betaFeedback

Reading v1View latest version

WorksComputer Science

TxBench: A Benchmark of Attack Transactions and Evidence-Backed Negatives on Seven EVM Chains

Abstract

Researchers evaluate detectors of malicious smart-contract transactions on datasets that each paper assembles anew. Their positives come from incident trackers, whose citations mix attacks with setup steps and transactions that never ran, or from a detector's own output, which makes the evaluation circular; their negatives are ordinary transactions. A shared benchmark needs positives whose own execution shows the attack and negatives that use the attacks' mechanisms, with no detector deciding either label. We present TxBench, a benchmark of attack transactions and evidence-backed negatives on seven EVM chains, with the five-step pipeline that builds it. The pipeline collects every transaction that BlockSec's incident library or the EVM Hack Registry cites, maps its weakness to OWASP's SCWE taxonomy, and rebuilds its execution from free data sources. A cited transaction qualifies as a valid TxBench attack only if it extracts value from others during execution, is consistent with all verified source claims, and is independently linked to the reported incident by corroborating evidence. Negative examples are sampled from each chain's traffic near the attacks and retained only when positive evidence confirms their legitimacy. Of 1,850 cited transactions, TxBench accepts 1,012 attacks from 602 incidents, worth USD 868.1 million: 846 contract exploits and 166 key compromises. A blind review under rules fixed in advance confirms 98 of 100 sampled attacks (97.7% weighted) and all 100 sampled negatives. As baselines, four zero-shot LLMs reading anonymized traces are precise but find at most 0.43 of the attacks and no key compromise, leaving ample room for better detectors.

Introduction

Attackers carry out smart-contract exploits through transactions [1]. DeFiLlama records 967 hacks of decentralized finance (DeFi) protocols worth USD 9.29 billion, 119 of them in 2025 [2]. Code analysis, rule-based or learned [3, 4], flags far more than attackers exploit: attackers exploited only 1.98% of the 23,327 contracts that six academic projects flagged as vulnerable [5]. Defenses therefore also inspect the transaction itself, with replay and rules [6, 7, 8], by lifting traces to DeFi actions [9, 10, 11], or with learned models of traces [12, 13, 14, 15, 16, 17].

Transaction-level detectors lack a shared ground truth. Code and agent benchmarks label contracts or vulnerabilities, not executed transactions [18, 19, 20, 21]. A benchmark of attacks alone also cannot measure false positives, which practitioners name a main barrier to adopting analyzers [4]. Transaction-level papers therefore build their own sets: DEFIER [8] and Horus [7] label the attacks their own analyses find, and Alhaidari et al. [22] and BridgeGuard [14] curate 402 and 203 attack transactions, BridgeGuard against 40,000 normal ones. These sets take one or more of three shortcuts: labels from a detector make the evaluation circular, labels copied from incident trackers inherit their errors, and random negatives reward flagging every large transaction.

The Euler Finance exploit of March 2023 shows what these shortcuts hide. BlockSec’s incident library lists six drain transactions and a loss of USD 200 million [23]. The EVM Hack Registry [24] cites only one of them, with a loss of about USD 197 million. On their own execution, the six transactions take USD 198.1 million together and the registry’s one takes USD 8.7 million, so a dataset built from the registry would credit one transaction with the whole loss. Euler’s attacker also used only calls that legitimate users make, so a fair test needs negatives that make them too.

We present TxBench, a benchmark of attack transactions and evidence-backed negatives on seven chains that run the Ethereum Virtual Machine (EVM), and the five-step pipeline that builds it from public data (Figure 1). Steps ❶–❸ collect the 1,850 cited transactions, map their weaknesses to OWASP’s Smart Contract Weakness Enumeration (SCWE) [25], rebuild their execution, and accept the 1,012 that take value from others, contradict no checked claim, and have evidence beyond their citation. Step ❹ keeps 2,024 negatives with positive evidence of legitimacy, and step ❺ has a person review samples of both classes blind. No detector decides any label, our own TxLens [17] included. Five splits test generalization by holding out incidents, time, protocols, attackers, or chains. As baselines, four zero-shot LLMs find at most 0.43 of the attacks.

