# The Audit Gap in Blockchain Security: A Four-Year Empirical Study of Public Audit Findings and Real-World Exploit Incidents

Crypto.world

> Blockchain security evidence is scattered: audit reports come from dozens of firms in incompatible formats, and exploit records come from a handful of trackers with inconsistent conventions.

**arXiv**: 2606.15465

**Published**: 2026-06-13

**Authors**: Stefan Beyer

## AI summary

## Overview

**Research area:** Web3 / blockchain security — empirical software engineering applied to smart-contract auditing and real-world exploit losses.

**Technical level:** Intermediate. The paper uses descriptive statistics (shares, counts, means, medians, concentration ratios) rather than formal modelling, but it assumes familiarity with audit severity tiers, DeFi protocol types and common attack classes.

**Scope:** A four-year-and-three-month empirical comparison of 23,818 public audit findings from 22 security firms against 218 documented exploit incidents totalling roughly US$7.76 billion, over 1 January 2022 to 27 March 2026.

## What This Paper Is About

Blockchain security evidence is scattered: audit reports come from dozens of firms in incompatible formats, and exploit records come from a handful of trackers with inconsistent conventions. The paper builds one aggregate view of both sides and asks whether the vulnerabilities auditors publicly report line up with the vulnerabilities that actually cause money to be lost. The answer is largely no — the two records describe different populations, and the paper quantifies exactly how they diverge.

## Key Contributions

1. **A large cross-firm audit corpus.** The author assembles and classifies 23,818 public audit findings from 22 independent security firms, drawn from both GitHub-published PDF archives and web-native publication channels (vendor APIs, CMS endpoints, report portals), and describes it as the largest cross-firm corpus of public audit findings analysed to date.
2. **A quality-filtered incident dataset.** 218 exploit incidents from rekt.news (filtered down from an initial archive of 242 entries), with aggregate losses of US$7.764 billion, each tagged with date, loss amount, chain, protocol type and a root cause re-coded into the author's own taxonomy.
3. **A measured misalignment between the two distributions.** The paper shows that the categorical distribution of audit findings does not correspond to the categorical distribution of realised dollar losses, and attributes the gap largely to the narrower scope of a conventional code audit.
4. **A characterisation of loss concentration.** Realised losses are shown to be heavy-tailed and concentrated by incident rank, by chain and by protocol type, with stated implications for risk provisioning.

## Main Findings

- **Stable audit picture.** Across all 23,818 findings, 17.2% are Critical or High (1,439 Critical, 2,659 High), 22.4% Medium, 37.5% Low and 22.9% Informational. The Critical-plus-High share stays inside a 15–17% band in every complete year (22.7% in 2022, 16.8% in 2023, 16.0% in 2024, 15.0% in 2025); the 22.2% figure for Q1 2026 is attributed partly to informational findings being backfilled to public archives later than severe ones.
- **Logic dominates audit output.** The five most frequent categories (logic/business logic 14.6%, code quality 13.0%, input validation 10.0%, access control 9.8%, initialisation/upgradeability 7.3%) make up 54.7% of findings; the top sixteen cover 92.1%, leaving 7.9% in an "Other" residual. Reentrancy is only 2.1%, and oracle/price-manipulation only 4.3%.
- **Volume more than doubled, then retreated.** Annual findings rose from 2,526 in 2022 to 7,412 in 2024, fell to 6,504 in 2025, and stood at 1,755 for Q1 2026. The author attributes much of the 2022–2024 growth to more firms publishing publicly (twelve in 2022, seventeen in 2024) rather than higher per-report detection.
- **Category drift under a stable total.** Logic findings fell from about 19% to about 11% (partly reclassification-driven), initialisation/upgradeability rose from about 4% to about 11%, oracle/price manipulation roughly doubled from about 2% to 6–7%, and reentrancy fell from about 6% to about 3%.
- **EVM dominance persists.** Solidity/EVM accounts for 79–84% of findings every year; Rust/Solana rose from about 3% in 2023 to a local peak of 8% in 2025 then 4% in Q1 2026; TON/FunC appears from 2023 at 4–5%; CosmWasm plus Cosmos SDK stay between 5% and 8%; Move reaches 4% of Q1 2026 findings.
- **Human-vector attacks dominate losses.** Private-key compromise (US$1,894M, 24.4%), phishing/social engineering (US$1,511M, 19.5%) and access control (US$994M, 12.8%) together account for 56.7% of cumulative losses. The four human-vector categories combined (adding supply chain and governance) total US$3,849M, or 49.6% of losses across 31.7% of incidents, and supplied the majority of annual losses in 2023 (65.6%), 2024 (74.6%) and 2025 (70.8%) — with 2022 the exception at 17.3%.
- **Only one category bridges both top-four lists.** Access control is the only category in the top four by audit frequency and by exploit losses. Logic errors, code quality and input validation are 37.6% of audit output but only 3.8% of realised dollar losses; conversely private-key compromise, phishing and supply chain are about 47% of losses but a negligible share of findings.
- **Extreme concentration.** The single largest incident (Bybit, February 2025, US$1.44B) is 18.4% of cumulative losses; the eight largest are 50.6% and the twenty largest 71.4%, while the remaining 198 incidents (91% of the count) hold under 29% of losses.
- **Chain concentration.** Ethereum and BNB Chain host 89% of incidents (199 of 218) and 94% of losses (US$7.35B of US$7.76B). The 2025 Cetus exploit on Sui (US$223M) is the first substantive sign that a single large loss can shift a chain's risk profile.
- **Protocol type beats vulnerability class as a tail-loss predictor.** Centralised exchanges and bridges are about 60% of cumulative losses across 21% of incidents (CEX: 24 incidents, US$2,284M; bridges: 22 incidents, US$2,401M). Lending is the most frequent protocol type by incident count (48) yet under 10% of losses (US$736M).
- **The "audited-but-exploited" pattern.** 105 of 218 incidents (48%) hit protocols with at least one prior public audit, accounting for about US$4.30B, or roughly 55% of losses. Of the twelve largest such incidents, nine have root causes outside conventional audit scope; the other three (Nomad, Euler, Qubit) involved post-audit code changes, out-of-scope paths or paths not fully enumerated.
- **Means mislead.** Phishing has a mean of US$125.9M against a median of US$6.5M (ratio about 19); bridge exploits have a mean of US$148.3M against a median of US$3.5M (ratio about 42); private-key compromise has a mean-to-median ratio of about 5. The author notes the private-key median is inferred from the reported ratio.

