DefiingerMulti-Chain DeFi Data, News & Research

Weekly Deep Dive: How a Range-Proof Cache Drained 95% of Liquid Network's Reserve

How roughly 3,996 BTC left Liquid Network's federation reserve without a single key being compromised: a range-proof caching flaw, a redemption flow that checked procedure rather than provenance, and a merged fix that had not shipped.

Weekly Deep Dive2026-09-079 min readDefiinger Research Desk1895 words

This week's report covers the largest single incident in crypto so far this year by funds moved: about 3,996 BTC leaving the Liquid Network federation reserve on 6 September 2026.

We chose it not only for its size but because it fails in an unfamiliar direction. There was no broken contract, no compromised key and no defeated multisig. The tokens were created upstream, and every system that later handled them behaved correctly.

1 1. In brief

On 6 September 2026, about 3,996 BTC left the Liquid Network federation wallet — roughly $320m at prevailing prices, and close to 95% of the reserve backing every L-BTC in circulation. No key was stolen. No multisig was defeated. The attacker minted tokens that should not have existed and then walked them through the official exit.

That sequence is the reason this report exists. Everything downstream of the mint behaved correctly: SideSwap processed a normal redemption, the federation validated a valid authorisation, and 11 of 15 members signed a legitimate withdrawal. The failure was a caching optimisation in the platform's core software, and it sat in a tagged release while a merged fix waited to ship.

~3,996 BTC
Reserve withdrawn
about $320m
~95%
Of the federation reserve
4,200 BTC to about 200 BTC
0
Keys compromised
per Blockstream
23 min
Mint request to payout
14:05 to 14:28 UTC

2 2. Background: what Liquid is, and what the reserve is for

Liquid launched in 2018 as a federated Bitcoin sidechain built by Blockstream. A user locks BTC in a wallet controlled by the federation and receives L-BTC on the sidechain. Burning L-BTC redeems an equivalent amount of mainnet Bitcoin from that same wallet. The arrangement buys confidential transaction amounts and fast settlement at the cost of a custody assumption: the peg is only as good as the federation's willingness and ability to honour it.

Two details matter for what follows. First, L-BTC is fungible and its amounts are hidden, so no observer can tell from the chain how much of it is properly backed. Second, other assets issued on Liquid — stablecoins, real-world-asset tokens — do not draw on the federation's Bitcoin. Only L-BTC holders were exposed to the reserve.

Structural differences that shape where risk sits. Liquid trades a verifiable settlement layer for confidentiality and speed, and the trade is carried by the federation's software, not by cryptography.
PropertyLiquid NetworkBitcoin mainnet
ConsensusFederated functionary setProof of work, permissionless
Who can validateFederation members onlyAnyone
Reserve controlMultisig across functionariesNone — self-custody
Amount visibilityConfidential transactionsAmounts public
Exit mechanismPeg-out authorised by federationNo exit needed

None of this made the incident inevitable. It did mean that a software bug could reach the reserve in a way that a bug on mainnet cannot: the chain does not independently verify that its own money supply corresponds to what the federation holds.

3 3. Mechanism: five steps, none of them exotic

Reconstructing the sequence from Blockstream's statement, SideSwap's account and independent technical analysis gives a path with no step that required breaking anything.

The attack path. A small transfer of about 2.5 BTC preceded the main withdrawal by roughly 27 minutes, consistent with a test before the full amount.
#StepWhat happenedWhy it passed
1MintMore than 4,000 L-BTC created on Liquid with no BTC locked behind themConfidential amounts hid the inflation
2TransferThe tokens moved to a SideSwap redemption addressIndistinguishable from legitimately issued L-BTC
3Burn4,000 L-BTC destroyed at 14:05 UTC to request a peg-outThe burn was valid; the tokens existed
4AuthoriseA peg-out authorisation was issued and 11 of 15 members signedThe process checked the process, not the origin
5PayAbout 3,996 BTC released at 14:28 UTC, less roughly 0.1% in feesFederation keys were used correctly

The step that fails safely on a well-designed bridge is step 4. A peg-out authority can check that tokens were burned; it rarely checks that the tokens were ever correctly minted, because until this week that seemed to be guaranteed upstream.

4 4. Root cause: a cache key that forgot what it was caching

Confidential transactions hide amounts while still proving validity. Liquid implements them through range proofs: cryptographic evidence that a hidden value is non-negative and within bounds. Verifying range proofs is expensive, so Elements caches verification results.

The reported defect is in how that cache was keyed. A cache key must bind every input that affects the answer. This one appears to have omitted the asset commitment and the output script, meaning a proof already accepted for one asset and destination could satisfy a different asset and a different destination.

cache_key = hash(range_proof) # what it actually keyed on
cache_key_required = hash(range_proof || asset_commitment || output_script)

In plain terms, the system remembered that a proof was valid and stopped asking what it was valid for. Reusing that answer let the attacker create L-BTC that passed verification without any Bitcoin behind it.

