# The Deception Delta: Adversarial Evaluation of LLM-Based Smart Contract Bytecode Forensics

> Evaluates how language models interpret raw EVM bytecode from purpose-built deceptive and control contracts. It distinguishes recognizing a hidden transfer mechanism from correctly judging its risk, using model judges alongside judge-independent checks.

- Authors: Timo Schefold
- Venue: ACM CCS 2026
- Status: Conference accepted
- Publication / posting date: 2026 (exact date not verified)
- Date source: Official ACM CCS 2026 accepted-paper list confirms acceptance, not a publication date. The reviewed arXiv v2 revision is dated 2026-10-06.
- First arXiv submission: September 12, 2026
- arXiv: 2609.14098
- Reviewed arXiv version: v2
- DOI: 10.48550/arxiv.2609.14098
- Sources reviewed: 2026-10-20
- Added to the catalog: 2026-10-20
- Original paper: <https://arxiv.org/abs/2609.14098>
- PDF: <https://arxiv.org/pdf/2609.14098v2>
- Project: <https://zenodo.org/records/23164989>
- BibTeX: <https://awesomerea.org/papers/citations/paper-deception-delta.bib>
- Catalog page: <https://awesomerea.org/papers/paper-deception-delta/>
- Record JSON: <https://awesomerea.org/papers/paper-deception-delta/record.json>
- Chinese version: [简体中文](https://awesomerea.org/zh/papers/paper-deception-delta/index.md)

## Scope and limitations

The main evaluation covers 22 models and 13 synthetic contracts using single-shot raw bytecode, without execution tools, source retrieval or on-chain context. Its 8,528 analyzed responses exclude refusals and pending runs. Gold labels come from one expert; some judges also belong to evaluated model families. Ground-truth checks vary across contracts, including owner/selector-only checks and no dedicated EVM tests for two injection variants. These results neither measure deployed-contract prevalence nor generalize to interactive reverse-engineering agents.

## Why it fits this collection

Adds a bounded failure-mode evaluation for interpreting executable bytecode, especially the gap between describing behavior and assigning a trustworthy safety verdict.

## Publication and review notes

The official CCS list confirms the single-author title; arXiv v2 corrects the September v1. The manuscript prints ACM DOI 10.1145/3830454.3846626 and future conference details, but its Crossref lookup returned 404 during review; no publisher publication date or container fields are asserted. The DOI field is the separately registered arXiv DOI. Article and Zenodo metadata declare CC BY 4.0. The concept artifact resolves to record 23164989, v1.0.2, dated 2026-10-05; its metadata describes ERRATA and a 53,006,308-byte ZIP.

## Primary sources

1. <https://arxiv.org/abs/2609.14098>
2. <https://arxiv.org/html/2609.14098v2>
3. <https://api.datacite.org/dois/10.48550/arxiv.2609.14098>
4. <https://www.sigsac.org/ccs/CCS2026/program/accepted-papers.html>
5. <https://doi.org/10.5281/zenodo.22686323>
6. <https://zenodo.org/records/23164989>
7. <https://zenodo.org/api/records/22686323>

## About this entry

Reviewed against the paper and its primary sources. Cite the original authors using the linked paper or BibTeX.
Original editorial content: MIT. Linked projects, papers and assets retain their own rights and licenses.