# Attack Patterns in Cross-chain Bridges: A Trust-based Empirical Analysis

Crypto.world

> Releases a 48-incident bridge-attack dataset and trust-violation taxonomy that maps each exploit to the assumption broken, giving auditors and protocol designers a reusable checklist.

**arXiv**: 2611.01587

**Published**: 2026-11-01

**Authors**: Ishraq Samiha, Yasir Rahman, Rahma Mukta

## AI summary

## Overview

**Research area:** Blockchain security, specifically the security of cross-chain bridges and cross-chain messaging protocols.

**Technical level:** Intermediate. The taxonomy and coding rules are readable without deep protocol expertise, but familiarity with how bridges lock, mint, relay and verify assets helps.

**Scope:** An empirical study of 48 evidence-supported real-world bridge attacks from January 2022 to August 2026, classified by the trust assumption each attack violated and by the bridge function where that trust boundary was breached.

## What This Paper Is About

Cross-chain bridges hold large amounts of locked value and depend on many moving parts, so attackers repeatedly find ways to make a bridge accept a transfer that never happened or to seize direct control of its assets. Existing studies organise bridge attacks mainly by vulnerability type, architecture, detection technique or individual incident, and rarely separate *how* a trust assumption is broken from *where* in the bridge it is broken. This paper builds a behaviour-based attack pattern architecture that does both, then measures how those patterns, functional domains and financial losses are distributed across 48 consistently coded incidents.

## Key Contributions

1. **Updated, auditable incident dataset.** 48 evidence-supported bridge incidents from January 2022 to August 2026, each recording gross bridge-specific loss, root cause, evidence level (E1–E4), pattern and domain labels with rationale, boundary-case flags, and a published codebook and analysis scripts (DOI: 10.5281/zenodo.23271026).
2. **Six trust-violation attack patterns in three groups.** Proof Forgery (PF), Logic Exploitation (LE), Event Injection (EI), Economic-State Manipulation (ESM), Credential Seizure (CS) and Infrastructure Hijacking (IH), grouped as Verification and Contract Logic (PF, LE), Cross-Chain Information and Economic State (EI, ESM) and Operational Trust Compromise (CS, IH). Each pattern names the trust assumption the attacker breaks, with explicit rules for choosing one primary pattern per incident.
3. **Separate functional-domain mapping.** Five domains — Deposit (FD-1), Relay (FD-2), Verification (FD-3), Execution (FD-4), and Governance and Operations (FD-5) — record where the compromised trust boundary sits, independent of the pattern.
4. **Empirical and sensitivity analysis.** Distribution of patterns, domains and losses over time, plus tests of how sensitive the results are to ambiguous classifications and to the choice of loss basis, together with a per-pattern discussion of mechanism, signs of attack and mitigations.

## Main Findings

- **Corpus size and losses:** 48 incidents across 43 distinct systems, totalling **$2.676 billion** in gross bridge-specific losses. The ten largest incidents account for **92.4%** of total losses.
- **Typical vs. average loss:** Median incident loss is **$4.25M** (IQR **$1.18M–$15.04M**) against a mean of **$55.7M**, showing heavy concentration.
- **Two dominant patterns:** Proof Forgery and Credential Seizure each account for **13 incidents** and together **54.2%** of incidents and **86.2%** of total losses, measured on the primary loss basis. CS has the highest median loss at **$10.7M**; PF accounts for **$1,233.4M (46.1%)** of losses with a median of **$7.00M**.
- **Logic Exploitation is frequent but cheaper:** LE ranks third with **9 incidents (18.8%)** but only **$270.4M (10.1%)** of losses, median **$3.30M**. Event Injection has **7 incidents (14.6%)** and **$90.5M (3.4%)**.
- **Two domains dominate losses:** On the primary loss basis, the **Verification** and **Governance and Operations** domains hold **93.9%** of all losses.
- **Concentration is robust:** The abstract reports that the concentration holds when switching to other loss bases or reclassifying boundary cases.
- **2022 dominates financially:** 2022 accounts for **70.8%** of total losses — 10 incidents totalling **$1,894.8M**, median **$90.0M** (IQR $3.3–292.0M). The 2023 total of $213.4M is consistent with the $1.5–2 billion figure cited for 2022 in prior work.
- **Loss sizes fell after 2023:** Median losses were $90.0M (2022) and $86.8M (2023), then dropped to **$2.7–$4.3M** during 2024–2026, while incident counts stayed at 8 in 2024 and 8 in 2025.
- **2026 has the most incidents:** **19 incidents** and **$427.5M** in January–August 2026 alone, more than any full earlier year; the authors caution that exposure data (number of active bridges, liquidity) is unavailable, so this cannot be read as rising attack frequency.
- **Pattern composition shifts over time:** 2022 incidents spanned five patterns; Logic Exploitation was most frequent in 2024; Proof Forgery was most frequent in 2026 with **8 of its 19 incidents**. Credential Seizure is the only pattern observed in every year.
- **Evidence quality:** 21 incidents are E1 (official post-mortem or fix plus an independent source), 23 are E2 (two or more independent sources), 3 are E3 (official report only) and 1 is E4 (third-party and on-chain evidence only, with classification inferred). Over 90% are E1 or E2.

