[ZIP-20] ZKsync v34 Upgrade for EraVM Chains

Proposal Type ZIP
One Sentence Summary This ZIP proposes the v34 upgrade for EraVM chains, which by default requires every batch to be proven by both Boojum and Airbender (dual proofs) and adds two safeguards that keep L1→L2 transactions moving: a 15M execution gas cap and operator execution omission.
Proposal Author Matter Labs
Proposal Sponsor Matter Labs
Date Created November 2026
Version v1
Summary of Action Upgrade EraVM chains to protocol version v0.34.0
Link to Contracts draft-v34-era-only branch

Abstract

This ZIP proposes the v34 protocol upgrade for EraVM chains. The release delivers:

  • Dual proofs, which by default require every batch to be proven by two independent proof systems, Boojum and Airbender.
  • Operator execution omission, which lets the operator skip the execution of an individual L1→L2 transaction, so that an unprovable one does not block the queue.
  • A 15M execution gas cap for L1→L2 transactions.

Motivation

v34 gives EraVM chains their strongest security model yet: to settle an invalid batch, an attacker would have to break two independent proof systems instead of one. It is a proactive step ahead of fast-growing AI-driven cyber capabilities.

Today, every batch on EraVM chains is proven by a single proof system, Boojum, so settlement relies on that one system being sound. v34 adds Airbender as a second, independent proof system. By default, a batch settles only with a valid proof from each.

Adding a second proof system significantly reduces soundness risk: a state transition is only accepted when two independent implementations agree on it, so a bug in one of them cannot be exploited unless the other has a matching bug. But it comes at the cost of increased liveness risk. Any small mismatch between the two systems, even one with no security impact, makes the affected L1→L2 transaction impossible to prove. Because these transactions are processed in order, one unprovable transaction would stall every transaction queued behind it.

Two changes mitigate this. The gas cap limits the execution and proving work a single L1→L2 transaction can request, which makes it much less likely that a transaction exceeds what either prover can handle. For a transaction that still cannot be proven, the operator can now skip its execution: the transaction is still included but never takes effect, and the queue moves on. An execution omission affects only that one transaction and is recorded on-chain, so it can unblock the queue, but no transaction can be skipped silently.

Specification

The implementation can be viewed in Github on the draft-v34-era-only branch of matter-labs/era-contracts. This upgrade applies to EraVM chains only.

Dual proofs: Boojum and Airbender

By default, ZKsync v34 requires every batch on EraVM chains to be proven by two independent proof systems, Boojum and Airbender (era-contracts #2275, #2451).

  • The new EraMultiProofVerifier accepts a batch only with a valid proof from each enabled system: a Boojum FFLONK proof and an Airbender PLONK proof.
  • Each batch now has a second commitment for Airbender. StoredBatchInfo, the operator’s commit data (CommitBatchInfo) and the genesis parameters for new chains (ChainCreationParams) each gain an Airbender field.
  • The chain admin can disable one proof system with setProofSystemStatus, for example to keep the chain live during a prover incident. At least one must stay enabled. There is no time limit, so a chain can run on one proof system until the admin re-enables the other.
  • The Boojum PLONK verifier is removed (era-contracts #2505). Boojum proofs are FFLONK only, so EraDualVerifier, which accepted a Boojum proof in either FFLONK or PLONK form, is no longer needed.

Operator execution omission for L1→L2 transactions

era-contracts #2500 lets the operator skip the execution of an individual L1→L2 (priority) transaction, so that one transaction that cannot be proven does not block the queue. It applies to every EraVM chain that settles on L1, with no opt-in, and covers every priority transaction, including the protocol’s own service transactions. Upgrade transactions cannot be skipped. Execution omission is a bootloader-only change.

  • A skipped transaction is handled like one that reverts on its first instruction: it ends with the same status as a reverted transaction, and the priority queue moves past it.
  • As with any reverted L1→L2 transaction, its full execution gas is charged and the rest of the deposited base token is refunded on L2 to its refund recipient. Any other token sent with it, such as an ERC-20 deposit, can be claimed back on L1.
  • For each skipped transaction, the bootloader sends an extra L2→L1 log with the transaction hash, so that on L1 an execution omission can be told apart from a normal revert.
  • On any settlement layer other than L1, a batch that contains an execution omission is rejected.

Gas cap for L1→L2 transactions

era-contracts #2450 caps the gas an L1→L2 transaction on EraVM chains can use for execution at 15M.

  • The cap applies to every L1→L2 transaction whose gas limit is set by its sender, including those sent by chain admins or governance. Requests above the cap revert on L1. Only service, upgrade and genesis transactions, whose gas limit the protocol sets, are exempt.
  • The cap is a protocol constant, so no per-chain configuration change is needed. Chains already configured with a lower limit keep it.

Additional changes

  • The EVM emulator no longer turns the offset of a zero-length memory access, which EVM allows to be any value, into an EraVM memory pointer (era-contracts #2487).

Note on Version Numbering

This upgrade moves EraVM chains from v0.33.0 (set by ZIP-18) to v0.34.0. ZKsync OS chains are not upgraded by this proposal, and the separate ZKsync OS v0.34.0 release is not part of it.

Rationale

Dual proofs: Boojum and Airbender

Until v34, every batch on EraVM chains was proven by Boojum alone, so a soundness bug in Boojum could be enough to settle an invalid batch. Requiring an Airbender proof as well adds a second, independent layer of protection against such bugs. The two systems commit to part of the batch data differently, so each batch carries a second commitment for Airbender; everything else in the two commitments is the same. Both commitments are always computed, even while one proof system is disabled, so the chain admin can switch a proof system off and back on while batches are in flight.