The optimisation was not wrong in principle. Caching range proofs is reasonable. The failure was in the key: performance work moved the security boundary without anyone re-examining what the guarantee actually depended on.

5 5. Three failures that all had to line up

It is tempting to file this under 'a bug was exploited'. That understates it. Three separate controls were available and none of them stopped the payout.

Independent controls are only independent if they fail independently. Here all three depended on the same assumption: that L-BTC in circulation equal BTC held.
LayerControl that existedWhy it did not fireCost of preventing it
CodeRange-proof verificationCache returned a stale verdict for a new contextCorrect cache key would have cost nothing at scale
ProcessFederation multisig reviewEvery signature was legitimate; no member was asked about provenanceSupply reconciliation check before release
ReleaseFix existed and was mergedNot included in the tagged build running at the timeBackporting the merge into the release

The third row deserves separate attention. Reporting indicates the fix was written over the summer and merged early this month, but was absent from the release running during the incident. A patch that exists is not a patch that protects anyone.

6 6. Data: the numbers, and what each one measures

This incident produced several quantities at once. They are different measurements, not competing estimates, and quoting any of them as 'the loss' without saying which is how reporting on it goes wrong.

All figures are provisional pending Blockstream's post-mortem. The BTC had not been returned at the time of writing.
QuantityValueWhat it measuresSource
L-BTC sent to redemption4,000 L-BTCTokens burned to request the peg-outSideSwap
BTC actually paid out~3,996 BTCPaid after roughly 0.1% in feesSideSwap
Value at request time~$320mQuoted at prevailing BTC priceMultiple
Reserve before~4,200 BTCFederation wallet balance before transferBlockstream
Reserve after~197-207 BTCSources differ slightly on the exact remainderCoinPost, others
Share of reserve taken~95%Withdrawal relative to prior balanceDerived
Weekly total crypto drains~$322m1-7 September across all reported protocol incidentsCoinPost Terminal
Liquid share of that total>99%One incident dominating the periodDerived

The derived items are worth stating plainly. If about 3,996 BTC left a reserve of about 4,200 BTC, roughly 95% of the backing went with it, and every remaining L-BTC is backed by about five cents of Bitcoin if no recovery occurs.

backing_ratio = remaining_BTC / L-BTC_outstanding
backing_ratio = ~200 / ~4,200 = ~0.05, about 5%

That ratio, not the dollar figure, is the operational exposure for anyone holding L-BTC.

7 7. The same shape, twice more in six days

Liquid was the largest instance of a pattern that appeared three times in this reporting window. In each case value was created upstream and accepted downstream by systems behaving as designed.

Three manifestations of one construction, separated by four orders of magnitude in realised loss.
DateVenueHow value was createdRealised lossContained by
31 AugankrFLOW / FlowAbout 8.6m unbacked ankrFLOW minted and used as collateral~$410kFlow Foundation reimbursement
5 SepRocketSelf-trades at inflated prices generated artificial profit~$287kPlatform pause, then socialisation
6 SepLiquid NetworkUnbacked L-BTC minted through a cache flaw and redeemed for BTC~$320mBridge nodes off, trading paused

Scale apart, the resembly is in the failure to ask provenance questions. Collateral checks verified that tokens existed and were transferable; redemption flows verified that tokens were burned. None asked whether they were legitimately created.

8 8. Impact assessment: four layers, different speeds

Immediate effects differ sharply by layer. The most concentrated loss sits with L-BTC holders; the broadest sits with every bridge whose security model resembles Liquid's.
LayerWho is affectedImmediate effectResolution depends on
L-BTC holdersAnyone holding the pegged assetBacking collapsed to about 5%Whether the attacker returns funds
Liquid users generallyStablecoin and RWA token holders on the chainUnaffected at the ledger levelNothing — these do not draw on the reserve
Exchanges and membersFederation participants and their customersDeposits and withdrawals haltedPatch deployment and network restart
Sidechain categoryOther federated and committee-run bridgesReputational re-pricingPost-mortem quality across the category

The clean separation between layers is the most useful thing here. Assets that never relied on the federation's Bitcoin were not impaired by its emptiness, which is a reminder that 'exposure to a chain' is not one risk but several, depending on which asset you hold on it.

9 9. The claim of white-hat status

The attacker left an on-chain message reading 'we are whitehats. contact us on chain', and subsequent exchanges included a request that the bug be fixed and nodes patched, and a question about whether returning most of the funds would be acceptable. Some reports describe a 20% bounty request; that figure has not been confirmed by either party.

