Initia Exploit: Script Caching Bug Enables Multimillion-Dollar Asset Drain
_A published local replay provides evidence of the incident: The Initia virtual machine granted approval to a script, and continued to honor that approval even after the underlying code was swapped out.
On October 23, a vulnerability in script caching allowed an attacker to credit themselves with unbacked assets, avoiding any increase in the affected tokens’ total supply.

In the replay, this stale approval permitted an attacker-created object to enter a deposit function as if it were an authentic token.
The attacker subsequently withdrew these balances through cross-chain bridges, receiving genuine assets on the destination networks: 1.4 million AUSD and 105 ETH were released via LayerZero, and approximately 1.08 million USDC through Noble and Circle’s CCTP.
The exploit started at 14:59 UTC. At 15:41:30 UTC, Initia suspended its L1 and rollup chains. Five seconds later, the attacker moved 1,135.52 ETH into a new Ethereum address.
Disabling the source chain did not reverse already-completed payouts on external networks.
According to Bitquery, the face value of the forged balances was $7.48 million, with about $2.73 million successfully withdrawn. Some of the generated value never left via bridges.
The attacker extracted 1.08 million USDC from Noble’s reserves, which backed Initia’s USDC. Other users quickly withdrew what remained, leaving only $63 by 15:37 UTC.
As the reserves emptied, routine withdrawals began to fail.
The initial flaw was located in Initia’s VM, while cross-chain bridges converted false balances into actual asset payouts.
When a bridge processes a withdrawal, who ensures the underlying balance is properly backed?
Credit: Canzi, Initia, Sprunky, D2 Finance, Bitquery, Spreek, Zon, Agora, Cryptopolitan, CoinStats, KuCoin, Binance, LayerZero, Code4rena, Zellic, Hexens, Liquid Network, Initia Foundation
Spreek publicly flagged the exploit 28 minutes after the first forged transaction.
By that time, around $2.73 million had already been drained.
At 15:27 UTC, Spreek estimated $2.6 million had been cashed out “thus far with more in flight.”
Bitquery’s later analysis shows an 80,000 USDC withdrawal at 15:24, three minutes before Spreek’s message.
Initia’s formal update recorded the halt of L1 and rollups at 15:41:30 UTC, about fourteen minutes after the public alert.
The statement noted the team was “working to establish the full scope of any impact with security teams, validators, and exchanges.” There was no mention of loss amounts, root cause, or a date for resuming operations.
Co-founder Zon provided additional context from his personal account: "To the best of our knowledge so far, this is a Day 0 bug related to Initia's MoveVM on CosmosSDK that allowed the exploiter to mint additional tokens on the Initia L1 but does not apply to native INIT tokens," he stated, estimating the exploit at about $2.7 million in AUSD, ETH, and USDC minted and bridged before the halt.
Initia’s official account confirmed the estimated $2.7 million loss the next day.
D2 Finance tracked the bridge payouts: 1.4 million AUSD was bridged to Ethereum, followed by a separate withdrawal of 105 ETH. According to D2, these actions depleted 71% of the AUSD adapter’s reserves and 96% of the native-ETH adapter’s.
Agora responded on October 23, announcing it would freeze hacker-associated AUSD on networks where its contracts are active, clarifying that it had no smart contracts on Initia itself.
This distinction separated Agora’s native assets and contracts from the bridged version on Initia, setting the boundaries of its intervention.
Agora used the batchFreeze function to freeze the attacker’s Ethereum address, but by the time the freeze was executed, the 1.4 million AUSD had already been swapped for ETH. With no AUSD left in the address when the freeze was applied, the measure could not reverse the payout or recover those funds.
Early reports suggested that the attacker had minted new tokens.
Spreek described the process as “infiniminted,” and Zon noted the vulnerability allowed the attacker to “mint additional tokens.”
However, the underlying mechanism was more complex.
Initia’s detailed explanation was published at 08:46 UTC on October 24, roughly 17 hours after the network halt, using the term “Spoofed.”
The attacker "artificially spoofed the MoveVM state," claiming unbacked balances in WETH, iUSD, and USDC, without altering the total supply for those tokens. No native INIT tokens were created.
The key issue was not the ability to unlock real value—bridge payouts had already demonstrated that—but rather the process by which these unbacked balances were established.
Bridge contracts processed withdrawals based on credits that had no legitimate issuance on the ledger.
How did data crafted by the attacker become valid spendable balances within the system?
