BlackHart
Go live
Hacks feed/Payy Network

Payy Network

September 24, 2026·Ethereum·Signature Replay
$1.9M
total loss
StatusConfirmed

Why it happened

Attack class: Signature replay / forgery

Payy's verifier, on an outdated proving library, accepted a forged burn proof, and the bridge paid out the fake withdrawal with no second check.

What would have stopped it

  • Track security advisories for the proving library and upgrade promptly.
  • Reject impossible values in the rollup contract itself.

Not scored before the exploit

BlackHart had not scored Payy Network before this exploit.

An attacker took about $1.92 million in USDC from Payy, a privacy payments network with its own rollup bridge on Ethereum. Payy's proof verifier accepted a fake withdrawal backed by no real balance, so the bridge paid out 1.83 million USDC in one batch and another 90,202 USDC hours later. The cause was a known bug in the third-party proving library Payy uses, fixed upstream months earlier, which Payy had not updated. Payy paused the whole network, freezing user balances and cards until a fix ships. The attacker sold most of the USDC for about 683 ETH and sent it through Tornado Cash.

Payy signing keyssafe(Payy ruled out a key compromise)
Other Ethereum appssafe
User balances and Payy cardspartially affected(frozen while the network is paused)
Rollup bridge USDCdrained(about 1.92 million USDC drained)
Exploit anatomy

Payy's verifier accepted a fake burn, and the Payy rollup bridge paid 1.83 million USDC to the attacker wallet in one settlement. The USDC went to a second wallet, was sold for about 683 ETH, and was mixed through Tornado Cash.

The attack in 3 steps. Step 1, Forged burn: Payy rollup bridge (bridge). Step 2, Settle: Attacker wallet (attacker), about $1.92M reached this step. Step 3, Mixer: Second wallet to Tornado Cash (exit), about $1.83M reached this step.

  1. Step 1: Forged burn

    • Bridge
      Payy rollup bridge
      USDC held for Payy users
      pays out every verified withdrawal
      0x367c1eaf...d85270
  2. Step 2: Settle

  3. Step 3: Mixer

    • Exit
      Second wallet to Tornado Cash
      sold for about 683 ETH
      682.3 ETH mixed on September 26
      0xb483b174...3f3d38

Why each step worked

    • Verifier accepted an invalid burn proofknown proving-library bug, fixed upstream in May 2026
    • Outdated pinned dependencyPayy ran an October 2025 build
    • No on-chain rule against a zero burn hashmissing controla value normal users cannot produce
    • No independent balance checkmissing controlaggregation reused the same verifier
    • Pause came ten hours after the first payoutmissing controlsecond settlement went through first
Full forensic detail

Step-by-step reconstruction, root cause, counterfactuals, remediation, and disclosure timeline.

Exploit anatomy

1.
Two days earlier the attacker made two small USDC deposits into Payy's rollup from the wallet that would later receive the stolen funds.
2.
At 03:25 UTC on September 24 a forged burn, a withdrawal request with no real balance behind it, was recorded on Payy's own chain. Payy's deployed proof verifier accepted the invalid proof, so Payy's prover included it in a batch and Payy's validator signed it.
3.
When that batch settled on Ethereum at 04:21 UTC, the rollup bridge paid 1,828,589.38 USDC to the attacker alongside a set of ordinary user withdrawals.
0x367c1eaf...d85270verifyRollup(uint256,bytes32,bytes,bytes32[],bytes32,(bytes32,bytes32,uint256)[])tx:0xf43abdac...1e1814
4.
A second forged burn settled at 09:30 UTC and paid another 90,202.82 USDC to the same wallet, plus 1 USDC through a burn substitutor. Total taken: about 1,918,793 USDC.
5.
The attacker moved the 1.83 million USDC to a second wallet within two minutes, sold it for about 683 ETH, and sent 682.3 ETH into Tornado Cash in 19 deposits on September 26. The last 90,202 USDC left the first wallet on October 30.

