Base Chain is a fast, low-cost, decentralized network that delivers sub-second, sub-cent global transactions. It is the foundation for a new economy where anyone, anywhere can participate.
Live
Triaged by Immunefi
Runnable PoC Required
KYC required
Select the category you'd like to explore
Assets in Scope
Impacts in Scope
Impacts In Scope
Only the following impacts are accepted within this program. All other impacts are not considered as in-scope, even if they affect something in the assets in scope table. Where a report demonstrates more than one impact, the highest applicable demonstrated impact governs. An impact that is claimed but not demonstrated does not raise the severity.
Proof of Concept (PoC) Requirements
A runnable PoC is required for every report, whatever its severity. Reports without a runnable PoC are closed. For more information, please read the Web3 PoC Guidelines.
The PoC must execute the in-scope code path that contains the bug, with the Production configuration in Assets in Scope, including signature checks and gossip validation. Outside the proof-component PoC exception below, a PoC that replaces a trusted component (preimage oracle, L1 provider, state database or verifier) with a mock that breaks that component’s contract is invalid. For enclave code, the local server harness is accepted if it runs the same request-handling code.
Where the PoC runs:
- Default: the designated local Everest devnet built from the competition commit (just devnet up everest), with the Production configuration, real binaries and default flags, and resource metering disabled (see Production configuration). This applies to Beryl, Cobalt, Denim and Everest code: Denim’s 200 ms cadence is active from L2 block 25 and Everest from block 100. PoCs for block production, derivation, timing or node resources report the measured timestamp lag or resource change against a baseline run. Mocks of the executor or the transaction pool are not accepted.
- Local forks: a local fork of Ethereum Sepolia, Base Sepolia, Ethereum mainnet or Base mainnet is allowed only when the exploit requires state that exists only on that network. Forks only, never the live network. A fork does not bring that network’s deployments into scope; the PoC must still execute in-scope code. A Base Sepolia fork for Denim timing starts at block 48,418,656 or later.
- Contracts at base/contracts v8.5.0-rc.1: a Foundry test in base/contracts at the competition commit (b9b76d97) that deploys the affected contracts with the Production configuration values, or the local devnet with base/base#6682 and #6680 applied (see Deployed contracts in scope).
- Proof components: the proof-component PoC exception below.
- Validity transactions: a local devnet only. Testing against the public sequencer endpoint is prohibited.
Proof-component PoC exception:
For in-scope proof programs, prover services, proposer/challenger logic and proof-verification contracts, runnable component-level PoCs are accepted. Researchers are not required to operate the complete proposer, prover and challenger stack or reproduce Base’s production proving infrastructure. The PoC must execute the affected in-scope code at the competition’s specified revision and applicable configuration.
Test fixtures, local harnesses and mocks may supply surrounding dependencies, provided they do not fabricate the vulnerability or bypass the security property being tested. For example, a local harness that bypasses hardware attestation may support a finding unrelated to attestation, but cannot establish an attestation bypass. Reports must identify mocked or substituted dependencies, explain what the PoC demonstrates, and distinguish demonstrated effects from remaining integration assumptions.
A report will not be rejected solely because the researcher cannot run the complete proof stack or access production proving infrastructure. Where end-to-end reproduction is limited by those requirements, Base and Immunefi will assess the runnable component-level reproduction and supporting evidence, with Base assisting in integration validation where needed. Final severity depends on the validated impact; a component-level defect does not automatically establish a higher-severity end-to-end outcome. In Proof soundness and Invalid output roots, severity follows the furthest stage that the PoC demonstrates or that Base’s integration validation confirms; a confirmed stage counts as demonstrated.
This exception applies only to the listed proof components that are already in scope. It does not expand asset scope, permit live-network testing, or relax the requirements for unrelated execution, sequencing, batching or validity-transaction findings. For findings spanning execution and proofs, the relevant execution behavior must still be faithfully reproduced. All other policy requirements remain unchanged.
Researchers may amend the PoC or the selected impact until the competition end date. Triage and mediation assess the report as it stands on that date.
Definitions
For this L2, “network not being able to confirm new transactions” means an unprivileged exploit prevents the sequencer from including otherwise-valid, adequately funded and fee-paying user transactions available to it for at least 30 seconds. Blocks that contain only the L1-info and BaseTime deposits count as not confirming user transactions.
A network processing node is shut down when it stops processing new blocks and does not recover without a manual restart. A total network shutdown is the Critical impact defined above.
An exploit through base_sendRawTransactionValidity is rated under the impact it demonstrates: Network not being able to confirm new transactions (Critical) when it stops user-transaction inclusion for at least 30 seconds; otherwise the node-shutdown or resource-consumption impacts. Degradation caused by volume alone is out of scope.
B20 impacts
Severity is assessed as if the token held mainnet value. A standalone policy or pause bypass without a demonstrated Critical impact is High. B20 balances, minting, roles, policies and pauses use the B20 and Smart Contract impacts, even though their implementation includes Rust precompiles. Execution or proving divergence and network availability use the Blockchain/DLT impacts.
Smart Contract impacts
Applies to base/contracts (multiproof system: AggregateVerifier, TEEVerifier, DisputeGameFactory, TEEProverRegistry), base/nitro-validator (NitroValidator, CertManager, P384Verifier) and base/contract-deployments (deployment and upgrade scripts).
Proof soundness
“Invalid TEE or ZK proof” means a proof accepted only through key compromise, forged attestation or a cryptographic break of the prover. Reports that rely on one are out of scope unless they show how to obtain it. A bug in NitroValidator, CertManager or P384Verifier that accepts an attestation AWS did not sign is a way to obtain one and stays in scope.
A valid proof that the deployed TEE or ZK program produces for a state transition different from canonical Base execution is a proof-soundness bug. It is Critical when the invalid root can also finalize despite Base’s Security Council and Guardian emergency actions (blacklisting or retiring the game), High when the PoC reaches AggregateVerifier acceptance or stops at the verifier library, and Medium when it stops at proof generation. Assumption (b) in Assumptions about Base’s response does not apply.