Checked Once, Trusted Twice
Move’s type system was designed to prevent attacker-created objects from passing as real FungibleAssets, which represent tokens in Move.
The VM is intended to verify a script’s types before running it, rejecting code if the types don’t match.
Initia leverages a MoveVM derived from Aptos, implemented in Rust and linked to the Cosmos SDK through Go bindings.
In its October 24 statement, Initia attributed the exploit to a MoveVM script caching flaw, which allowed malicious code to sidestep type and safety checks. The attacker manipulated VM state to obtain unbacked balances, without affecting the official supply.
The announcement did not point to a specific code file, commit, or function.
0xCanzi, a security researcher, published a local replay using real Move VM code, modeling transaction rollback locally to illustrate a possible attack vector.
While this is a demonstration of a plausible method, it does not definitively establish the exact steps taken in the live attack.
In the reconstruction, a script was validated once, then reused after the underlying module had changed.
Move scripts are code segments submitted with transactions. In 0xCanzi’s replay, the attacker first published a contract module named Router, whose swap function was declared to return a FungibleAsset but aborted when invoked.
Another script would generate a metadata object from a currency’s address, call the swap, and pass the result to the deposit function. The types matched, so the VM verified and cached the result.
The Router module’s first version then aborted, rolling back the publication, but the cached verification for the script remained.
Next, the Router module was republished at the same address, with swap returning an attacker-constructed object called Quote, embedding currency metadata and a claimed amount.
The attacker reran the same script. The VM relied on the previous verification, not recognizing the updated module’s changed return type.
Consequently, the Quote object was accepted by the deposit function under the incorrect type assumption. In the replay, deposit credited the attacker with 3 million units and emitted a DepositEvent.
A newly-initialized VM without the cached approval would reject the same sequence with a TYPE_MISMATCH error.
The $3 million figure in the replay is illustrative, not the actual sum lost.
The replay demonstrates the deposit function crediting the claimed balance, giving insight into how balances increased without adjustments to the tokens’ official supply, as described by Initia.
The official statement did not confirm that the real-world attack reproduced these exact steps.
After the VM accepted forged balances, subsequent checks failed to detect the issue.
Burn Here, Pay There
The attacker’s Ethereum gas funds arrived more than six hours before the exploit.
At 08:26:59 UTC on October 23, a 0.1 ETH withdrawal from Tornado Cash sent about 0.098 ETH to the Ethereum address that later received bridge payouts.
Sprunky identified the transaction relayer and the submitter.
Tornado Cash Gas Funding Transaction:
0x3fb58feaaff3abdc5d55e3c6128ea03b6527665d702c983630f7f1d2f26c4d49
Attacker EOA on Ethereum:
0xDD922c841d85E12dac2dc2eef5a5E29cFFf6edFA
Starting at 14:59:17 UTC, a single Initia account generated unbacked balances in four rapid transactions. Bitquery calculated these as 3 million iUSD, 2 million AUSD0, 2 million USDC, and 200 WETH, with a face value of $7.48 million.
Bitquery published one of the transaction hashes.
Spreek posted the same transaction as a “sample fake mint” from Initia’s explorer.
Initia Account Used for Forged Balances:
init1mkfzepqashsjmtpdcth0tf0znnlldm06a5m89u
Initia L1 Forged-Balance Transactions:
- 14:59:17 UTC: 712111A48BA4D7F830133BCDE2B2E0ECED1962CED9D8CE118381CFBD1F1C14F5
- 14:59:21 UTC: 7F0A72758F661A5D7C057B9DEFBBB9F8E0D769CA91DD478C677D9D8FF3720EEF
- 14:59:26 UTC: 67423E84815CEC63632BD4676CA22D33567F3E02C3187281C9B6BFD593CBA7AF
- 14:59:30 UTC: 3F165329DE9C4415C58C76FDA9F7F6E7CAA5F76159F287B109812ECADD0919A3
AUSD0 represents Initia’s bridged version of Ethereum-based AUSD, transferred through LayerZero.
The attacker sent the unbacked AUSD0 from Initia to Ethereum via LayerZero, and the Ethereum bridge contract released actual AUSD from reserves.
At 15:01:59 UTC, the bridge contract paid out 1.4 million AUSD to the attacker.
LayerZero AUSD Release (1.4M AUSD):
0x034c1076a08091b3221ef1e7bbcfb47dfeda93a6e460888f2d1bc124be17f3b6
Ethereum Bridge Contract Holding AUSD for Initia:
0x38b8b7ca28e672018a299b20565aca8cf200c844
At 15:03:23 UTC, the attacker swapped the AUSD for 581.55 ETH using 1inch.