Root cause

Payy's rollup bridge on Ethereum pays out whatever withdrawals a verified batch contains, so its safety rests entirely on the zero-knowledge verifier rejecting invalid proofs. Per Payy's September 29 post-mortem, the deployed verifier, built on Aztec's Noir and Barretenberg proving library, accepted an invalid burn proof. Certora's analysis traces this to a soundness bug in Barretenberg that Aztec fixed in May 2026 and shipped in its v5.0.0 release on July 13; Payy had pinned its services to an October 2025 build and never picked up the fix. Once the forged burn passed verification, Payy's own prover and validator processed it like any other withdrawal, which is why early reports, including ExVul's, pointed at Payy's infrastructure. Payy later ruled out a compromised key, social engineering or an attack on its off-chain systems. Two further gaps made it worse: the forged burn carried an all-zero burn hash that ordinary users cannot produce, yet the rollup contract had no rule rejecting it, and no later aggregation step re-checked note balances. Payy paused the network the same day.

Prevention analysis

Track security advisories for the proving library and upgrade promptly.

The bug was fixed upstream four months before the attack and in a stable release more than two months before. Upgrading the pinned verifier build would have closed the hole.

Reject impossible values in the rollup contract itself.

The forged withdrawal carried an all-zero burn hash that normal users cannot create. A simple on-chain check for that value would have blocked both payouts even with the verifier fooled.

Add an independent balance check before settlement.

Payy's aggregation re-verified each proof with the same verifier, so nothing confirmed that burned notes existed. A second check against note balances, or a different proof system, would have caught the fake burn.

Limit outsized withdrawals per batch.

One withdrawal of 1.83 million USDC sat in a batch of small user withdrawals. A per-withdrawal cap or delay on unusually large exits would have limited the loss.

Similar incidents

Aztec Connect (2026)

A retired Aztec rollup lost about $2.18 million to forged proofs its genuine verifier accepted. Same class: a proving-system soundness failure turned straight into a withdrawal.

FOOMCASH (2026)

A flaw in a Groth16 verifier let an attacker forge proofs and drain about $2.26 million. Same root: the contract trusted a verifier that could be satisfied without a valid statement.

Taiko Bridge (2026)

Rogue provers finalized fake state and the bridge paid out about $1.75 million. Same lesson: a rollup bridge is only as strong as the weakest check that can finalize state on Ethereum.

Remediation

1.Payy paused all network transactions on September 24 and published a post-mortem on September 29.
2.Payy said a user remediation plan would follow; as of the latest public reports no plan or restart had been published, so users should follow Payy's official channels.
3.As of the latest public reports, a confirmed verifier patch and an independent review of that fix were still pending.pending
4.Any team running a Noir or Barretenberg verifier should confirm it is on a release that includes Aztec's May 2026 soundness fix and add on-chain sanity checks so a fooled verifier cannot pay out alone.

Timeline

2025-10-30Payy pins its services to a Barretenberg build from October 2025.
2026-05-14Aztec commits the Barretenberg soundness fix.
2026-07-13Aztec ships the fix in Barretenberg v5.0.0.
2026-09-22The attacker's wallet makes a small deposit into Payy's rollup.
2026-09-24First forged burn is recorded on Payy's chain.
2026-09-24Batch settles on Ethereum and pays 1,828,589.38 USDC to the attacker.
2026-09-24Attacker moves the USDC to a second wallet and sells it for about 683 ETH.
2026-09-24Second forged burn settles and pays 90,202.82 USDC.
2026-09-24Payy announces the exploit and pauses the network.
2026-09-26682.3 ETH is sent into Tornado Cash in 19 deposits.
2026-09-29Payy publishes its post-mortem.
2026-10-30The remaining 90,202 USDC leaves the first attacker wallet.
Payy (post-mortem). ExVul (initial alert). Certora (root-cause analysis). DC Posch of Daimo (traced the payout to the verifier)
Continuous adversarial monitoring

Get your protocol scored across 12 dimensions, or request ongoing coverage.