, marketOrder, currencyID, marketOwnerAmount)    // <-- skips when <= 0

computeMarketOwnerAmount (L540-542) — the silent skip:

func (m *marketKapp) computeMarketOwnerAmount(... marketOwnerAmount int64) (... , error) {
	if marketOwnerAmount <= 0 {
		return transaction.Transaction_Ok, nil   // negative seller share dropped, NO error
	}
	// ... AddToBalance(marketOwnerAmount) ...
}

Meanwhile computeReferralAmount (L376) and computeRoyaltiesAmount (L515) each AddToBalance(...) the full computed amount with no matching debit from the buyer beyond the single bidderAcc.SubFromBalance(amount) taken in Buy (market.go:301).

Conservation breaks: buyer is debited bid once; recipients are credited referralAmount + royaltiesAmount. When that sum > bid, the surplus (referralAmount + royaltiesAmount − bid) is minted.

2. The combined ≤100% invariant is enforced only at listing time

Sell (market.go:908) correctly rejects a listing whose combined cut exceeds 100%:

if asset.Royalties.MarketPercentage + marketplace.ReferralPercentage > core.HundredPercent {
	return transaction.Transaction_ParameterInvalid, common.ErrInvalidValue
}

…and snapshots referral into the order, but not royalty (market.go:997-998):

marketOrder := &kapps.MarketOrderData{
	// ...
	ReferralPercentage:    marketplace.ReferralPercentage, // snapshotted
	RoyaltiesFixedDeposit: asset.Royalties.MarketFixed,    // snapshotted
	// NOTE: asset.Royalties.MarketPercentage is NOT snapshotted -> read live at buy
}

MarketOrderData has no field for the royalty percentage (kapps/market.pb.go), so settlement always re-reads it live from the (mutable) asset.

3. Royalty % is mutable after listing

core/kapp/kda/trigger.go — handleUpdateRoyaltiesNFTandSFT (L613+) lets the asset owner overwrite asset.Royalties.MarketPercentage (L670) with only a per-field <= 100% check (CheckValid100Params, L651) — it has no knowledge of any outstanding marketplace listing's snapshotted referral. So the owner can list at, e.g., referral 100% / royalty 0% (sum 100%, passes Sell), then raise royalty to 100%, making the buy-time sum 200%.

The shipped emergency-guard source documents this exact vector: "The royalty percentage is read live at buy time, so a listing made now can be weaponised later via UpdateRoyalties." (common/emergencyGuard.go)

Net effect: referralAmount + royaltiesAmount = bid + bid = 2·bid; marketOwnerAmount = −bid (skipped); bid KLV minted per settlement, paid to attacker-controlled addresses.

Proof of Concept

A. Committed regression test (deterministic, runnable today)

core/kapp/market/market_test.go — TestMarketKApp_ExecuteBuyMarket_RoyaltyReferralInflation. It builds an order with ReferralPercentage = 100% and an asset with MarketPercentage = 100% (the attacker is both the referral and the royalty address), then settles a bid of 25,600,000 KLV (25600000000000 base units):

go test ./core/kapp/market/ -run TestMarketKApp_ExecuteBuyMarket_RoyaltyReferralInflation -v

• FixDisabled_MintsKLVFromThinAir: settlement returns Ok; the attacker address ends with 2·bid credited while only bid was paid in — i.e. bid KLV minted.
• FixEnabled_RejectsInflation: with FixMarketBuyOverflow on, settlement returns Transaction_AmountInvalid and the attacker balance stays 0 — no payout runs.

B. End-to-end on a local node (the real attack path)

A single-node local network is sufficient. The exploit is four transactions from one ordinary funded account; nothing privileged is required.

1. Create an NFT collection you own, with royalties.marketPercentage = 0 and a royalties address you control.
2. Create a marketplace with referralPercentage = 10000 (100%) and a referral address you control (CreateMarketplace).
3. List one NFT for sale (Sell) on that marketplace. The Sell check passes because 0 (royalty) + 10000 (referral) = 10000 = HundredPercent. The order snapshots ReferralPercentage = 10000.
4. Raise the royalty on the asset to 100% (AssetTrigger / UpdateRoyalties, marketPercentage = 10000). Allowed: the per-field check passes and the live combined invariant is never re-evaluated against the open listing.
5. Buy the listing (MarketBuy) from a second account (or settle the auction via Claim). referralAmount = bid, royaltiesAmount = bid, marketOwnerAmount = −bid (skipped). Your referral + royalty addresses receive 2·bid; the buyer paid bid; bid KLV is minted.