AUSD-to-ETH Swap via 1inch (581.55 ETH):
0xeb5e85e76f00a7a84dfe4d2d6aab146a769785b3d80307c67dde14c79f36c5ae
For USDC, Bitquery tracked 1 million USDC withdrawn from Noble’s reserve and bridged to Ethereum via Circle’s CCTP.
Noble USDC Reserve for Initia:
noble1j8rrc5epz0wppvl7jpy9p7q0u4ky9dcegaw3cu
At 15:04:59 UTC, the attacker routed about 1 million USDC through 1inch, receiving 415.70 ETH. The swap involved Uniswap V3 and included converting some USDC to USDT. Bitquery’s analysis links this to the initial 1 million USDC withdrawal.
Second 1inch Swap, USDC and USDT (415.70 ETH):
0x6981eed05757cfbce5ac7818449b28132dcb400361029ed3452037936c66c3b0
At 15:06:47 UTC, the native-ETH bridge paid out 105 ETH, with no swap needed.
Native-ETH Bridge Release (105 ETH):
0x5cd422fd8166abf5d8632e9919bfedf6e26f5b77cf732cc797d863450887d7e9
Ethereum Bridge Contract Holding ETH for Initia:
0x66a503a1060ab3f2b1aaabed613fe30babbc1bde
A subsequent request for another 1 million USDC at 15:07 failed; the reserve was insufficient.
The final successful withdrawal occurred at 15:24 UTC: 80,000 USDC from the Noble reserve, which was eventually swapped for about 33.18 ETH.
Initia confirmed its network halt at 15:41:30 UTC. Five seconds later, 1,135.52 ETH, valued at about $2.72 million, was consolidated into a new wallet.
Consolidation Transaction (1,135.52 ETH):
0x60f15b7e96111f11af54a2a0c9adc2ff33fef6b6a3c1a3b073a55c1421043e06
Consolidation Wallet:
0x358eF0130f7A5aA516dCd81Bdc80B18CdE41CAC8
Bitquery’s summary matches the consolidated ETH: 581.56 ETH from AUSD, 415.70 ETH from USDC, 105 ETH from a bridge, and 33.18 ETH from the 80,000 USDC withdrawal, plus roughly 0.098 ETH from initial funding, with some spent as fees.
Agora’s freeze was executed at 16:31:59 UTC in block 26,256,539, 1 hour and 28 minutes after the AUSD-to-ETH swap. At that moment, the attacker’s AUSD balance was zero.
Agora batchFreeze Transaction:
0x9df56874a61efc7c84aecc9263a384578d70ec28734990cb74febc3003b794db
batchFreeze Caller:
0xcF7D2a525057555d7b4816941185b7ae10E94681
AUSD Token Contract:
0x00000000eFE302BEAA2b3e6e1b18d08D69a9012a
Bitquery noted that the Ethereum AUSD bridge blocked a further 200,000 AUSD payout at 17:51 UTC. Four out of five verifiers signed, but LayerZero Labs’ verifier did not.
Initia’s October 24 statement identified the consolidation wallet.
Since then, funds have begun to move. On October 27, the consolidation wallet sent two 200 ETH transactions to the same recipient address.
First Outgoing Transfer from the Consolidation Wallet - 200 ETH:
0xf35dcdd4c23816adcd5e1fb6578a75c2d43b4816bf78203b6721cfb59b9cc7f1
Second Outgoing Transfer from the Consolidation Wallet - 200 ETH:
0x383a55d8caf78c926a5022b8b7a386d1cc31369ca095cdb6b2dc093703b3f038
Recipient of Both Transfers:
0xA764b87FCB14bC850B68DfE41720Cba3b24a89C0
The recipient subsequently distributed ETH to multiple addresses, and made a 100 ETH deposit to Tornado Cash.
The attacker had managed to transform counterfeit balances on Initia into tangible assets on Ethereum. Agora’s freeze occurred after the proceeds had been converted to ETH and the consolidation wallet later started dispersing the funds.
On Initia, Bitquery calculated that $4.55 million in unbacked balances remained when the chain was frozen, based on face value.
These were not assets already withdrawn, but forged on-chain balances, raising questions about their treatment before any restart.
If the network resumes, what will happen to these unbacked balances?
Halted, Not Settled
Recovery efforts for Initia had moved into the auditing phase, with no relaunch scheduled as of reporting.
On October 27, the team reported that a security patch had been drafted and was under review by security partners.
