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

Crypto.world

> Cross-chain bridges work by having a contract on the destination chain verify data that came from somewhere else — a message, a state proof, a signature, a validator set.

**arXiv**: 2610.13170

**Published**: 2026-10-09

**Authors**: Yuying Wang, Yi Guo, Jiachi Chen, Hanwen Zhang, Yi Sun

## AI summary

## Overview

**Research area:** Blockchain security — smart contract and cross-chain bridge vulnerability analysis, empirical software-security measurement, and LLM-based vulnerability detection.

**Technical level:** Intermediate. Readers should know what smart contracts, Merkle proofs, validator signatures and cross-chain bridges are; the paper's taxonomy and evaluation metrics (precision, recall, F1, Cohen's κ) are explained in enough detail to follow without prior audit experience.

**Scope (1 sentence):** A systematic empirical study that derives a taxonomy of Cross-chain Verification Vulnerabilities (CVVs) from 588 audit reports across 36 real-world bridge projects, builds an 830-instance dataset from them, and benchmarks 12 existing detection tools and 4 LLMs against that dataset.

## What This Paper Is About

Cross-chain bridges work by having a contract on the destination chain verify data that came from somewhere else — a message, a state proof, a signature, a validator set. If that verification logic is wrong or incomplete, the destination chain accepts data it should have rejected and executes an unauthorized transfer or mint. The authors call these Cross-chain Verification Vulnerabilities (CVVs) and argue that existing vulnerability classifications, which describe implementation-level defects (access control, integer overflow), do not capture *which cross-chain verification guarantee* was broken. The goal is to characterize CVVs systematically, publish a labelled dataset that maps audit findings to the exact audited source-code versions and files, and measure whether today's tools and LLMs can actually find them.

## Key Contributions

1. **First systematic empirical study of CVVs.** The authors screen 6,069 security findings from 588 audit reports covering 36 real-world bridge projects and identify 1,482 CVVs, from which they derive 11 CVV patterns organized into 5 high-level categories.
2. **A labelled CVV dataset.** 830 vulnerability instances traced to reliably recovered audited source-code versions, across 209 audit reports and 27 projects, mapped to 786 source files, with links among reports, versions, affected files and pattern labels (publicly released at an anonymous repository URL given in the paper).
3. **Characterization of how CVVs manifest.** Analysis of pattern distribution, normalized severity, and cross-project diversity, including per-project pattern heat maps for the six most represented bridges.
4. **A benchmark of existing detection capability.** A documented applicability assessment of 12 detectors (8 traditional single-chain tools plus 4 cross-chain tools) against the 11 patterns, and an evaluation of 4 LLMs, showing only two of four beat a 29.3% Macro-F1 all-pattern constant baseline, with the best at 51.2%.