Invalid output roots. A report that causes AggregateVerifier to accept an invalid output root, or a game to resolve or a withdrawal to finalize against one, is Critical, whatever the path: a forged or bypassed proof, a schedule defect or a timing defect. Assumption (b) may lower it by one level, once. A proof-soundness bug or an execution divergence is the exception: acceptance alone is High, and it is Critical only when the invalid root can also finalize despite Base’s Security Council and Guardian emergency actions. A bug that stops before AggregateVerifier accepts the root is rated by the stage it reaches, as set out above.
Optional Project Info
Where might Security Researchers confuse out-of-scope code to be in-scope?
Given Base’s previous status as an OP Stack chain, participants may confuse OP Stack components (op-node, op-geth, op-batcher) as in scope. Only Base-native code in the Assets in Scope section is in scope. Optimism contracts that Base does not deploy are out of scope. The ZK host and its circuits (the zkVM program), including the changes Base made to OP-Succinct, are in scope. Soundness bugs in the zkVM are not in scope. See Networks and code in scope.
Is this an upgrade of an existing system? If so, which? And what are the main differences?
Everest adds Base Transactions (EIP-8430) and gas-limit billing. The competition also covers the Beryl, Cobalt and Denim components listed in Assets in Scope. Base Azul introduced the earlier changes below:
- Client consolidation: migrating from op-geth and op-node to base-reth-node (EL) and base-consensus (CL).
- Proof system migration: replacing Optimistic fault proofs (Cannon) with a dual-proof system (TEE and ZK) for faster finality.
- Custom features: Fusaka EIP support and future protocol innovations. See https://specs.base.org/upgrades/v1/overview for more details.
Where do you suspect there may be bugs and/or what attack vectors are you most concerned about?
Primary areas for review, subject to Assets in Scope:
- Offchain consensus and execution logic: state root derivation bugs in base-consensus and base-reth-node.
- Proof system integration: TEE and ZK dispute game submission logic, verifier contract bugs, and soundness issues in ZK recursion.
- Protocol transitions: incorrect hardfork activation logic, backwards compatibility breaks, and edge cases in EIP-1559 fee calculation.
- Multi-client discrepancies: execution differences between CL and EL that cause chain forks.
What ERC20 / ERC721 / ERC777 / ERC1155 token standards are supported?
At the execution layer, Base can run EVM-compatible ERC20, ERC721, ERC777 and ERC1155 contracts. The canonical bridge interfaces listed in the deployment manifest cover ERC20 and ERC721, subject to their documented token requirements; this does not imply dedicated ERC777 or ERC1155 bridging support. B20 exposes its documented ERC20-compatible interface and Base-specific extensions. Compatibility does not make third-party tokens or bridges in scope. A report must demonstrate an eligible impact in an in-scope Base component under the competition configuration.
What emergency actions may you want to use as a reason to downgrade an otherwise valid bug report?
Base does not close reports on the basis of its own response. The assumptions Base relies on, and the one-level downgrades they allow, are listed in Assumptions about Base’s response in Impact Terms.
What addresses would you consider any bug report requiring their involvement to be out of scope, as long as they operate within the privileges attributed to them?
Every trusted role listed in the next answer, when acting within the privileges attributed to it.
What addresses would you consider any bug report requiring their involvement be out of scope, even if they exceed the privileges attributed to them?
- The trusted roles: the sequencer, batcher, proposer, TEE signers and registrar, Security Council and Guardian. See Trust boundaries in Out of Scope & Rules.
- L1 Ethereum validators and stakers, which are not relevant to Base consensus.
- Third-party bridge validators (Relay, Across, Axelar and others), which are outside Base’s control.
- Base corporate infrastructure that is not part of the protocol (KMS, deployment pipelines and similar).
Which chains and/or networks will the code in scope be deployed to?
- Base Sepolia testnet for Beryl, Cobalt and Denim, and the designated local Everest devnet for Everest.
- Base mainnet, Zeronet and other devnets are out of scope, except for the local environments expressly permitted in the PoC Requirements.
What external dependencies are there?
- Ethereum mainnet (L1): state root anchoring, EIP-4844 blob data availability and gas price feeds.
- SP1 ZK prover: SP1 v6.10.0, with SP1 cluster v2.9.0.
- AWS Nitro Enclaves: TEE attestation infrastructure.
- Optimism upstream libraries (op-batcher, optimism/contracts) for bridge and withdrawal logic. No modifications to these libraries are in scope.
- Base's Reth fork, github.com/base/reth at 2b20e42de9c42b51ea0fb56599fda3c2f22640ef: the execution client dependency recorded in Assets in Scope.
Are there any unusual points about your protocol that may confuse Security Researchers?
The full list is in Intended behavior in Out of Scope & Rules. The points researchers most often misread:
- Single-proof finality model: a TEE or ZK proof is needed to finalize. A ZK proof can be used to challenge a TEE proof, but not the reverse.
- Intermediate output roots: proofs cover block ranges and commit to intermediate roots within a proposal, so disputes can cover a much smaller range of blocks.
- Soundness alert mechanism: if two different state roots both have valid proofs of the same type (both TEE or both ZK), the game can be automatically nullified.
What are the most valuable educational resources already available? (Ie. Documentation, Explainer videos or articles, etc)
- https://specs.base.org/upgrades/v1/overview: full V1 (Base Azul) specification.
- https://docs.base.org/base-chain/node-operators/base-v1-upgrade: operator migration guide.
- https://github.com/base/base: open-source code with README and architecture docs.
- https://blog.base.dev: technical blog posts on proofs and the stack migration.
Previous Audits
Base’s completed audit reports are listed below. Unfixed vulnerabilities mentioned in these reports, including their accepted-risk sections, are not eligible for a reward.