Operator execution omission for L1→L2 transactions

L1→L2 transactions are processed strictly in order, so one that cannot be proven blocks every L1→L2 transaction queued behind it. Two proof systems make this more likely, since each transaction now has to fit within the proving limits of both. Disabling the proof system that cannot prove it would unblock the queue, but only by giving up the dual-proof guarantee while it is disabled. Execution omission is a fallback for such exceptional cases: the operator skips the execution of that one transaction, and the queue keeps moving with both proof systems enabled. It is a single per-transaction flag that reuses the existing revert path.

Gas cap for L1→L2 transactions

The cap is a liveness safeguard: it limits the execution and proving work that the sender of an L1→L2 transaction can request. Before v34, senders could request up to their chain’s own limit, 72M gas by default. The 15M value was chosen from observed L1→L2 usage and is not expected to affect existing protocols.

Implementation & Backwards Compatibility

The upgrade modifies the bootloader and the EVM emulator on L2, the L1 contracts that settle batches, verify proofs and validate L1→L2 transactions, and the server-side protocol implementation and provers. Once this proposal is enacted, each chain is upgraded separately, at a time set by its chain admin, and only after every batch it has committed has been executed on L1. The breaking changes for node operators, chain operators and integrators are listed below.

Breaking Changes

  1. Node software and external nodes must upgrade to a zksync-era release that supports protocol version v0.34. The minimum release and the external node image tag are TBC.
  2. By default, every batch needs both a Boojum and an Airbender proof. Chain operators must have both provers running when their chain upgrades, unless the chain admin disables one with setProofSystemStatus.
  3. The L1 batch data format changes. StoredBatchInfo gains airbenderCommitment, so stored batch hashes change from v34. Commit data adds airbenderBootloaderHeapHash, and the proof carries a Boojum FFLONK proof and an Airbender proof. Commit, prove and execute data move from encoding version 1 to 5. Tools that build or decode this calldata, or recompute batch hashes, must be updated.
  4. EraDualVerifier is replaced by EraMultiProofVerifier, and Boojum PLONK proofs are no longer accepted. On the new verifier, verificationKeyHash() returns the Boojum FFLONK key hash instead of the PLONK one, verificationKeyHash(1) (PLONK) reverts, and the Airbender key hash is verificationKeyHash(2). Integrations that read the chain’s verifier or its key hashes must update.
  5. L1→L2 transactions that request more than 15M gas for execution now revert. Scripts and integrations that request more, including governance and chain-admin calls, must lower the gas limit. getPriorityTxMaxGasLimit() is unchanged and can return more than 15M. The new getUserPriorityTxMaxGasLimit() returns the limit that actually applies.
  6. The operator can skip the execution of an L1→L2 transaction. Integrations that assume an L1→L2 transaction with enough gas always executes, such as relays or cross-chain governance, must handle it being skipped, which looks the same as a revert.

Security Considerations

The v34 upgrade’s trust surface centers on the second proof system (Airbender), the chain admin’s ability to disable one proof system, and operator execution omission.

  • Dual proofs trade liveness for safety. While both proof systems are enabled, a soundness bug in one of them is not enough to settle an invalid batch. This does not cover code that both proofs rely on, such as the bootloader and the L1 contracts. In exchange, a failure in either prover stops batches from being proven, which is why the chain admin can disable one proof system.
  • The dual-proof guarantee depends on the chain admin. The chain admin can switch either proof system off at any moment and for as long as it chooses, including for batches already committed but not yet proven. Settlement then relies on the other alone: on Boojum, as before v34, or on Airbender. On the ZKsync Era chain itself, the chain admin is a contract controlled by a 4-of-7 multisig operated by Matter Labs. Governance cannot directly re-enable a disabled proof system; it can freeze the chain or push an upgrade to it. Chain admins should announce when they disable and re-enable a proof system. Matter Labs, as chain admin of the ZKsync Era chain, will publicly record each such change within 24 hours in a dedicated thread on the ZK Nation forum.
  • Execution omission keeps the L1→L2 queue moving, at the cost of more trust in the operator. It is intended only for exceptional cases, such as an L1→L2 transaction that cannot be proven or one whose execution would exploit a vulnerability. Before v34, an operator of a chain could keep one L1→L2 transaction from executing only by stalling all L1→L2 processing. From v34, it can skip the execution of a single one while everything else keeps running: the transaction’s funds, less the gas fee, remain recoverable, but the transaction never takes effect. Anyone relying on an L1→L2 transaction, from users sending transactions through L1 to governance and chain-admin calls, has to trust the operator not to skip it. Each use can be proven on L1, but nothing on-chain limits it. Operators should disclose each use. Matter Labs intends to disclose uses on the ZKsync Era chain and the reason, except where disclosure would create a security risk.

Audit

OpenZeppelin audited the main v34 changes from September 25 to September 30, 2026. The audit report is publicly accessible here. It covers #2451, #2505, #2500, #2450 and #2487.

The report finds the core dual-proof design sound and no critical, high or medium severity issues. Its main concern is the trust that execution omission places in the operator (see Security Considerations). One low severity issue, L-01, was acknowledged: a skipped transaction is charged its full execution gas, like any reverted L1→L2 transaction, so the operator earns more than if it had executed. All other findings were resolved in #2559 and #2560, one of them partially.

The audited pull requests are in the upgrade unchanged. Since the audit, the contracts changed only through these fixes, the rename from v33 to v34 (#2542), and two regenerations of the Airbender verifier contract (#2599, #2623). The Airbender verifier contract and the ZKsync OS-only changes added in #2275, the proof systems themselves and the upgrade calldata were not part of this audit.