## Methodology in Plain English

The authors built the incident list from four kinds of sources: academic literature (searched across IEEE Xplore, ACM Digital Library, Scopus, Web of Science and Google Scholar), security-firm reports (Halborn, CertiK, SlowMist, BlockSec), project post-mortems and fix commits, and on-chain data from block explorers and incident trackers. They also searched reference lists backwards to catch missed cases, merged duplicate records, and screened 60 candidates down to 48 by excluding exchange or custodial hacks, non-bridge components of bridging protocols, multi-chain apps, incidents with no bridge loss, prevented attempts, and anything starting before 1 January 2022.

Each incident needed evidence: an official post-mortem, two or more independent sources, or corroborated on-chain evidence with third-party reporting. Losses were normalised as **gross bridge-specific loss** in USD at the time of attack — no deduction for recoveries, bounties or returned funds, and the lower bound taken whenever sources gave a range. Data fields separate the first-reported loss, the normalised loss, a note on any adjustment, and a "loss basis" field stating what the number actually measures (bridge assets withdrawn, wallet funds drained, value of unbacked minted assets, real assets acquired with unbacked tokens, or reported with no stated basis).

Classification used a fixed codebook developed in four phases (concept generation, refinement, codebook fixing, coding). Every incident gets exactly one primary pattern — the attacker behaviour that directly caused the compromise — and one primary domain — the component whose trust was broken. Supporting techniques such as flash loans or phishing are recorded in the root-cause and rationale fields; a flash loan is coded as ESM only if it manipulated an economic state the bridge relied on. Boundary rules resolve overlaps (for example, a flawed or bypassable verifier is PF, while an initialisation or upgrade flaw is LE; stolen credentials that directly authorise a release are CS, while credentials that merely shift what the verifier trusts are PF). One author coded each incident and a second checked it; because the codings were not fully independent, no agreement score is reported.

## Why This Matters

**Impact on research.** The paper supplies an incident-level coding scheme that treats the violated trust assumption as the unit of classification and keeps it separate from the affected bridge function, with an explicit loss-normalisation rule and published codebook. This lets studies compare different bridge attacks on the same terms and makes the corpus auditable and reproducible across a longitudinal window, which the authors identify as a gap in architecture-focused, taxonomy-focused, detection-focused and forensic-focused prior work alike.

**Real-world applications:**

- **Threat modelling for bridge designers:** mapping each known attack to a pattern and domain gives teams a checklist for hardening verification logic, relay inputs, economic state and the governance/operations layer.
- **Audit and review scoping:** auditors can prioritise the Verification and Governance and Operations domains, which together hold 93.9% of losses in this corpus.
- **Incident triage and post-mortem writing:** the pattern definitions and "signs that reveal it" descriptions give a common vocabulary for describing what went wrong.
- **Detection tooling:** runtime monitors such as BridgeGuard, XChainWatcher, BridgeShield and XScope could label alerts by trust-violation pattern and domain, adding an interpretive layer to existing detection.

**Industry relevance.** The finding that credential seizure carries the highest median loss ($10.7M) and that Proof Forgery and Credential Seizure together explain 86.2% of losses points industry attention at key management, validator-set control, upgrade authority and operational infrastructure, not only at smart-contract bugs. The dominance of a handful of very large incidents (ten incidents at 92.4% of losses) also matters for insurers, risk desks and treasury teams sizing bridge exposure.

## Future Directions

- **Extend and rebalance the corpus.** Several categories have very few cases, 2026 covers only January–August, and the taxonomy was built from the same corpus it is applied to; independent replication on a wider incident set would test external validity.
- **Add exposure denominators.** The authors note the absence of measures such as how many bridges were active or how much liquidity they held, which is needed before any claim about attack frequency over time.
- **Refine multi-failure incidents.** Each incident currently receives a single primary pattern, even when several behaviours combine (TeleSwap, E4, had to be classified from scarce public evidence); multi-label or weighted coding is an open design question.
- **Standardise loss measurement.** Unbacked minting can be valued by what was minted or by the real assets obtained; the paper records which basis was used and tests alternatives, leaving a shared convention for the field as an open problem. The full results of the sensitivity analysis are not reported in the excerpt reviewed here.

## Target Audience

Bridge and interoperability protocol developers, smart-contract auditors, and blockchain security researchers who want an empirically grounded classification scheme rather than a vulnerability list. It also suits incident-response and forensic teams, risk and insurance analysts assessing bridge exposure, and graduate students studying DeFi security, since the methodology section explains coding rules, evidence levels and loss normalisation in enough detail to be reused. Detection-tool builders will find the pattern-to-domain mapping useful, though readers seeking formal proofs or protocol-level cryptographic constructions will not find them here.

## Authors’ abstract

The rapid growth of multiple blockchain ecosystems has made cross-chain bridges essential for transferring assets and information across independent blockchains. Their multi-component architectures, cross-chain trust assumptions, and concentration of valuable assets expose them to diverse security threats. This paper presents an empirical study of cross-chain bridge attacks based on 48 evidence-supported real-world incidents from January 2022 to August 2026. We propose a…

[The full abstract on arXiv](https://arxiv.org/abs/2611.01587)

[Read the original paper](https://arxiv.org/abs/2611.01587)

---

Canonical: https://crypto.world/paper/attack-patterns-in-cross-chain-bridges-a-trust-based-empirical-analysis-2611.01587