## Methodology in Plain English

The author merged two evidence streams rather than sampling one population.

For audits, reports were pulled from firms' public GitHub archives and from machine-readable web channels. For repository sources, the publication date is the date a report was first committed, obtained with `git log --diff-filter=A`, so that old reports sitting in a repo are not counted as new. Findings were then sorted into about twenty categories in three passes: title pattern matching, then reclassification using surrounding report text, then targeted large-language-model reclassification for leftover Critical/High findings and the "Other" bucket. The "Other" rate on the full dataset is 7.9%, and web-native findings classify more cleanly than PDF-extracted ones because their titles are not truncated or line-broken.

For incidents, entries came from rekt.news with written permission. Each was tagged with date, loss, chain, protocol type and root cause, with root causes re-coded into the author's own taxonomy rather than rekt.news's free text. A quality filter kept only entries with a present, theft-consistent loss amount and excluded rug-pulls and editorial entries, narrowing 242 entries to 218.

The two datasets are explicitly presented in parallel, not as comparable samples: audit findings are vulnerabilities found in reviewed code, incident records are vulnerabilities exploited in deployed systems that may or may not have been audited. Firm identities are withheld and no firm-level results are reported. Construction and classification were AI-assisted under human editorial oversight, with methodology and quality thresholds set by the author; analytical code and intermediate datasets are retained and available on request.

## Why This Matters

**For research:** The paper supplies a multi-year, cross-firm empirical baseline that complements tool-evaluation and taxonomy work. It aligns with prior results — Perez and Livshits finding that of 23,327 tool-flagged contracts fewer than 2% were ever exploited, and Chaliasos et al. finding that five widely used tools would have prevented only a minority of 127 high-impact attacks — and extends them from single contracts to the whole public record.

**Real-world applications:**

- **Audit scoping and engagement design:** teams can see which finding categories their code review reliably surfaces and which loss categories require separate operational-security work.
- **Risk provisioning and underwriting:** the concentration figures (eight incidents = 50.6% of losses, twenty = 71.4%) argue against Gaussian loss models, which the author states will systematically underestimate worst-case annual outcomes.
- **Key-management and signer-workflow review:** the Bybit, Ronin, DMM Bitcoin, WazirX and Poloniex incidents show where the largest losses came from, and the approval interface in the Bybit case was not itself audited.
- **Multi-chain security planning:** the non-EVM tail (Rust/Solana, TON/FunC, CosmWasm, Cosmos SDK, Move) is small but changing shape, so teams assuming single-stack audit availability accumulate coverage gaps.

**Industry relevance:** The core message is a portfolio view — audits for bugs in the code as written, and parallel assessments of key management, signer workflows, build pipelines and dependency supply chains for the categories that drive the larger share of realised losses. Audits are described as narrowly effective but broadly incomplete relative to the risk surface they are often implicitly relied on to cover.

## Future Directions

- **Extend operational-security measurement.** The paper documents that human-vector attacks dominate losses from 2023 onward but does not evaluate operational-security engagements the way prior work evaluated static analysis tools; measuring their effectiveness is the obvious next study.
- **Track the non-EVM expansion.** Whether incident concentration in Ethereum and BNB Chain (89% of incidents, 94% of losses) reflects genuine resilience or merely lower total value locked elsewhere is left open; the Cetus incident is offered as a partial test case, and continued monitoring would resolve it.
- **Address classification drift.** Part of the decline in "logic" findings is attributed to reclassification rather than real change, and 7.9% of findings remain in "Other" — a cross-firm, standardised finding taxonomy would make longitudinal comparison cleaner.
- **Correct for backfill lag.** Informational and Low findings reach public archives later than severe ones, which distorts partial-year readings such as Q1 2026; modelling or correcting this lag would improve early-year comparisons.

## Target Audience

Security auditors and audit-firm leads who need to know what their reports do and do not capture; protocol risk, treasury and underwriting teams sizing tail exposure; operational-security and key-management practitioners at exchanges, bridges and custodians; and academic researchers in empirical software engineering or blockchain security who want a multi-year aggregate dataset and a clear statement of its limits. Readers wanting incident-level forensics or firm-level comparisons will not find them here — the analysis is deliberately aggregate and firms are not identified.

## Authors’ abstract

This paper presents an empirical analysis of the Web3 security landscape over the four-year and three-month period from 1 January 2022 to 27 March 2026. The dataset combines 23,818 public audit findings produced by 22 independent security firms with 218 real-world exploit incidents documented by rekt.news, representing aggregate losses of approximately…

[The full abstract on arXiv](https://arxiv.org/abs/2606.15465)

[Read the original paper](https://arxiv.org/abs/2606.15465)

---

Canonical: https://crypto.world/paper/the-audit-gap-in-blockchain-security-a-four-year-empirical-study-of-2606.15465