TxBench pipeline

Figure 1. The TxBench pipeline: steps ❶–❸ accept attacks, step ❹ selects negatives, and step ❺ reviews both. Numbers count transactions.

Building TxBench showed how far the public record is from ground truth. The two sources share only 197 of the

ETHBSCARBBaseOPPOLAVAXTotal
Incidents (union of the sources)4054066763212814–
Cited attack txs9045601831074336171,850
Accepted attack txs5443094365251791,012
value taken (M USD)692.467.830.412.542.310.812.0868.1
Negatives (every target met)1,088618861305034182,024

Table 1. TxBench by chain: Ethereum, BSC, Arbitrum, Base, Optimism, Polygon, Avalanche; values rounded per chain.

1,850 transactions (10.6%), and five of the registry’s cited hashes exist on no chain. Only 72.2% of the cited transactions with a full call tree take at least USD 100 from others; the rest include setup steps, follow-ups, and attacks whose gain is unpriced. Random blocks near the attacks are neither clean nor hard: 2.0% of the 11,293 traced candidates from them involve an attacker or a main victim of a cited incident, and only 1.0% of their transactions use more than a million gas, against 54.5% of the accepted attacks. Datasets that copy the sources’ citations or draw random negatives carry these flaws into every evaluation.

From Incident Citations to Accepted Attacks

Steps ❶–❸ turn incident citations into attacks that their own execution confirms; Table I gives their results per chain.

Two Public Sources, Matched on the Transaction

No public source covers every incident, so step ❶ collects the attack transactions that two of them cite. BlockSec’s Security Incidents Library documents incidents with their transactions, attackers, root causes, and losses [23]; our copy holds its 308 incidents up to id 768, dated from 2023-01-03 to 2026-07-24. The EVM Hack Registry extends DeFiHackLabs [26] with 932 dated incidents from 2017 to 2026 at commit 5e310fc, each a Foundry proof of concept with a report that cites its transactions and parties [24]. TxBench matches the sources on the transaction hash and keeps the seven chains with more than ten incidents in their union: Ethereum, BNB Smart Chain (BSC), Arbitrum, Base, Optimism, Polygon, and Avalanche. Hacks on these chains alone account for 75.4% of the DeFi losses that DeFiLlama records [2]. The result is 1,850 cited transactions: 197 in both sources, 780 in BlockSec’s only, and 873 in the registry’s only. On chain, 1,840 succeeded, five reverted, and five do not exist.

Weaknesses in a Shared Taxonomy

To compare the sources’ weakness labels, step ❶ maps each incident to OWASP’s SCWE v1.0 [25], a taxonomy of 156 weaknesses. Claude Opus 5.5 read each incident’s root cause under a fixed codebook, which the artifact ships, with one rule per class of attack. For the 844 registry incidents, the model read the report, and for the others, the traces, a post-mortem, or BlockSec’s label. A reading names a specific entry, a related entry, a general class, or none, and each transaction takes its incident’s reading; Euler’s six share SCWE-125, a missing health check. Overall, 1,166 of the 1,850 transactions (63.0%) map to a specific entry, and 228 exploit no contract flaw. An author read 30 incidents blind under the codebook and matched the level for 28 (Cohen’s κ=0.90\kappa= 0.90 [27]) and the entries for 26.

Execution Records from Free Data Sources

A receipt shows a transaction’s events but not the internal calls where an attack happens. Step ❷ therefore builds an enriched trace of each transaction: its receipt, decoded call and events, and internal call frames. To stay rebuildable without paid services, the builder reads only public endpoints, free API tiers and tracers, Sourcify, and public signature databases. BSC and Arbitrum offer no free tracer, so the builder replays the block in a local Anvil fork and keeps the replay only when its status and logs equal the chain’s. Of the 1,845 on chain, 1,797 (97.4%) have the full call tree that step ❸ needs.