Because the attacker controls buyer, seller, referral and royalty addresses, the only real cost is transaction fees; the cycle is repeatable until supply targets are met.

Evidence

Regression test (local, verbatim)

=== RUN   TestMarketKApp_ExecuteBuyMarket_RoyaltyReferralInflation
=== RUN   TestMarketKApp_ExecuteBuyMarket_RoyaltyReferralInflation/FixDisabled_MintsKLVFromThinAir
=== RUN   TestMarketKApp_ExecuteBuyMarket_RoyaltyReferralInflation/FixEnabled_RejectsInflation
--- PASS: TestMarketKApp_ExecuteBuyMarket_RoyaltyReferralInflation (0.00s)
    --- PASS: TestMarketKApp_ExecuteBuyMarket_RoyaltyReferralInflation/FixDisabled_MintsKLVFromThinAir (0.00s)
    --- PASS: TestMarketKApp_ExecuteBuyMarket_RoyaltyReferralInflation/FixEnabled_RejectsInflation (0.00s)
PASS
ok  	github.com/klever-io/klever-go/core/kapp/market	0.279s

FixDisabled asserts the attacker balance equals 2·bid = 51,200,000 KLV for a single settlement (bid = 25,600,000 KLV), with bid of that minted. FixEnabled asserts rejection and a 0 balance.

Mainnet exploitation (observed)

The bug was exploited in production, and was detected and characterised externally by the community monitoring project KleverPuls / kpulse.tech before the root cause was known internally. Over a ~24h window kpulse isolated a single wallet (opened 2026-06-04, ~204 transactions in ~24h, funded only by a ~242K KLV KuCoin withdrawal, no treasury/foundation funding) that:

• self-issued 3 NFT collections named "InflationPOC" and wash-traded one (NFLATION-ESGO/1) 29 times through self-created marketplaces — ~$450K of artificial, economically empty NFT volume;
• swapped the proceeds KLV → USDC / USDT / WBTC / WETH on KleverSwap and bridged ~$72K of value to Ethereum via wrapped-asset burns over 24h (USDC −12,487 ≈ $12.5K; USDT −26,362 ≈ $26.4K; WBTC −0.31 ≈ $19.6K; WETH −7.55 ≈ $14K);
• surfaced a spurious "1.86B KLV outflow" headline that kpulse correctly identified as a wash-trade receipt-doubling artifact with small real net KLV flow.

That "doubling of marketplace receipts" is precisely the on-chain signature of this bug: each abusive settlement pays out a referral cut (bid) plus a royalty cut (bid) while the buyer paid only bid once — the market contract emits ~2× the value it took in, which is the mint. The attacker's self-issued collection and self-created marketplaces are exactly the self-dealing setup the regression test reproduces (the test reuses the real on-chain identifiers: collectionID = "NFLATION-ESGO", asset 1, market name "Inflation Market").

kpulse could not determine the cause from on-chain data alone and flagged the activity for confirmation; the Klever core team then traced it to the referral+royalty settlement defect described above and shipped the emergency guard + protocol fix.

Each abusive settlement minted one bid of KLV; the observed bid was 25,600,000 KLV, repeated and funnelled through a short hop chain before being swapped and bridged. The emergency guard (common/emergencyGuard.go) blocks the following observed sender public keys (hex):

The single-bid per-settlement size (25.6M KLV) matches the "direct recipient, 25.6M" entry, and the ~125M at the collector hop is consistent with roughly five abusive settlements.

Impact

• Unbounded inflation of KLV (and of any sale currency used for the listing), repeatable for only transaction fees, by any account that creates its own collection + marketplace.
• The minted KLV is created by direct AddToBalance to attacker addresses (no tracked Mint), so the asset's booked supply does not change — the inflation is off the books and only detectable by summing balances / auditing receipts (it surfaces on-chain as doubled marketplace receipts).
• Realized impact (observed): the attacker minted KLV via ~29 self-dealt settlements, swapped to stable/wrapped assets on KleverSwap, and off-ramped ~$72K to Ethereum via the bridge (USDC, USDT, WBTC, WETH) before the emergency guard halted the activity, alongside ~$450K of artificial NFT wash-trade volume.
• Total loss of economic integrity for all KLV / token holders.

