Crypto exploits

Aztec V1 exploit analysis

An attacker called the permissionless escapeHatch function on Aztec's deprecated Private Rollup Bridge three times on June 17, 2026, using Turbo-PLONK proofs whose declared root matched the real on-chain state despite proving membership in a fabricated tree, draining 1,158 ETH, 150,000 DAI and 0.4696 renBTC, about $2.2 million per DeFiHackLabs.

David

Security ·

Reported loss
$2.2M
Occurred
Logged
Chain
Ethereum

Logged as “2026-06-17: Escapehatch Proof Forgery Permissionless Rollupprocessor Exit on Aztec V1”.

Analysis3 min read

The contract the attacker hit, 0x737901bea3eeb88459df9ef1BE8fF3Ae1B42A2ba, is labelled on Etherscan as "Aztec: Private Rollup Bridge" and carries verified source code, an exact match for Solidity 0.6.10. It is a deprecated Aztec 2.0 rollup contract, a separate deployment from the Aztec Connect contract drained three days earlier, though press coverage often groups the two under the Aztec Connect name. Aztec Labs had already stopped operating the product and administrative control was renounced, so nobody could pause it, upgrade it, or patch a bug once one surfaced.

The escapeHatch route

escapeHatch(bytes proofData, bytes signatures, bytes viewingKeys) is a permissionless emergency exit built into the rollup: inside a reserved block window, anyone can submit a Turbo-PLONK rollup proof, and the contract pays out whatever withdrawal the proof's public inputs encode, checked against the verifier contract at 0x48cb7ba00d087541dc8e2b3738f80fdd1fee8ce8. Rekt's write-up on the incident traces the break to a binding gap between the proof's private witness and its public input: the circuit proves that a withdrawal note sits inside a Merkle tree, but never constrains that tree's root to equal the root the proof declares to the verifier. An attacker can build a private, fabricated tree holding a withdrawal note backed by no real deposit, generate a valid proof over that fake tree, then hand the verifier the real on-chain data root as the public input instead. The circuit accepts the proof. The on-chain check compares the declared root to the genuine state root, finds a match, and pays out, because nothing tied the fabricated tree to the input that was actually checked.

Attacker EOA 0x6952d9246e9aFE8B887B2877225163436F78E97F called escapeHatch directly, with no attack contract, three times on June 17, 2026:

  1. A proof encoding a withdrawal of 1,158 ETH, confirmed on Etherscan in block 25,339,094 at 18:34:47 UTC.
  2. A second proof for 150,000 DAI, called fifteen minutes later.
  3. A third, for 0.4696 renBTC, three blocks after the DAI call according to the proof of concept's own fork targets.

Each call drained what DeFiHackLabs' reproduction treats as the rollup's full remaining balance of that asset. Nothing in the contract distinguished a forged proof from a genuine one, so there was no partial failure anywhere in the sequence to stop it.

The reported losses, and the exploit three days earlier

DeFiHackLabs puts the total at about $2.2 million: 1,158 ETH, 150,000 DAI and 0.4696 renBTC. Rekt's own tally lands close, at about $2.197 million. Etherscan's transaction page shows the ETH leg at about $2.60 million, but that is priced at today's ETH rate; its estimated value on the day of the transfer is about $2.03 million, in line with Rekt's roughly $2.04 million for the ETH. The headline gap is a pricing-date artefact, not a disagreement over the asset amounts.

This was not the first time the same contract family paid out against a forged proof that week. Three days earlier, on June 14, a different attacker, address 0x0F18D8b44a740272f0be4d08338d2b165b7EdD17, drained an amount sources put between $2.19 million and $2.28 million from the rollup's settlement path by a different route: a mismatch between the number of transactions a rollup proof validated and the number the settlement code actually processed, letting declared deposits outrun real ones. BlockSec's analysis, reported by The Block, found that despite the different entry points, both bugs reduce to the same class of failure: a public input the circuit never bound tightly enough to the on-chain state it was supposed to represent. That earlier incident has its own write-up at Aztec Connect's proof-of-concept analysis.

Aztec Labs' response, as reported by The Block, was that it holds no admin keys or control over the contract, that the bridge cannot be paused or upgraded by the team, and that the incident has no bearing on the current Aztec network or the AZTEC token. No bounty, patch or compensation followed either exploit.

What would have stopped this specific call is narrow: a constraint inside the escape hatch circuit binding the Merkle root used to prove note membership to the same root declared as the proof's public input, so a fabricated tree could never be paired with a genuine on-chain root and still pass verification.

Sources read (10)

Other exploits

Query every incident yourself — free, no signup

Updated 9 Nov 2026. Figures are as reported by the sources listed above and are not independently verified.