A Value Oracle Instead of a Detector

A citation is a claim, and so are the attacker, victim, date, and loss that a source names. Step ❸ therefore judges each transaction by its execution. The value oracle computes each address’s net change per asset from the trace, prices it in USD an hour before the block [28], and measures what the attacker side takes from others. The attacker side holds the named attackers, the sender of an exploit, the contracts they create, and wallets that only receive from them. Of the 1,797 judged transactions, 1,297 (72.2%) take at least USD 100.

Four checks then compare each value-taking transaction with the sources’ claims: the sender of an exploit is a named attacker, a named victim loses value, the block falls within seven days of BlockSec’s incident time or a month of the registry’s, and the incident’s value is within a factor of ten of the reported loss. Of these, 123 fail a check, most often the loss check (62), whose figure can cover a whole campaign.

Passing the sender and loss checks confirms little, because a source usually takes its attacker and loss from the cited transactions themselves. An accepted attack therefore also needs one of five kinds of evidence beyond its citation: both sources cite it, the registry’s proof of concept forks within an hour of it, the registry names its sender an attacker, a victim that the registry names loses value, or a source names the thief who gains from a key compromise. The 1,012 transactions that take value, fail no check, and have evidence are TxBench’s accepted attacks: 846 exploits and 166 key compromises. The 162 without evidence, all cited by BlockSec alone, form a secondary set.

Evidence-Backed Negatives from Each Chain’s Own Traffic

To measure precision as well as recall, step ❹ adds negatives where attacks and legitimate activity look alike. TxBench keeps discovery, validation, and sampling apart, so sampling never lowers the evidence bar to fill a quota.

Targets and Discovery

TxBench targets two negatives per accepted attack on each chain, 2,024 in all. Since 18.4% of the accepted attacks run their exploit in the constructor of a contract they deploy, each chain’s negatives include two contract deployments per such attack. The rest split 50/30/20 over three categories: behaviorally hard (swaps, arbitrage, liquidations, and other DeFi operations), financially hard (values at or above the 25th percentile of the chain’s attack values), and verified ordinary (transfers, approvals, and simple calls).

Candidates come from each chain’s traffic, not from pairing with single attacks: TxBench draws seeded random windows within 30 days of the chain’s attacks, both single blocks and windows that target what random blocks rarely hold, such as liquidations and contract creations. The 8,091 windows hold 316,235 successful candidates; TxBench traces a stratified 14,994 with the attacks’ builder, so both classes share fields; one trace fails to decode.

Positive Evidence, Not Absence from the Sources

A known protocol proves nothing alone, since attackers also route loot through Uniswap and borrow from Aave, and neither does absence from the incident sources. TxBench therefore labels a candidate tier 1 only when it passes every applicable rule. The candidate must have succeeded with a full call tree, and no third party may lose value that its events do not explain, beyond a small tolerance. Every venue that emits a market event in it must be one of 300 deployments of 48 protocols that TxBench checked on chain, or a pool or market that such a deployment lists. The sender may gain at most 5% of the volume (15% for a liquidation), and the candidate must involve no cited transaction, no named or oracle-found attacker, and no incident’s main victim near the incident. A missing piece of evidence makes a candidate tier 2, which TxBench ships separately, and a contradiction makes it tier 3. Of the 14,993 candidates with a trace, 7,715 (51.5%) are tier 1, 4,005 tier 2, and 3,273 tier 3.

Sampling under Diversity Caps

Step 4 samples tier-1 candidates only, in cells of chain, category, and year. To match the attacks’ complexity, it draws each behavioral and financial cell evenly across the quartiles of the chain’s attack gas use, and diversity caps limit the negatives per sender, called contract, block, and protocol. Since 43.9% of the accepted attacks come from senders with fewer than 10 prior transactions, each cell takes such fresh senders in the same share, with 28 flagged exceptions. The result is 2,024 negatives: 372 deployments, 826 behavioral, 496 financial, and 330 ordinary, at a median of 2.9 days from an attack on the same chain.