Blockstream's language has stayed at 'purported white-hat'. That caution is justified. Responsible disclosure is a process agreed in advance, not a label applied after funds have left. Roughly 95% of a reserve was moved without authorisation, regardless of intent, and the only commitments offered so far are messages.

Intent is not custody. Until the coins move back under terms the users can verify, the prudent assumption for anyone holding L-BTC is that the funds are gone.

10 10. Risk matrix

Likelihood reflects the position described in public reporting; it changes with Blockstream's post-mortem.
RiskLikelihood nowSeverityWhat would reduce it
Permanent loss of the reserveUnresolvedSevereVerifiable return before network restart
L-BTC trading far below pegHigh while unresolvedSevere for holdersReserve restoration or a stated compensation plan
Restart before all nodes patchMediumSevereAttested patch deployment by federation members
Same flaw in Elements forksUnverifiedHighIndependent audit of the fix across downstream projects
Contagion to L-BTC integrationsMediumModerateIntegrations disabling deposits until backing is restored

11 11. What this changes for due diligence

  • For bridged and pegged assets: ask what verifies that supply was legitimately minted, not just that it exists and transfers.
  • For committee-run systems: ask who can pause a withdrawal and whether any participant is required to reconcile total supply against total reserves before release.
  • For any dependant product: check whether your asset's backing is the same pool someone else's redemption draws from.
  • For performance-critical code paths: treat caching layers as security boundaries and review what the cache key binds.
  • For release processes: ask whether merged fixes are in the running tagged build, not whether they are in the repository.

12 12. What would change our read

Our assessment rests on the account available at publication. It would change materially if Blockstream's post-mortem attributes the incident to a different cause than a cache-key defect, if the fix proves to have been present in the running release, or if the funds are returned under verifiable terms. Any of those would move specific sections above. Until the post-mortem is published, every figure quoted here should be treated as provisional.

The broader conclusion does not depend on those details. Three incidents in seven days turned on inputs that were never checked for provenance, and the largest of them nearly exhausted a reserve because every downstream control verified procedure rather than substance.

DE
Defiinger Research Desk

The Defiinger Research Desk compiles multi-chain DeFi data and commentary from public on-chain sources and vetted industry publishers. Our editorial process prioritizes verifiable figures and clearly dated references.

Sources & Methodology

  1. Blockstream — statement on the 6 September incident, key compromise findings and Elements flaw attribution.
  2. SideSwap — peg-out timeline, including the 14:05 UTC request and 14:28 UTC settlement.
  3. CoinPost Terminal — weekly protocol-drain totals for 1-7 September and confirmations of the smaller incidents.
  4. SlowMist — classification of the Rocket perpetual-market manipulation.
  5. Flow — the ankrFLOW minting incident and reimbursement commitment.
  6. crypto.news — reporting on the Rocket perpetual-market manipulation and the platform's response.
  7. CN-SEC weekly review (issue 25) — incident inventory for 31 August to 6 September 2026.
  8. DefiLlama — chain-level data used for the same-week context.

Headlines and figures on this page are drawn from the outlets listed above; commentary is clearly labelled opinion and is not investment advice. Last reviewed 2026-09-07.

Frequently Asked Questions

Was Bitcoin itself affected?
No. Bitcoin's protocol and network were unaffected. The incident was confined to Liquid's sidechain implementation and its federation reserve.
Were any keys compromised?
Blockstream stated that the peg-out authorisation key and all other federation keys remained secure. No signatures were forged; every member who signed did so legitimately.
How much was taken?
About 3,996 BTC, roughly $320m, from a reserve of approximately 4,200 BTC — around 95%. Sources differ slightly on the exact remaining balance.
Are other assets on Liquid safe?
Tokens that do not draw on the federation's Bitcoin are not directly impaired by the missing reserve, because the chain's ledger entries for them are unaffected. L-BTC holders carry the direct exposure.
Is the money coming back?
The party behind the withdrawal described itself as a white-hat and discussed returning funds after a patch, with an unconfirmed bounty request reported. Nothing has been returned, and no verifiable arrangement has been published.
What was the underlying bug?
A range-proof verification cache in Elements whose key did not bind the asset commitment and output script, letting an accepted proof be reused for a different asset and destination. A fix was reportedly merged days earlier but absent from the running release.
How does this compare with last week's incidents?
Last week's three exploits manipulated the price of thin collateral. This one created the asset itself. The difference matters: no amount of collateral diligence catches an upstream mint.
Should I avoid federated sidechains?
We do not make recommendations. The useful takeaway is understanding which asset you hold and what backs it — on any chain, an asset's exposure depends on the specific reserve it draws on.