Who can exploit it / prerequisites

• Any account that can pay the one-time collection-create + marketplace-create fees and tx fees. No roles, admin, or allowlist.
• Any client. The settlement is triggered by standard MarketBuy / Claim contracts POSTed to the public, unauthenticated /transaction/send RPC (network/api/transaction/routes.go, SendTX / BroadcastTX). The operator CLI, the SDKs, or a hand-signed curl all work.
• Deterministic; the abusive state is reached with one extra UpdateRoyalties after a normal listing.

Remediation

Shipped as a layered response (embargoed):

Layer 0 — emergency guard (deployed first, fork-proof) — GHSA-p7gw rc1

common/emergencyGuard.go + data/transaction/emergencyGuard.go: matching transactions are kept out of blocks this node proposes (core/process/block/preprocess/transactions.go) and refused at the node API (node.go SendTransaction / SendBulkTransactions). It never changes block validity, so a partial-fleet rollout cannot fork the chain. It blocks the known attacker senders (all contract types) plus all MarketBuy, Sell, and CreateMarketplace / ConfigMarketplace operations while the protocol fix rolls out. Enforcement is by proposer cooperation, not protocol — coverage equals the share of block producers running the guard.

Layer 1 — protocol fix (consensus, epoch-gated) — GHSA-p7gw rc2

core/kapp/market/market.go:593 rejects the settlement when marketOwnerAmount < 0, before any payout runs, returning Transaction_AmountInvalid. Gated behind the new FixMarketBuyOverflow activation-epoch flag (config/enableEpochs.*, core/fork/forks.go, core/interface.go) so historical blocks reprocess identically. Covered by TestMarketKApp_ExecuteBuyMarket_RoyaltyReferralInflation.

Recommended hardening (defense in depth)

1. Snapshot the royalty % into MarketOrderData at Sell (as referral already is) and pay from the snapshot, eliminating the TOCTOU entirely; or re-evaluate the combined referral% + royalty% <= 100% invariant at settlement.
2. Treat a negative seller remainder as a hard error everywhere, and only treat exactly 0 as a no-op skip (computeMarketOwnerAmount), so a future regression aborts the tx instead of minting.
3. After splitting a payment pool, assert conservation (referral + royalties + ownerShare == bid) so any drift aborts the transaction.

Notes

• Triggered by both BuyItNow (Buy) and auction settlement (Claim).
• The same family of "pay full cut, silently drop the negative remainder" minting also exists in the royalty-split paths and is tracked separately under GHSA-cgc5-v3f2-8m2v (split-royalty uint32 overflow). This advisory covers the top-level referral+royalty > bid case; the FixMarketBuyOverflow guard here only checks marketOwnerAmount, not intra-split over-payments.

Acknowledgments

• KleverPuls / kpulse.tech — community monitoring project that first detected and characterised the exploitation in the wild. kpulse isolated the attacker wallet and its "InflationPOC" collections, identified the marketplace wash-trading of NFLATION-ESGO/1 and the KleverSwap → bridge off-ramp (~$72K to Ethereum), and flagged the anomalous doubling of marketplace receipts — the exact on-chain signature of this bug — prompting the incident response that led to this fix. The root cause was then identified and remediated by the Klever core team.

Source

• Vulnerable / fixed code: core/kapp/market/market.go:540-545,575-596,908,997-998, core/kapp/kda/trigger.go:613-676, core/process/kda/assetHelper.go:101, tools/converters.go:102, core/constants.go:18.
• Emergency guard: common/emergencyGuard.go, data/transaction/emergencyGuard.go, core/process/block/preprocess/transactions.go, node/node.go.
• Fork flag: config/enableEpochs.go, config/node/enableEpochs.yaml, core/fork/forks.go, core/interface.go.
• Regression test: core/kapp/market/market_test.go (TestMarketKApp_ExecuteBuyMarket_RoyaltyReferralInflation).