Measuring Label Quality before Trusting It

Rules label both classes, so label errors bound every metric that a detector reaches on TxBench. Step 5 therefore has a person review stratified samples of both classes blind, under decision rules that we wrote and hashed before drawing them.

ModelPRF1AUCVictimLossRaw F1
Gemini 3.1 Flash Lite0.930.430.590.760.250.210.44
Qwen 3.7 Flash0.670.040.080.680.010.010.04
DeepSeek V4 Flash0.880.240.380.740.220.200.43
GLM 5.3 Flash0.810.290.420.720.220.180.66

Table 2. Zero-shot LLMs on 193 reviewed transactions, anonymized: precision, recall, F1, AUC, the share of attacks for which each names a victim and the loss (within 2×), and F1 on raw traces.

The rules ask one question of each transaction: does it take, or directly enable taking, assets or control from a victim without consent, through a contract flaw, a stolen key or signature, or deception? A reviewer answers attack_yes or attack_no, each under a named rule, or unsure. Explorer tags and incident lists, TxBench’s sources included, are pointers, never evidence.

The sample holds 100 accepted attacks and 100 negatives, stratified by chain and period and interleaved without labels in one sheet. The reviewer read a seeded fifth of the rows again at least seven days later, without seeing the first decision, and settled each disagreement against the rules. We estimate label accuracy as the stratified share of rows that the review confirms, counting unsure as not confirmed, with a Wilson interval [29] at the effective sample size [30].

Weighted by stratum, the review estimates that 97.7% of the accepted attacks (95% interval 92.5–99.3%) and 100% of the negatives (96.3–100%) carry the right label. The two readings agree on 87.5% of the re-read rows (κ=0.75\kappa= 0.75, interval 0.55–0.95), and adjudication settled the five disagreements. The two rejected attacks, a key-compromise transfer and a deployment that performs a transfer, keep the rules’ labels; every reviewed row’s decision ships, so users can drop them.

Zero-Shot LLM Baselines

The baselines are deliberately zero-shot LLMs: TxBench evaluates existing detectors as they are, so its baselines train nothing, and an LLM can judge a trace as shipped. We ask four such models to judge the reviewed transactions against the review’s final decisions, so the two attacks that the review rejects count as negatives. Each prompt holds a transaction’s trace file as JSON, untruncated, with every address replaced by a label consistent within the transaction and the hashes, block, and entity labels removed, so that a model cannot recognize a known incident by its identity. With a token table and the 156 SCWE entries as separate context, the model answers whether the transaction is an attack, which addresses the attacker controls and which lose value, the USD taken, the attack type, and the SCWE entry. The 193 transactions whose prompts fit every model’s context, 91 attacks (77 exploits and 14 key compromises) and 102 negatives, form the test; seven exploits with thousands of internal calls do not fit. Answers vary between runs at temperature zero, so the artifact stores them; 13 that ran out of output tokens count as misses.

The models are precise but miss most attacks (Table II); Gemini 3.1 Flash Lite has the best F1, 0.59. None catches a key compromise, and even the best finds only 0.51 of the exploits. When a model flags an attack, it always names an attacker address, a low bar since the sender is one in 85% of the attacks. In most attacks they flag, three of the four models also name a victim, so what the models lack is recall. On the unanonymized traces, GLM 5.3 Flash rises from F1 0.42 to 0.66, suggesting it remembers known incidents.

Storage, Use, and Research Directions

Three CSV files form the benchmark: the accepted attacks with their sources, values, checks, and evidence; the negatives with their categories, evidence, and tiers; and combined.csv, which joins both classes for evaluation. Each row of combined.csv names the transaction, its label, category, and mechanism tags, its incident, protocol, and attacker groups, and its five splits, with no group on both sides of a split; the data dictionary defines every column.

Exploits form the primary evaluation and key compromises a separate track, selected by the category column, since most key compromises are plain transfers (C4). Each transaction’s enriched trace sits in one flat folder of the benchmark, named by its hash, so that a trace’s path never reveals its class. A script, evaluate.py, scores a detector’s predictions per track and split, weights incidents equally, and, given the share of attacks in traffic, reweights precision.