- Multiproof contracts audit 1: https://cantina.xyz/portfolio/25ba64ea-d6f3-411e-8338-419ffc385ba6
- Multiproof contracts audit 2: https://cantina.xyz/portfolio/b72c7078-f6da-4074-a3bd-4f938f469fb7
- TEE contracts audit 1: https://cantina.xyz/portfolio/423a9f33-b710-4445-a125-950b0a7771d7
- TEE contracts audit 2: https://cantina.xyz/portfolio/a4f952cf-1c5b-4e3c-8153-c3adff899613
- Denim 200 ms blocks (Cantina): denim-200ms-blocks.pdf
- Dynamic Upgrades contracts (Cantina): cantina_coinbase_dynamic_upgrades_sep2026.pdf
- Cobalt Dynamic Upgrades (Cantina): cobalt-dynamic-upgrades.pdf
- Beryl B20 (Cantina): beryl-b20.pdf
- Beryl precompile storage (Cantina): beryl-precompiles-storage.pdf
- Cobalt B20 (Cantina): cobalt-b20.pdf
- Nitro validator (Cantina, 3 August 2026): cantina-nitro-validator-05294ec0-2026-08-03.pdf
- Audit reports for Optimism smart contracts and components, including Kona: https://github.com/ethereum-optimism/optimism/tree/develop/docs/security-reviews#security-reviews
- Base Azul Audit Competition results: https://immunefi.com/audit-competition/audit-comp-base-azul/
Public Disclosure of Known Issues
Bug reports for publicly disclosed bugs are not eligible for a reward. Known issues are:
- The issues listed in Known Issues, below.
- Unfixed findings in the audit reports listed above.
- Limitations stated in code comments or NatSpec at the in-scope tag.
- Issues publicly disclosed in the in-scope repositories before the competition starts.
Entries added after launch apply only to reports submitted after they are published.
Known Issues
Reports of the issues below are not eligible for a reward. Each entry covers only the behavior it describes, within its stated scope; it does not exclude other defects in the same component.
Issues reported in public repositories
| ID | Issue | Description and scope | Status and reference |
|---|---|---|---|
| KI-02 | Proof-store initialization omits storage | Two separate MDBX defects: cursor errors are treated as end-of-table, and interrupted unordered writes can cause resume to skip unwritten addresses. Incomplete proof data impairs proving and recovery. RocksDB avoids the second defect. No invalid-root acceptance is established. | Open; fix not merged. #6586 |
| KI-03 | Validator startup waits indefinitely for execution readiness | After an initial SYNCING response, startup lacks a periodic execution-readiness retry. Unsafe blocks may arrive while safe and finalized derivation remains gated. Distinct from the fixed catch-up-buffer issue. | Open. #6566 |
| KI-04 | Pruned heads cause a SyncStart crash loop | Following snapshot catch-up, saved safe and finalized heads can reference pruned blocks; backward recovery then repeatedly fails on unavailable history. A pruned finalized block can also be read as an unset label, so recovery restarts from genesis instead of the saved checkpoint. Reported with minimal pruning; applies only to supported configurations. | Open; no complete recovery fix. #6505 |
| KI-05 | Derivation resets stall finalization | A reset clears finalization candidates without rebuilding all discarded entries. Repeated resets can stop finalized-head progress while safe-head and L1-finality progress continues. Covers this queue-reset mechanism, not every finalization stall. | Open; the proposed fix was closed without merging. #6228 |
| KI-06 | Lost validity transactions are not resent before expiry | Resend suppression lasts at least twice the configured validity lifetime and at least the standard resend interval. If a builder loses a successfully forwarded transaction through restart or eviction, retransmission can remain suppressed for its remaining lifetime. No general inclusion guarantee is implied. | Closed as not planned; not fixed. #6450 |
| KI-07 | Mixed nonce runs are selected incompletely | Selection can stop after validity nonce 0 and ordinary nonce 1, omitting executable validity nonce 2. The bridge traverses ordinary transactions but not the next validity run; deferred successors can miss expiry. | Closed as not planned; not fixed. #6451 |
| KI-08 | Resting validity transactions add build work | Native builds still visit resting validity transactions with false predicates. Work grows with the resting population and can delay eligible transactions. Does not establish the cause of every inclusion failure or a resource impact under this policy. | Open; optimization not merged. #6439 |
| KI-09 | Deposit receipts report incorrect DA usage | After Jovian, deposit receipts include blobGasUsed although block-level accounting excludes deposits. Receipt totals can exceed the header's data-availability footprint. This is RPC misreporting, not deposit overcharging. | Open. #6602 |
| KI-10 | Residual nonce-free admission race | A same-replay-ID transaction mined between validation and watcher admission can leave an unusable twin occupying sender, payer or pool capacity until expiry. Replay protection remains intact. Only this residual race is included, not the fixed broader eviction issue. | Residual gap after the main fix. #6596 |
| KI-11 | Stale TEE-signer cohorts recover serially | Recovery handles the invalid signer at the submission head; discovering later stale cohorts can wait for earlier replacement proofs, delaying proposals. Excludes the fixed retry-exhausted-session issue. The initiating signer change or RPC outage is not itself a vulnerability. | Bulk-recovery proposal closed without merging; per-signer cleanup remains. #5755 |
| KI-12 | Background tasks survive shutdown or failure | Separate lifecycle gaps: discovery loops on a closed channel; the proof host can detach a sibling task; witness failure can bypass server cancellation. Work and resources can remain active. Unprivileged reachability and a material availability impact require separate proof. | Open; fix not merged. #6537 |
| KI-13 | Batcher draining and startup-sync gaps | Separate defects: fatal work errors skip orderly flush and drain; optional startup synchronization can finish before the target L1 block is processed, permitting reposting. The latter requires --wait-node-sync, which is off by default. | Open; fix not merged. #6534 |
| KI-41 | Malformed batch arithmetic can panic in debug builds | Unchecked unsigned arithmetic in span-batch origin and timestamp derivation and related sequence-window checks can overflow on malformed values. Debug builds can panic; optimized release builds wrap. Covers these overflows only; it is not an exclusion for other malformed-batch findings. | Present at the competition commit; fix proposed, not merged. #6538 |
