Understanding Verification Vulnerabilities in Cross-chain Bridges: An Empirical Study

What it is

An empirical taxonomy of Cross-chain Verification Vulnerabilities (CVVs): flaws where a bridge contract accepts invalid external-chain data (messages, proofs, signatures, sender identities) and authorizes an operation on it. It is built from 588 audit reports covering 36 bridge projects (from Chainspot and DefiLlama listings, retrieved May 2026). Reports come from firms including Trail of Bits, OpenZeppelin, Zellic, Quantstamp and Code4rena. The dataset is at https://anonymous.4open.science/r/CVV-Dataset/.

How it works

The reports contain 6,069 findings. Each finding is kept only if it passes three screens:

That leaves 1,482 CVVs. Two reviewers labeled them independently (Cohen's κ 0.85 for screening, 0.88 for pattern assignment) and sorted them into patterns by the guarantee each violates. The five guarantees are message authenticity, execution integrity, replay and freshness, verification soundness, and trust management.

Counts for the patterns visible in the fetched text (the table is cut off after C3):

• B3 Improper Handling of Unexpected Input: 705 (47.6%), in 31 projects. The paper says this means failed protocol-level validity checks on cross-chain data, not generic input validation.
• B1 Missing Execution Parameter Binding: 198 (13.4%), 19 projects.
• A1 Missing Message Identity Binding: 115 (7.8%), 21 projects.
• C1 Message/Proof Replay: 109 (7.4%), 17 projects.
• B2 Missing Destination Validation: 108 (7.3%), 18 projects.
• C2 Signature Replay/Reuse: 35 (2.4%), 8 projects.

The paper counts an access-control flaw as a CVV only if it lets data from an unauthorized cross-chain sender be accepted.

Preconditions / Affected

• The findings are Solidity-style bridge contracts across multisig, light-client, ZK and optimistic designs.
• The motivating incidents are Kelp DAO's rsETH bridge (about $292M, forged LayerZero messages) and Hyperbridge (forged state proofs, one billion wDOT minted).
• The authors count at least 11 bridge attacks of this class as of June 2026, with about $330M in losses.

What to do

• Use the pattern list as an audit checklist. For each destination-side entry point, check that it verifies the following:
• the message's source identity;
• the execution parameters;
• the destination domain;
• replay of the message, proof and signature;
• the validity of the incoming data itself (B3, the largest class).
• Don't rely on general-purpose scanners. The authors report that existing tools document limited CVV coverage.
• LLM triage is weak. The best of four models reached 51.2% Macro-F1 against a 29.3% constant baseline, and only two of the four beat that baseline. The authors attribute most errors to missing cross-component context and weak cross-chain semantic reasoning, so give a reviewer or model the gateway, verifier and executor together.

Caveats

• The dataset is audit-reported findings only, so unaudited or unreported bugs are absent.
• The 1,482 CVVs and the 830 dataset instances (786 files) are different counts. The fetched text does not explain the gap.
• The page is truncated, so the remaining patterns (C3 onward), the severity analysis and the detector comparison details are not covered here.