## Main Findings

- **CVVs are a real, costly attack class.** As of June 2026 the authors identify at least 11 bridge attacks involving such vulnerabilities, with about $330 million in losses. Examples cited include the Kelp DAO rsETH bridge attack (estimated $292 million, via forged LayerZero cross-chain messages) and the Hyperbridge attack (forged state proofs used to mint one billion wDOT).
- **Execution integrity dominates, not cryptography.** In the 830-instance dataset, Execution Integrity accounts for 577 instances (69.52%), Trust Management 87 (10.48%), Replay and Freshness 82 (9.88%), Message Authenticity 63 (7.59%), and Verification Soundness only 21 (2.53%). Improper Handling of Unexpected Input (B3) alone is 409 instances (49.28%); Missing Execution Parameter Binding (B1) is 100 (12.05%).
- **Prevalence does not equal severity.** Across the normalized severity labels, 199 CVVs (23.98%) are High. B3 is the most widespread pattern but not the most severe; Improper Trust Root Management (E1, 75 instances) and Message/Proof Replay (C1, 57 instances) reach High-severity ratios of 42.7% and 36.8% respectively.
- **CVVs are spread across projects and architectures.** Each of the six most represented projects covers at least six patterns: Optimism (196 instances, 9 patterns), Polygon (143, 11), Axelar (99, 8), Scroll (66, 7), ZKsync (64, 6) and Wormhole (54, 11). B3 accounts for 313 of their 622 CVVs (50.32%), while the residual composition is project-specific (e.g., Polygon leans toward B1, E1 and A1; Axelar toward C1 and B2; Scroll toward E1).
- **The full screened population skews even more toward B3.** Among the 1,482 screened CVVs, B3 has 705 instances (47.57%) across 31 projects, and E1 has 119 (8.03%) but appears in only 10 projects.
- **Existing tools barely cover the taxonomy.** Of the 12 assessed tools, none of the 8 traditional tools (Slither, SmartCheck, Solhint, Mythril, Securify, ConFuzzius, Oyente, Manticore) maps to any of the 11 patterns (0/11). Among cross-chain tools, SmartAxe maps to 2/11, BridgeGuard 1/11, XGuard 1/11, and ChainSniper 0/11.
- **LLM detection is inconsistent.** Composer-2.5 is best (Macro-F1 51.2%, recall 75.6%), then GPT-5.5 (F1 33.8%, precision 45.4%, recall 29.8%), DeepSeek-V4-Pro (25.3%) and Opus-4.8 (20.7%). Against the 29.3% all-pattern constant baseline, Composer-2.5 is +21.9 percentage points, GPT-5.5 +4.5, DeepSeek-V4-Pro −4.0, and Opus-4.8 −8.6; the four-model mean is 32.7% (+3.4).
- **High raw F1 can be an artifact of prevalence.** B3 has the highest raw mean LLM F1 (57.5%) but appears in 71.8% of evaluated reports, giving the constant baseline an F1 of 83.6% for that pattern, which only Composer-2.5 exceeds. Raw F1 correlates strongly with report-level prevalence (Spearman's ρ = 0.955); low-prevalence patterns such as D2 (16.1%), C2 (19.3%) and D1 (23.5%) look hard largely because they are rare.
- **LLM errors trace to missing context and weak semantics.** The paper attributes most failures to incomplete cross-component context and limited cross-chain semantic reasoning; a detailed manual false-negative/false-positive analysis is posed as RQ3, but the supplied content is truncated before that section's results.
- **Annotation was reproducible.** Two researchers with over three years of smart contract security experience screened findings independently with 92.4% agreement and Cohen's κ of 0.85, and assigned CVVs to patterns with 90.8% agreement and κ of 0.88.

## Methodology in Plain English

The authors start from two bridge aggregator sites (Chainspot and DeFiLlama, data as of May 2026), deduplicate the project list, drop projects with no public audits, and end up with 36 projects. They gather every public audit report they can find — 588 in total, from firms including Trail of Bits, OpenZeppelin, Zellic, Quantstamp and Code4rena — yielding 6,069 individual findings.

Each finding is put through a three-question filter: does it involve a cross-chain operation, does it involve data that participates in verification, and does the defect actually break verification correctness or the execution it authorizes. Only findings passing all three count as CVVs; that leaves 1,482 of the 6,069 findings (3,942 failed the first filter, 456 the second, 189 the third). Two researchers label independently and reconcile disagreements.

To build the taxonomy they use open card sorting: they write down each instance's description, affected code, trigger, exploitation mechanism, consequence and fix, group the instances by the underlying verification failure, and iterate — merging and splitting candidate patterns until they settle on 11 patterns under 5 guarantees (message authenticity, execution integrity, replay and freshness, verification soundness, trust management).

For the dataset, they try to pin each CVV to the exact source-code version that was audited, since reports may cite commits, tags, branches, pull requests or ranges. GPT-5.5 is used only to extract version clues, which humans then verify against report text, affected files, code snippets and patch records; unrecoverable cases are dropped, keeping 830 of 1,482 across 209 reports and 27 projects. They then keep the affected source files and exclude tests, scripts and libraries.

Evaluation runs in two tracks. For tools, they read each of the 12 tools' papers and documentation and manually map its stated scope onto the 11 patterns, marking a pattern only where the tool explicitly targets that vulnerability or directly captures its underlying verification failure. For LLMs, they use one shared taxonomy-guided multi-label prompt giving the auditor role, all 11 pattern definitions, the source code, analysis requirements and a JSON output schema; each model reports whether a CVV exists, and for each predicted pattern the vulnerable function, code evidence, reasoning and a short exploit path. Unparseable outputs count as negative. Because reports may contain several patterns and frequencies are skewed, they compare against an all-pattern constant baseline that predicts all 11 patterns for every report, whose F1 for a pattern of prevalence *p* is 2p/(1+p), macro-averaged to 29.3%.