| KI-42 | Preimage response length truncates on 32-bit proof targets | The preimage oracle casts a channel-supplied u64 response-length prefix directly to usize. A value exceeding usize on a 32-bit target truncates, potentially desynchronizing subsequent reads. This narrowing does not occur on a 64-bit host. Covers this cast only; it is not an exclusion for other preimage-oracle or proof-soundness findings. | Present at the competition commit; fix proposed, not merged. #6538 |
| KI-47 | Simulation accepts a delegation off the protocol nonce | eth_estimateGas and eth_call accept a Base transaction that carries a delegation entry on a non-zero nonce key. The pool and block execution reject such a transaction, because a delegation must use the protocol nonce, so a wallet can get a successful simulation for a transaction that is then rejected at submission. RPC and wallet UX only; the rejected transaction never executes. | Open; fix not merged. #6652 |
| KI-52 | Simulation ignores sub-second block time | Base-transaction eth_call/eth_estimateGas check validity bounds at whole-second precision, and eth_simulateV1 skips the per-block BaseTime update after Denim. Simulations near a bound or reading BaseTime can differ from execution. Pool admission and block execution use exact milliseconds. | Known; fix planned. |
| KI-53 | Estimates ignore a lowered gas-limit override | For Base transactions, eth_estimateGas computes its ceiling before applying a caller's lowered block gas limit override, so the estimate can exceed that override. The real block limit, execution and ordinary-transaction estimates are unaffected. | Known; fix planned. |
| KI-54 | eth_simulateV1 skips the per-transaction gas cap | In validation mode, eth_simulateV1 does not apply the 16,777,216 per-transaction gas cap that blocks enforce, matching upstream Reth. A simulation can succeed for a transaction a block would reject. Base transactions are rejected by eth_simulateV1. | Known; matches upstream Reth. |
| KI-55 | base_transactionStatus omits included transactions | base_transactionStatus reports mempool presence only; once a transaction is included and leaves the pool, it returns Unknown, as for a dropped one. Use eth_getTransactionReceipt to confirm inclusion. Inclusion and receipts are unaffected. | Known; fix planned. |
| KI-56 | Channel heads bypass the sender slot limit | The next transaction on a channel with queued transactions is exempt from the per-sender slot check but still counted. With several channels, a sender can hold up to twice the configured slot limit in one pool. Global capacity, eviction and fees are unchanged. | Accepted pool limit; extends Reth's next-nonce exemption. |
| KI-57 | Invalid batch recovery differs from proofs | After Holocene, certain invalid user transactions in batch data make the node retry the derived block indefinitely, while the proof program falls back to deposits only. Requires the trusted batcher to post the data; no invalid-root acceptance is established. | Known; fix planned. |
| P-01 | RPC and metrics hang reported on v1.5.1 | On v1.5.1, blocking-thread exhaustion after about six hours made state RPCs and metrics hang while block following continued. Whether it reproduces at the competition commit is not established. Covers this blocking-thread mechanism only; other RPC hangs are not excluded. | Closed as not planned; no fix identified. #6230 |
Audit findings, documented limitations and specification differences
| ID | Issue | Description and scope | Status and reference |
|---|---|---|---|
| KI-14 | Policy arrays decoded before size checks | B20 account arrays are decoded before the 64-account cap. The audited Alloy 1.6.1 decoder can reserve memory before checking that all declared elements exist. The cap still applies before state updates; no out-of-memory exploit or corruption is established. | Accepted limitation. Beryl B20 §3.2.11; allocation only: Cobalt B20 §3.4.12 |
| KI-15 | Storage-layout hashing is unmetered | Mapping and dynamic-storage layout helpers perform keccak derivations outside the metered hashing path. Storage operations still incur gas. A material resource impact needs separate proof. | Accepted limitation. Beryl precompile storage §3.1.5 |
| KI-16 | Static-write failures revert instead of halt | Native precompiles reject static-context writes with catchable reverts rather than exceptional halts. State changes remain prohibited; only failure classification, return data and related gas semantics differ. | Accepted limitation. Beryl storage §3.3.19; Cobalt B20 §3.4.13 |
| KI-17 | Permits accept high-s signatures | B20 permits accept the high-s equivalent of a valid ECDSA signature. Owner nonces still prevent replay. This is a signature-canonicality limitation, not an authorization or replay bypass. | Accepted limitation. Beryl B20 §3.3.13 |
| KI-18 | Upgrade identity and coordination assumptions | Schedules use numeric positions and ignore unknown tail entries; commitments do not bind semantic upgrade identities. Independently launched consensus and execution services can diverge. Integrated startup shares a schedule; metrics-only mode does not apply it. No unprivileged schedule control is established. | Accepted coordination limitations. Cobalt Dynamic Upgrades §§3.3.1, 3.3.3, 3.4.3 |
| KI-19 | Incomplete initial upgrade schedule | A truncated contract schedule can disagree with a node's static suffix and cause rejection of honest proofs. Requires incorrect initialization or deployment; no current deployment misconfiguration is established. | Accepted deployment limitation. Dynamic Upgrades contracts §3.1.1 |
| KI-20 | BaseTime proxy hash is not validated | Activation checks proxy code presence and administrator, but not a canonical runtime-code hash. This is a deployment and activation assumption, not proof that an unprivileged user can replace the proxy or that a correctly deployed proxy is defective. | Accepted activation limitation. Denim 200 ms blocks §3.2.6 |
| KI-21 | Millisecond timestamp overflow differs across paths | RPC timestamp extraction wraps, consensus scheduling saturates, and Solidity truncates to uint64. They agree for realistically reachable timestamps; divergence requires values outside normal consensus bounds. | Acknowledged. Denim 200 ms blocks §3.2.2 |
| KI-22 | Zero Denim activation has different meanings | An explicit zero activation can mean fast blocks from genesis to the node but 'never' to AggregateVerifier. A misconfigured chain can produce proofs with inconsistent block-to-time mappings. Correct nonzero deployments are not affected. | Acknowledged configuration limitation. Denim 200 ms blocks §3.2.3 |
| KI-23 | Verifier intervals must match prover sampling | The verifier's configurable intermediate interval is not intrinsically tied to the prover's sampling cadence. A mismatched deployment can reject honest proofs. This requires a configuration mismatch, not an unprivileged change to the verifier. | Acknowledged deployment limitation. Denim 200 ms blocks §3.2.4 |