Upon completion of these audits, the team planned to detail steps for restoring asset backing and restarting L1 and rollups in a controlled process.
The team anticipated completion “in the coming days,” but provided no firm date.
On October 28, Initia confirmed that asset backing would be fully restored with support from the Initia Foundation when the chain resumes.
The Foundation authorized funding to restore 1:1 backing for affected Noble USDC, WETH, and AUSD0 (used as iUSD collateral). This commitment covers confirmed deficits up to 1.08 million USDC, 105 ETH, and 1.4 million AUSD0, subject to reconciliation and final approval.
This announcement formalized a commitment to replenishment, not completed transactions.

Exchanges took action independently. KuCoin suspended INIT deposits and withdrawals on October 24.
On October 26, Upbit and Bithumb flagged INIT as a trading-caution asset.
[INIT’s price fell to around $0.115 on October 23, rebounded above $0.15, and recovered to $0.24 by October 28](https://www.kucoin.com/news/flash/initia-halts-l1-and-rollups-after-2-7m-mint-exploit; https://www.coingecko.com/en/coins/initia).
The more direct loss occurred within the reserves that honored withdrawals against forged credits.
The Ethereum contract storing AUSD for Initia lost 1.4 million AUSD, and the ETH contract lost 105 ETH—representing 71% and 96% of their holdings respectively, per D2’s figures.
The extent of holders’ exposure depended on the affected token and its collateral arrangements, as the AUSD, ETH, and USDC reserves suffered varying degrees of depletion.
Separately, Bitquery reported $4.55 million in remaining forged balances on Initia at the time of the halt, including 3 million iUSD. Two failed attempts were made to swap some of this iUSD for AUSD0.
Face value does not equal realizable value, and not all forged assets were necessarily transferable or redeemable.
The Foundation committed to funding specified shortfalls, subject to reconciliation and approval. Details on treatment of the remaining forged balances were not disclosed.
Timing for the restart, resolution of unbacked balances, and completion of asset replenishment remain open issues.
Zon characterized the incident as a “Day 0 bug.” As of October 28, Initia’s movevm repository had no public security advisories.
The VM had undergone external review. Code4rena’s January 2025 Initia Move contest reviewed code changes to movevm between two versions, focusing on the JSON interface and IBC bridge hook.
That same brief references a prior Zellic audit of Initia, which recorded 34 findings, including two critical issues. The review scope included the movevm codebase and its differences from upstream Aptos.
This does not confirm whether the specific script cache flaw was tested.
The published local replay illustrates a vulnerability involving cached type verification, transaction rollback, and module updates, modeled on a local instance.
The replay suggests a targeted test approach: Compare script execution in a VM with cached verification to execution in a fresh VM after rollback and module replacement. The cached approval permitted the forged object, while the new VM raised a TYPE_MISMATCH error.
Such a test could highlight stale-verification mismatches as seen in this case.
Hexens documented a related, though distinct, stale type-tag cache vulnerability in Aptos’s Move VM, involving partial cache flushes.
This issue differs in cache mechanism from the script-verification bug in Initia, but both highlight risks of inconsistent cache states.
Restoration of asset backing and network operations are linked but separate issues: Fixing the VM does not recover funds already withdrawn.
The Foundation has committed to restoring the affected asset backing; delivery depends on reconciliation and approval processes.
When does a funding commitment become finalized backing, so holders can redeem their assets?
The Initia incident underscores a dangerous pattern: Relying on an outdated approval without validating its context.
As demonstrated in the local replay, a virtual machine reused a cached verification after the underlying code had changed.
Meanwhile, on-chain evidence shows that bridge exits allowed withdrawal of real assets based on forged balances.
LayerZero processed the AUSD and ETH withdrawals; USDC left via Noble and Circle’s CCTP.
Nearly seven weeks earlier, Liquid Network’s BTC reserve dropped from about 4,205 BTC to 197 BTC due to an exploit involving cached range-proof checks and further withdrawals prior to a halt.
Though the technical details differ, the underlying lesson is similar: Cached approvals or verifications are only safe if their context remains unchanged.
In the Initia scenario, Move's type system protections failed when operations relied on a check tied to previous code.
A fix can eliminate the vulnerability, and a restart can resume operations, but neither one independently replaces assets that were already lost.
The Foundation has pledged to make up the losses. Fulfilling that pledge will be the next test.
How many other systems are still relying on approvals whose underlying assumptions may have changed?
Get new scam files the moment we publish them — usually 2–3 emails a week.