## Why This Matters

**Impact on research.** Prior bridge-security work targeted specific protocols, scenarios or detection rules; this paper supplies a shared vocabulary (5 guarantees, 11 patterns) and a reproducible, version-pinned dataset, which lets future detectors be compared on identical data and lets researchers measure whether a claimed detector actually addresses verification semantics rather than generic Solidity bugs. The constant-baseline comparison is also a methodological warning: prevalence-correlated F1 can make a detector look competent when it is only exploiting class imbalance.

**Real-world applications.**

- **Pre-audit triage for bridge teams:** the 11-pattern checklist, and especially the dominance of B3 and B1, tells developers where to focus review effort on destination-chain execution paths rather than only on signature checks.
- **Audit-firm scoping:** the taxonomy's mapping from patterns to the specific files in real reports gives auditors concrete precedents, and the severity data flags which rare patterns (E1, C1) deserve escalation despite low frequency.
- **Bug-bounty and monitoring design:** because verification bugs cut across gateway, messenger, validator, relayer, decoder and execution contracts, the results argue for tooling that spans components rather than per-contract scanning.
- **Procurement of security tooling:** the documented 0/11 coverage of mainstream single-chain detectors and the at-most 2/11 coverage of cross-chain tools is a concrete basis for deciding that off-the-shelf scanners are insufficient for bridge code.

**Industry relevance.** The $330 million in cited losses, concentrated in forged messages and forged state proofs, makes this a direct financial-risk topic for bridge operators, L2 teams running canonical bridges, and insurers or risk analysts assessing bridge exposure. The evaluation of frontier LLMs (including GPT-5.5, Opus-4.8 and DeepSeek-V4-Pro) gives practitioners a realistic picture of what AI-assisted auditing can and cannot yet deliver on heterogeneous multi-language codebases (the dataset spans 490 Solidity, 159 Rust, 76 Go and 61 other files).

## Future Directions

1. **Detectors built for verification semantics rather than generic defects.** The authors explicitly call for CVV-specific detection that models the verification-to-execution path and carries cross-component context, since 69.52% of CVVs are execution-integrity failures and existing tools map to at most 2 of 11 patterns.
2. **Handling heterogeneous, multi-language bridge stacks.** With 236 of 786 mapped files in Rust or Go, and most tools limited to Solidity source or EVM bytecode, cross-language and source-plus-off-chain analysis is an open requirement.
3. **Fixing the rarity problem in evaluation and detection.** With Spearman's ρ = 0.955 between raw F1 and report-level prevalence, and low-prevalence patterns like D2 (6 reports, 2.9%) and C2 (8 reports, 3.8%) hard to measure, better sampling, reporting and training strategies for rare verification bugs are needed.
4. **Extending coverage of recovered instances and of LLM error analysis.** Only 830 of 1,482 CVVs could be pinned to an audited version, leaving 652 instances outside the dataset; and the paper's truncated RQ3 error analysis — incomplete cross-component context and limited cross-chain semantic reasoning — invites deeper study of what context (messages, validator sets, proof generators, relayer logic) an LLM or analyzer must be given to reason correctly.

## Target Audience

Smart contract security researchers and auditors; bridge and L2 protocol engineers who own verification code; developers of static-analysis, symbolic-execution and LLM-based vulnerability detectors; and risk or security analysts at firms that assess bridge exposure. It is also a useful methodological reference for empirical software-security researchers interested in building audit-report-derived datasets, since it documents screening criteria, inter-rater agreement and version-recovery procedures in detail. Readers need only moderate blockchain background; no deep cryptography is assumed.

## Authors’ abstract

With the rapid development of blockchain technology, numerous heterogeneous blockchains have emerged, each featuring distinct architectures, consensus mechanisms, and application ecosystems. Cross-chain bridges have consequently become a critical infrastructure component, typically relying on smart contracts to facilitate message passing and asset transfers across heterogeneous blockchains. Unlike single-chain contracts, bridge contracts introduce additional verification…

[The full abstract on arXiv](https://arxiv.org/abs/2610.13170)

[Read the original paper](https://arxiv.org/abs/2610.13170)

---

Canonical: https://crypto.world/paper/understanding-verification-vulnerabilities-in-cross-chain-bridges-an-2610.13170