| KI-24 | Verifier assumes schedule slot 13 is Denim | AggregateVerifier reads the fast-block activation from numeric index 13 without independently binding it to Denim's identity. Correct ordered schedules satisfy the assumption; reordered or inconsistent deployments can yield incorrect timing. Related to, but more specific than, KI-18. | Acknowledged deployment assumption. Denim 200 ms blocks §3.2.5 |
| KI-25 | Additional BaseTime setter can disagree with metadata | A privileged unsafe producer can place the scheduled setter at tx[1] and another authorized setter later, leaving executed BaseTime different from the reported value. Ordinary users cannot create these protocol deposits; normal safe derivation does not construct this sequence. | Acknowledged privileged-producer behavior. Denim 200 ms blocks §3.2.8 |
| KI-26 | BaseTime setter can appear before activation | A privileged producer with synthetic or prelinked pre-activation state can execute a BaseTime setter before Denim; RPC can expose the premature millisecond value. Normal safe derivation does not create this condition. | Acknowledged privileged and configuration limitation. Denim 200 ms blocks §3.2.9 |
| KI-27 | Redirected first deposit can observe stale BaseTime | A privileged unsafe producer can redirect otherwise decodable L1-info calldata at tx[0] to a reader, which runs before the BaseTime update at tx[1]. It can observe the prior time even though final state and metadata agree. Normal safe derivation uses the canonical first deposit. | Acknowledged privileged-producer behavior. Denim 200 ms blocks §3.2.10 |
| KI-28 | Upgrade signal trusts the configured L1 endpoint | Schedule reads are pinned to an RPC-reported block hash but are not independently authenticated to the intended L1 chain and configured ProtocolVersions contract. A misconfigured or compromised trusted endpoint can supply the wrong schedule to applying modes. This is not arbitrary public RPC input. | Acknowledged L1 and configuration trust assumption. Cobalt Dynamic Upgrades §3.3.4 |
| KI-29 | Runtime activation can lack Beryl's required admin | A custom chain can start without Beryl or an activation admin, then enable Beryl through runtime-admin without the startup invariant being revalidated. Administrative calls revert and proof boot may reject the configuration; this does not grant unauthorized admin access. | Acknowledged custom-chain limitation. Cobalt Dynamic Upgrades §3.4.4 |
| KI-30 | Activation-admin fallback remains configuration-dependent | Beryl uses the chain-config admin. Cobalt supports a state-backed admin, but falls back to chain configuration while the stored admin is zero. Nodes with inconsistent fallback metadata can therefore disagree on authorization before a nonzero state admin is established. | Residual configuration assumption after partial mitigation. Beryl B20 §3.3.2 |
| KI-31 | Activation marker survives feature deactivation | Deactivation clears the feature flag but does not remove the registry's initialization bytecode marker. Code presence can mislead integrations that use it as an activation signal. isActivated and checkActivated remain authoritative; this is not a bypass of those APIs. | Acknowledged integration limitation. Beryl B20 §3.3.6 |
| KI-32 | Nitro library leaves debug-PCR policy to callers | A cryptographically valid debug-mode attestation with zero PCRs can pass the generic library's structural and signature verification. Integrators must enforce trusted production measurements. Base's registry integration mitigates this; the entry does not cover acceptance of debug enclaves into the production signer set. | Documented caller responsibility. Nitro validator §3.2.1; nitro-validator #56 |
| KI-33 | Nitro certificate owner and revoker are trusted roles | Privileged roles can revoke and unrevoke cached certificate identities, including a global validation halt. Deployment controls and operational safeguards mitigate key compromise; they do not remove the underlying trusted-role powers. No unprivileged authorization bypass is asserted. | Accepted trust model. Nitro validator §3.3.2; nitro-validator #78 |
| KI-40 | EIP-8037 state-gas accounting gaps | Native precompile paths do not fully enforce the EIP-8037 reservoir, SSTORE state-gas and existing-account code-deposit accounting. The audit deferred these because EIP-8037 was not planned for activation. Findings that depend on these gaps are known issues wherever EIP-8037 is enabled. | Deferred by the audit. Beryl precompile storage §§3.2.2–3.2.4 |
| KI-48 | Signing domains differ from the public EIP-8430 draft | Sender, payer and replay-ID hashes use the 0x79-prefixed domains 0x7900, 0x7901 and 0x7902. The public EIP-8430 draft still lists older domain bytes, so a signature computed from the draft is rejected. The implementation is authoritative; this is a specification lag, not a signature-validation defect. | Documented specification difference. #6552, included in v1.6.0-rc.9 |
| KI-49 | Pool holds Base transactions below the current base fee | Pool admission does not reject a Base transaction whose max fee per gas is below the current base fee. Protocol-nonce transactions wait in the base-fee subpool; sidecar transactions that cannot pay the current base fee are evicted first when the sidecar is full. Such transactions are only included once the base fee falls to their max fee. This is a pool-policy difference from the EIP-8430 draft, not a fee bypass. | Intended pool policy. |
| KI-50 | Call phases are capped at 1,024 per transaction | A Base transaction with more than 1,024 call phases is rejected at decoding, so it cannot enter the pool, be derived from a batch or appear in a valid block. The public EIP-8430 draft sets no phase limit. Every Base component applies the same cap; this is a specification difference, not a disagreement between nodes, derivation and proof programs. | Documented specification difference. |
| KI-51 | Validity window is checked at millisecond precision | validAfter and validBefore are compared with the block's exact millisecond time (timestamp plus the BaseTime millisecond part), while the public EIP-8430 draft compares them with the whole-second timestamp times 1,000. A bound that falls inside a block's second can be accepted or rejected differently from a literal reading of the draft. Pool admission, block building and execution apply the same rule; the implementation is authoritative. | Documented specification difference. |
Cantina review of EIP-8430 (Base Transactions): acknowledged findings
This review report is not public; each entry below describes the acknowledged finding in full.