All outputs, including the traces and the caches, take 5.4 GB, and each step keeps its evidence. From these caches, run_pipeline.py --offline rebuilds every output byte for byte in standard-library Python, showing the pipeline is deterministic; an online rerun requeries the sources.

TxBench supports four lines of research. Realistic negatives: categories, tags, and tiers measure precision when legitimate transactions use the attacks’ mechanisms. Generalization: the temporal split trains on the 531 attacks before 2025 and tests on the 335 of 2025 and 146 of 2026, and the cross-chain split tests on the 159 outside Ethereum and BSC. Weakness- and value-aware recall: SCWE entries and the value per attack, up to USD 118.1 million, support recall per weakness and weighted by USD. Reliability of public incident data: the oracle’s checks record where a source contradicts the chain. As a new resource, TxBench has no usage history.

Limitations and Improvements

Internal validity. I1. The review has one reviewer, an author, so its agreement measures consistency only, and with 100 rows per class the attacks’ interval stays wide (92.5–99.3%). I2. A registry of 300 protocol deployments which venues TxBench recognizes, so no tier-1 negative trades through an unrecognized venue, while 210 accepted attacks (20.8%) do. The tier-2 negatives include such venues: a detector that scores higher on tier 1 than on tier 2 has likely learned the registry. I3. The SCWE mapping reads reports, not exploit code; BlockSec’s coarse label alone decides 128 transactions, and the blind check covers only 30 incidents.

External validity. E1. TxBench holds the attacks that two public sources cite, so incidents that both miss are absent, and

BlockSec’s library starts in 2023, so only 77 accepted attacks are older; the other public trackers we examined list incidents without their transactions or block automated reading. E2. Free data sources limit what TxBench observes: the oracle judges only the 1,797 cited transactions with a full call tree. E3. Acceptance needs evidence beyond a citation, which mostly the registry supplies, so the accepted attacks lean toward its incidents; the 162 BlockSec-only secondary attacks measure the lean.

Construct validity. C1. The oracle reads one transaction at a time, so a campaign’s value splits over its transactions, and unpriced assets make values lower bounds. C2. Two negatives per attack is a sampling choice, so precision on TxBench does not transfer to traffic without reweighting. C3. Certifying legitimacy excludes transactions that pay their sender, such as reward claims, so value-aware detectors may score better on tier 1 than in traffic; the tier-2 negatives include such transactions. C4. Of the 166 key compromises, 121 are plain transfers that no trace can catch, and BtcTurk’s 153 are 15.1% of the attacks; weight incidents equally.

Improvements. Certified negatives that pay their sender, multi-transaction attacks, and a wider protocol registry matter most.

Code and agent benchmarks, such as SmartBugs, DApp-SCAN, EVMbench, and CyberChainBench, label contracts or vulnerabilities [18, 19, 20, 21]. Incident trackers cite attack transactions, but none checks every citation on chain. The per-paper sets of Section I label by a detector or a curated list. Only TxBench has all four properties of Table III.

ResourceUnitOn chainTracesValueNegatives
Code and agent benchmarks [18, 19, 20, 21]code✗✗✗✗
Incident trackers [23, 24, 26]incident○✗○✗
Per-paper sets [7, 8, 14, 22]transaction○✗○○
TxBenchtransaction✓✓✓✓

Table III. TxBench and related resources: transactions confirmed on chain, traces, value per transaction, and evidence-backed negatives (○: partly).

Conclusion

TxBench gives transaction-level attack detection a benchmark of 1,012 attacks and 2,024 evidence-backed negatives on seven EVM chains, labeled by execution and evidence, not by a detector, and checked by a blind review. The pipeline rebuilds TxBench from public data, so TxBench can grow with each release of its sources.

Data Availability: We archive the dataset, pipeline, tests, and review protocol on Zenodo at doi:10.5281/zenodo.23758032, with files restricted until publication; reviewers can access them through its share link, zenodo.org/records/23758032. The code is MIT-licensed and our annotations CC BY 4.0; third-party material, such as BlockSec’s library and the registry’s proofs of concept, keeps its source’s terms, per the artifact’s dataset notice.