| ID | Issue | Description and scope | Status and reference |
|---|---|---|---|
| KI-34 | Transaction Context exposes additional view functions | The Transaction Context precompile exposes getTransactionSender() and getTransactionSenderActorId() in addition to getTransactionPayer(), the only view function defined in the reviewed EIP-8430 text. Base retained the additional functions. Covers that interface difference only; it does not cover unauthorized state changes or authentication bypasses, or other Transaction Context defects. | Low; acknowledged, intentionally retained. |
| KI-36 | Sidecar replacement can discard a fresh sender balance | Replacing a sender's sole nonce-free sidecar transaction, or rolling back a guard-rejected replacement, can temporarily reduce its usage count to zero and delete its cached balance. If a canonical balance update advanced the balance epoch during validation, reconciliation can leave the balance unknown. A high coinbase-tip, low-fee-cap transaction can lose tip-backed ordering priority until a balance change or another submission refreshes the cache. No transaction drop or consensus impact. Covers this replacement bookkeeping mechanism, not every sidecar inclusion delay. | Low; acknowledged. Fix not merged: #6653 |
| KI-37 | Declared coinbase tip can boost priority without being paid | Pool admission can accept a transaction at the sender-intrinsic gas floor even though its remaining call gas cannot execute the phase-0 tip transfer through the SequencerFeeVault proxy. Balance-only tip-backing checks and declared-tip ordering can give it priority, but the tip call runs out of gas: ordinary gas fees are paid, the declared tip is not. This can displace or delay other transactions when blocks are full. Covers this gas-floor and tip-execution mismatch, not free execution or a general fee-payment bypass. | Low; acknowledged. |
| KI-43 | eth_estimateGas ignores the DA-footprint limit | After Jovian, pool admission and block execution reject a transaction whose data-availability footprint exceeds the block gas limit. The Base transaction estimator only bounds execution gas, so a request with very large metadata can get a successful estimate and then be rejected at submission. Wallet and relayer UX only. | Informational; acknowledged. |
| KI-44 | eth_estimateGas can return a gas limit that is too low or too high | The estimator builds the transaction from placeholder values where the request omits fields (gas limit, fee caps) and always uses nonce sequence 0. The real values can lengthen the encoding and raise the intrinsic calldata floor above the returned estimate, so the pool rejects the signed transaction with GasTooLow. Conversely, on a nonce key that was already used, the nonce update is priced as first use, so the estimate is 17,100 gas too high and a transaction signed with it pays for that margin under gas-limit billing. No authorization or replay bypass. | Informational; acknowledged. |
| KI-45 | Delegated-sender in-flight limit is not applied to the sidecar | Reth limits in-flight transactions from delegated senders (--rollup.txpool-max-inflight-delegated-slots, 4 by default in Base) and applies a nonce-gap rule. These rules apply only to the protocol pool; Base transactions in the sidecar are not subject to them. Per-payer and sidecar capacity limits still apply. Pool space only. | Informational; acknowledged, not fixing for now. |
| KI-46 | Pooled Base transaction memory size is undercounted | The memory size reported for a pooled Base transaction omits four fixed-size fields, so pool byte limits, sidecar capacity accounting and pool size metrics undercount each pooled transaction by a small constant amount. No unbounded memory growth is established. | Informational; acknowledged. |
Fixed at the competition commit, listed for reference: KI-01, inconsistent transaction-expiry precision (#6617, backported in #6646), and KI-35, pool removal racing readmission and releasing a live transaction's guard (#6253).
Private Known Issues Reward Policy
Private known issues, meaning known issues that were not publicly disclosed, are valid for a reward at their full severity and unlock the corresponding reward pool without any downgrade.
A private known issue that is publicly disclosed during the competition becomes a public known issue at the timestamp it is listed in the Mid-Contest Changelog. Reports of it submitted before that timestamp remain valid; reports submitted after it are invalid.
Public known issues are invalid.
Bug Fix Policy
Base may make bug fixes during the competition. Base develops these fixes in a private repository. A fix becomes public, and the bug out of scope, when Base lists it in the Mid-Contest Changelog with its UTC timestamp.
- Fixed bugs become out of scope once the fix is public, as defined in Code Update Assurance.
- Duplicate submissions of a bug are only valid if they are submitted before the fix is public.
- Once a bug fix is made public, Base provides an explanation of the bug and proof of its resolution, typically a pull request. Immunefi lists all bug fix information in the Mid-Contest Changelog and notifies researchers in the competition Discord channel.
- If a bug fix can be bypassed, the bypass is considered a new, valid, in-scope bug.
- If a code change unrelated to a bug introduces a bug, the introduced bug is considered new, valid, and in scope.
- In the rare exception where a bug fix must be made public before it can be safely disclosed, the bug is treated as a known issue and disclosure may be delayed until the fix is deployed. This prevents issue snipers from claiming other researchers' findings.
Feasibility Limitations
A report may be valid, cite an in-scope asset, and demonstrate an in-scope impact, while still facing real-world obstacles to executing the attack. Feasibility limitations define what may and may not be cited when evaluating those obstacles.
Assess demonstrated impact first. Feasibility affects payout, except where this policy expressly permits a severity adjustment. Any adjustment requires evidence that its stated conditions are met.
The following standards apply to this competition. Where this program page defines a specific rule, the program page prevails.
Attack Investment Amount A high capital requirement is generally not a valid reason to downgrade a report. Attacks requiring USD 100 million or more to execute are considered case by case. Flashloan-based attacks are valid where sufficient liquidity exists at submission or is expected within 12 months. https://immunefisupport.zendesk.com/hc/en-us/articles/17243068885265
Attacks With A Financial Risk To The Attacker Financial risk to the attacker only justifies a downgrade where the risk massively outweighs the reward. Low return on investment alone is not sufficient, and high risk alongside high return is not sufficient either. Evaluation is made from the perspective of an external user, not from information only the project holds. https://immunefisupport.zendesk.com/hc/en-us/articles/17454897136401
When Is An Impactful Attack Downgraded To Griefing? Applies to High or Critical findings where the attacker gains no profit. The baseline is that spending USD 1 to cause USD 10 or less in damage is Griefing, while spending USD 1 to cause USD 100 or more in damage is not. Ratios between these are evaluated case by case. https://immunefisupport.zendesk.com/hc/en-us/articles/17455102268305
Chain Rollbacks The ability to roll the chain back to a previous state is not a valid reason to downgrade impact, severity, or payout. Projects intending to rely on rollbacks as a safety mechanism should reflect that in their reward amounts rather than citing it after a report is submitted. https://immunefisupport.zendesk.com/hc/en-us/articles/16913153448721
Pre-Impact Bug Monitoring Where a project can objectively prove that automated tooling would prevent an exploit with 100 percent certainty, this may justify a downgrade of one severity level. Where prevention depends on human action following automated detection, the project must demonstrate 100 percent objective certainty of both detection and prevention. https://immunefisupport.zendesk.com/hc/en-us/articles/19430444320401
Attacks Involving Access to Privileged Addresses for Smart Contract Impacts A report is out of scope where the attack vector itself requires the use of a privileged address. It is not out of scope merely because a privileged address is affected by the exploit. Exceptions apply where a code vulnerability grants privileges the address did not originally hold, which may result in a one-level downgrade, and where an in-scope impact arises from a logic error that a privileged address can trigger, which remains valid. https://immunefisupport.zendesk.com/hc/en-us/articles/27383440951313
Attacks Involving Undeployed Code on GitHub or Equivalent for Smart Contract Impacts Bugs in undeployed code are valid only where the project has publicly stated the code will be deployed without further changes and the asset is in scope. In those cases the impact is treated as real, but any scaling element of the reward, such as funds at risk or freezing duration, is scaled to the minimum. https://immunefisupport.zendesk.com/hc/en-us/articles/27871612917649
Theft of Unclaimed Yield https://immunefisupport.zendesk.com/hc/en-us/articles/18753869210641
Theft Of Unclaimed Royalties https://immunefisupport.zendesk.com/hc/en-us/articles/18753902288785
Calculating the value of NFTs at risk https://immunefisupport.zendesk.com/hc/en-us/articles/19055304658065
For the full set of standards, see Feasibility Limitation Standards: https://immunefisupport.zendesk.com/hc/en-us/articles/16913132495377
Assumptions about Base’s response
No assumption below closes a report on its own. Each downgrade lowers severity by exactly one level.
- (a) Out of scope: impacts that require a proposal with an invalid root to resolve without being challenged, blacklisted or retired during its full dispute window (5 days with one accepted proof; 1 day after acceptance of a second proof of the other type; as re-extended by challenge() or nullify()), unless the reported bug is itself what gets that root past proof verification or the dispute window.
- (b) Base runs base-challenger, which recomputes every proposed output root and alerts on any divergence, and the Guardian can blacklist or retire a game within its dispute window. Under Immunefi's Pre-Impact Bug Monitoring standard (see Feasibility Limitations), Base may lower the severity of a report that gets an invalid root past proof verification by exactly one level, only when Base objectively demonstrates, for that report, 100 percent certainty of both detection by base-challenger and prevention by blacklisting or retiring the game within the window.
- (c) A restart lowers severity by exactly one level only when restarting with the documented configuration clears the condition. Defects in the default or production configuration that come back after a restart, and attacks that can be repeated, are not downgraded.
- (d) Attacks that require compromising Base-operated infrastructure are out of scope under Trust boundaries in Out of Scope & Rules. This is an exclusion, not a downgrade.
Direct loss to Base or users ≥ 2.5% of funds held within Bridge
Unintended permanent chain split requiring hard fork (network partition requiring hard fork)
Network not being able to confirm new transactions (total network shutdown)
Permanent freezing of funds held in Base’s canonical bridge, or of users’ native ETH balances on Base (fix requires hardfork). Funds held in third-party protocols are not covered.
An attacker causes a node in supported metrics-only mode to change fork activation from contract data, resulting in total network shutdown, a permanent chain split requiring a hard fork, or acceptance or finalization of an invalid root. Other effects are rated under the applicable impact.
Bypassing required sender or payer authorization, or replay protection, resulting in unauthorized execution of a Base Transaction, delegation or charging
Unauthorized minting of a B20 token, or total supply above its supply cap
Unauthorized transfer or spending of B20 balances, including through unauthorized creation or modification of an allowance enabling theft
Attacker-induced permanent freezing of B20 balances, excluding restrictions legitimately imposed through the issuer’s authorized roles, policies or pause
Unauthorized acquisition or restoration of a B20 role that enables unauthorized minting, seizure, destruction of balances or an equivalent Critical outcome
Bypassing a required B20 policy restriction or pause, causing unauthorized transfer, unauthorized minting or permanent freezing of B20 balances
Draining or stealing funds from the L1 bridge portal.
Out of scope
Base-specific:
- Findings demonstrated only on Base mainnet, the public Base Sepolia network, Zeronet or other devnets. The exceptions are the designated local Everest devnet, local forks and component-level PoCs, as permitted in the PoC Requirements.
- Contracts inherited from Optimism that Base does not deploy, and deployment scripts inside base/contracts.
- Findings that only appear under a configuration different from the Production configuration in Assets in Scope, unless the report shows an unprivileged way to reach it.
- Impacts that require a proposal with an invalid root to resolve without being challenged, blacklisted or retired during its full dispute window, as set out in Assumptions about Base’s response.
- Theoretical impacts without any proof or demonstration.
- EIP-8130 Keystore code (Zenith account configuration, non-secp256k1 authenticators, account-change types 0x02 and 0x03), removed from the recorded rc.9 baseline, and any other code gated to Zenith, unless shown reachable with only Everest active.
- crates/common/evm2, which is not the executor for Everest.
- The Base-operated token-gas payer service, which is deferred from the Everest launch. Core payer authorization and settlement, support for third-party ERC-8168 payers, and the node’s payer-rpc proxy when enabled as documented are in scope.
- Flashblocks components (builder mode, RPC methods, the websocket-proxy and flashblock_index predicates), when unused in the approved competition configuration. Shared code remains in scope when reachable in that configuration.
- The legacy NitroEnclaveVerifier, and the RISC Zero and Boundless attestation path.
Trust boundaries:
- Trusted roles: the sequencer, batcher, proposer, TEE signers and registrar, Security Council and Guardian. Attacks that require any of them to act maliciously or be compromised are out of scope.
- Trusted inputs: unsafe-block gossip signed by the unsafe-block signer, the Engine API, and L1/L2 RPC endpoints configured by the operator. Attacks that require malicious content from these sources are out of scope, unless the attacker can inject that content without controlling the source.
- Base-operated infrastructure: attacks that require compromising it are out of scope.
- Non-public endpoints: the admin, debug, engine, miner and txpool namespaces of base-consensus and base-reth-node, builder-internal methods such as base_insertValidatedTransaction, the prover service and the websocket-proxy origin. Public traffic reaches nodes only through proxyd method allowlisting. A missing authentication or flag check on a non-public endpoint is out of scope, unless the report shows the endpoint is reachable from an untrusted network in Base’s documented deployment. Public RPC allowlist (Base mainnet and Base Sepolia): eth_chainId, eth_blockNumber, eth_gasPrice, eth_maxPriorityFeePerGas, eth_feeHistory, eth_getBlockByHash, eth_getBlockByNumber, eth_getBlockReceipts, eth_getBlockTransactionCountByHash, eth_getBlockTransactionCountByNumber, eth_getBalance, eth_getStorageAt, eth_getTransactionCount, eth_getCode, eth_getProof, eth_call, eth_estimateGas, eth_createAccessList, eth_simulateV1, eth_getTransactionByHash, eth_getTransactionByBlockHashAndIndex, eth_getTransactionByBlockNumberAndIndex, eth_getTransactionReceipt, eth_getLogs, eth_newFilter, eth_newBlockFilter, eth_uninstallFilter, eth_getFilterChanges, eth_getFilterLogs, net_version, web3_clientVersion, eth_sendRawTransaction, eth_sendRawTransactionSync, base_sendRawTransactionValidity and base_transactionStatus. The local devnet exposes every method for testing. Local availability of an RPC method does not establish that an attacker can reach it in the supported deployment. A finding requiring a non-public method must demonstrate an unprivileged path to that method or the affected behavior. Using privileged local RPCs solely to prepare a test does not itself invalidate the PoC, provided the exploit does not depend on privileges unavailable to the attacker.
- RPC resource consumption that matches the behaviour of upstream Reth at the version Base’s fork tracks (base/reth 2b20e42d), for example eth_getLogs or debug_* methods, is out of scope. RPC code specific to Base, such as the BaseTime fields, is in scope.
- Attacks that rely on heavy volumetric flooding of the public RPC ingress nodes are out of scope, unless they cause a lasting effect that needs manual recovery, such as a chain halt. Temporary unavailability caused by a denial of service is out of scope.
- Only eth_sendRawTransaction, base_sendRawTransactionValidity, eth_sendRawTransactionSync and base_transactionStatus reach the nodes that build blocks. Other, read-only RPC methods are served by a separate set of nodes, so an issue limited to the read path cannot affect block building or the sequencer and is not rated Critical.
Intended behavior (not vulnerabilities):
- The L1 head (l1OriginHash) only attests that blob data came from a real L1 block. Which recent L1 block is used does not matter; the proof binds the L2 block number and state root.
- Nullifying a prover is a global kill-switch. It blocks all games and needs a Security Council upgrade to restore. Bonds stay recoverable through DelayedWETH.
- Challenges exist only to counter incorrect TEE proofs. A faulty ZK verifier is handled by nullifying it, which can be done through any unchallenged game.
- challenge() and nullify() have no gameOver() check, so invalid roots can be disputed for as long as possible.
- Child games of an invalid parent are invalid. Bonds are refunded when a game is blacklisted or retired, and are awarded only in the earliest parent game. Blacklisting stops a game from becoming the anchor; it does not claw back bonds.
- Additional proofs can only shorten a game’s resolution time.
- After retirement, recovery resets the AnchorStateRegistry anchor and deploys a new AggregateVerifier, as in contract-deployments/archive/legacy/sepolia/2026-04-20-activate-multiproof.
- TEE proofs are centralized: only Coinbase generates and submits them. The registrar treats signer discovery as the source of truth, so deregisterSigner is not a permanent revocation. The proposer discards and retries while it waits for a valid TEE signer.
- The sequencer and op-conductor cluster favours consistency: when the conductor cannot commit, block production stalls by design.
- Behaviour that follows the OP Stack specification and matches upstream op-node is not a Base-specific vulnerability.
- Gas-limit billing: a receipt’s gasUsed is the billed gas (the gas limit, plus the payer authentication cost for a sponsored Base Transaction), and the header’s gasUsed is the sum of the receipts. Unused gas and EIP-3529, EIP-7702 and precompile refunds do not reduce the fee.
- Documented exceptions: deposits keep metered gas and pay no L2 execution fee. eth_call, eth_estimateGas, eth_simulateV1 with validation=false, debug_traceCall and trace_call run without fee charging and report metered gas. Estimates for Base Transactions exclude the payer authentication cost, which settlement adds once. eth_simulateV1 rejects Base Transactions.
- The transaction hash is keccak256 of the full EIP-2718 encoding, including sender and payer authentication, so fee bumps and re-signed authentication change it. Nonce-free deduplication uses replay_id, not the transaction hash.
- A transaction that carries a delegation must use the protocol nonce (nonce key 0); a delegation on any other nonce key is rejected at inclusion. Delegation needs no separate EIP-7702 authorization, and Base never delegates automatically.
- Validity predicates are not allowed on EIP-7702 transactions, on senders delegated through EIP-7702, or on Base Transactions that carry a delegation; the transaction pool rejects them at admission.
Blockchain/DLT specific
- Incorrect data supplied by third party oracles
- Not to exclude oracle manipulation/flash loan attacks
- Impacts requiring basic economic and governance attacks (e.g. 51% attack)
- Lack of liquidity impacts
- Impacts from Sybil attacks
- Impacts involving centralization risks
All categories
- Impacts requiring attacks that the reporter has already exploited themselves, leading to damage
- Impacts caused by attacks requiring access to leaked keys/credentials
- Impacts caused by attacks requiring access to privileged addresses (including, but not limited to: governance and strategist contracts) without additional modifications to the privileges attributed
- Impacts relying on attacks involving the depegging of an external stablecoin where the attacker does not directly cause the depegging due to a bug in code
- Mentions of secrets, access tokens, API keys, private keys, etc. in Github will be considered out of scope without proof that they are in-use in production
- Best practice recommendations
- Feature requests
- Impacts on test files and configuration files unless stated otherwise in the bug bounty program
- Impacts requiring phishing or other social engineering attacks against project's employees and/or customers