References

References

  1. [1]L. Zhou, X. Xiong, J. Ernstberger, S. Chaliasos, Z. Wang, Y. Wang, K. Qin, R. Watterhofer, D. Song, and A. Gervais, “SoK: Decentralized finance (DeFi) attacks,” in Proceedings of the IEEE Symposium on Security and Privacy (S&P), 2023, pp. 2444–2461.arxiv.org/abs/2208.13035
  2. [2]DeFiLlama, “DeFiLlama: DeFi hacks and exploits database,” https://defillama.com/hacks, 2026.
  3. [3]T. Abdelaziz and A. Hobor, “Smart learning to find dumb contracts,” in Proceedings of the 32nd USENIX Security Symposium (USENIX Security 23), 2023, pp. 1775–1792.
  4. [4]T. Abdelaziz, M. Alsaghiir, and K. Ali, “Where do smart contract security analyzers fall short?” in Proceedings of the 23rd International Conference on Mining Software Repositories (MSR), 2026, pp. 49–61.DOI
  5. [5]D. Perez and B. Livshits, “Smart contract vulnerabilities: Vulnerable does not imply exploited,” in Proceedings of the 30th USENIX Security Symposium, 2021, pp. 1325–1341.
  6. [6]M. Zhang, X. Zhang, Y. Zhang, and Z. Lin, “TxSpector: Uncovering attacks in Ethereum from transactions,” in Proceedings of the 29th USENIX Security Symposium, 2020, pp. 2775–2792.
  7. [7]C. Ferreira Torres, A. K. Iannillo, A. Gervais, and R. State, “The eye of Horus: Spotting and analyzing attacks on Ethereum smart contracts,” in Financial Cryptography and Data Security (FC), 2021, pp. 33–52.
  8. [8]L. Su, X. Shen, X. Du, X. Liao, X. Wang, L. Xing, and B. Liu, “Evil under the sun: Understanding and discovering attacks on Ethereum decentralized applications,” in Proceedings of the 30th USENIX Security Symposium, 2021, pp. 1307–1324.
  9. [9]S. Wu, Z. Yu, D. Wang, Y. Zhou, L. Wu, H. Wang, and X. Yuan, “DeFi-Ranger: Detecting DeFi price manipulation attacks,” IEEE Transactions on Dependable and Secure Computing, vol. 21, no. 4, pp. 4147–4161, 2024.DOI
  10. [10]M. Xie, M. Hu, Z. Kong, C. Zhang, Y. Feng, H. Wang, Y. Xue, H. Zhang, Y. Liu, and Y. Liu, “DeFort: Automatic detection and analysis of price manipulation attacks in DeFi applications,” in Proceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis (ISSTA), 2024, pp. 402–414.DOI
  11. [11]J. Zhong, D. Wu, Y. Liu, M. Xie, Y. Liu, Y. Liu, and N. Liu, “Detecting various DeFi price manipulations with LLM reasoning,” in Proceedings of the 40th IEEE/ACM International Conference on Automated Software Engineering (ASE), 2025, pp. 1781–1793.arxiv.org/abs/2502.11521
  12. [12]Y. Gai, L. Zhou, K. Qin, D. Song, and A. Gervais, “Blockchain large language models,” arXiv:2304.12749, 2023.arxiv.org/abs/2304.12749
  13. [13]Z. Wu, J. Wu, H. Zhang, Z. Zheng, and W. Wang, “Hunting in the dark forest: A pre-trained model for on-chain attack transaction detection in Web3,” in Proceedings of the ACM Web Conference (WWW), 2025, pp. 4519–4530.DOI
  14. [14]J. Wu, K. Lin, D. Lin, B. Zhang, Z. Wu, and J. Su, “Safeguarding blockchain ecosystem: Understanding and detecting attack transactions on cross-chain bridges,” in Proceedings of the ACM Web Conference (WWW), 2025, pp. 4902–4912.DOI
  15. [15]B. Cai, W. Bai, Y. Lu, H. Xu, Y. Yang, Y. Zhou, and K. Lu, “GenDetect: Generalizing reactive detection for resilience against imitative DeFi attack cascade,” in Proceedings of the 48th IEEE/ACM International Conference on Software Engineering (ICSE), 2026, pp. 2006–2018.DOI
  16. [16]D. Meng, F. Gritti, R. McLaughlin, N. Ruaro, I. Grishchenko, C. Kruegel, and G. Vigna, “HOUSTON: Real-time anomaly detection of attacks against Ethereum DeFi protocols,” in Proceedings of the Network and Distributed System Security Symposium (NDSS), 2026.DOI
  17. [17]T. Abdelaziz and K. Ali, “TxLens: Scalable real-time detection of malicious Ethereum transactions,” in Proceedings of the IEEE International Conference on Blockchain and Cryptocurrency (ICBC), 2026, pp. 1–10.DOI
  18. [18]T. Durieux, J. F. Ferreira, R. Abreu, and P. Cruz, “Empirical review of automated analysis tools on 47,587 Ethereum smart contracts,” in Proceedings of the 42nd ACM/IEEE International Conference on Software Engineering (ICSE), 2020, pp. 530–541.DOI
  19. [19]Z. Zheng, J. Su, J. Chen, D. Lo, Z. Zhong, and M. Ye, “DAppSCAN: Building large-scale datasets for smart contract weaknesses in DApp projects,” IEEE Transactions on Software Engineering, vol. 50, no. 6, pp. 1360–1373, 2024.DOI
  20. [20]J. Wang, A. Bigger, X. Xu, J. W. Lin, A. Applebaum, T. Patwardhan, A. Yukseloglu, and O. Watkins, “EVMBench: Evaluating AI agents on smart contract security,” arXiv:2603.04915, 2026.arxiv.org/abs/2603.04915
  21. [21]J. Huang, F. Jiang, R. Poovendran, and Z. Lin, “CyberChainBench: Can AI agents secure smart contracts against real-world on-chain vulnerabilities?” arXiv:2606.26216, 2026.arxiv.org/abs/2606.26216
  22. [22]A. Alhaidari, B. Palanisamy, and P. Krishnamurthy, “On-chain decentralized learning and cost-effective inference for DeFi attack mitigation,” in Proceedings of the 7th Conference on Advances in Financial Technologies (AFT), ser. LIPIcs, vol. 354, 2025, pp. 35:1–35:27.DOI
  23. [23]BlockSec, “Security incidents library,” https://blocksec.com/security-incident, downloaded 2026-10-11, 2026.
  24. [24]evm-hack-registry contributors, “evm-hack-registry,” https://github.com/sanbir/evm-hack-registry, commit 5e310fc, 2026.
  25. [25]OWASP Smart Contract Security Project, “Smart contract weakness enumeration (SCWE), version 1.0,” https://scs.owasp.org/SCWE/, commit fefd476 of OWASP/owasp-scs, 2026.
  26. [26]SunWeb3Sec, “DeFiHackLabs: Reproduce DeFi hacked incidents using Foundry,” https://github.com/SunWeb3Sec/DeFiHackLabs, 2026.
  27. [27]J. Cohen, “A coefficient of agreement for nominal scales,” Educational and Psychological Measurement, vol. 20, no. 1, pp. 37–46, 1960.DOI
  28. [28]DeFiLlama, “DeFiLlama coins API: Historical token prices,” https://api-docs.defillama.com, 2026.
  29. [29]E. B. Wilson, “Probable inference, the law of succession, and statistical inference,” Journal of the American Statistical Association, vol. 22, no. 158, pp. 209–212, 1927.
  30. [30]E. L. Korn and B. I. Graubard, “Confidence intervals for proportions with small expected number of positive counts estimated from survey data,” Survey Methodology, vol. 24, no. 2, pp. 193–201, 1998.

Paper details

vlnc:W_wjn5wyh5
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 14

downloads · 1

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents