ethiack/ethibench-gt

EthiBench Ground Truth

Expert-annotated ground-truth vulnerability entries for the EthiBench evaluation framework — a practical evaluation protocol for AI pentesting agents, shifting assessment from task completion to validated vulnerability discovery.

📄 Paper: From Controlled to the Wild: Evaluation of Pentesting Agents for the Real-World

Dataset Summary

This dataset contains 108 expert-annotated ground-truth vulnerability entries across 3 open-source web application targets. These entries serve as the reference set for evaluating AI pentesting agents using the EthiBench framework.

Target	Config	GT Entries	Repository
vuln-bank	vulnbank	60	Commando-X/vuln-bank
PAYGoat	paygoat	28	stuxctf/PAYGoat
XBEN-090	xben	20	xbow-engineering/validation-benchmarks

⚠️ Target Version Alignment

The ground truth was annotated against specific pinned versions of the target apps. Two of the upstream repos have since diverged, so git clone <latest main> will not reproduce the benchmarked target — use the pinned versions below.

Target	Pinned version to use	Why
vuln-bank	Commit b1962823 ("disabled query in transactions", 2025-07-23) + de-hinting patch (see EthiBench docs).	main later added a merchant-payments module, a GraphQL API, marketing pages and tests — extra attack surface not in the GT.
PAYGoat	Commit 69589b4 ("Add ticket scenario", 2025-08-04), i.e. before "update to v2" (0cd0077).	"v2" adds a whole business-accounts module (extra controllers, routes, models and dashboard pages).
XBEN-090-24	Latest main is fine — the XBEN benchmarks are frozen.	Unchanged upstream.

See the full EthiBench repository for the vuln-bank de-hinting patch and detailed setup instructions.

Data Schema

Each entry is a JSON object with the following fields:

Field	Type	Description
id	string	Unique UUID identifier
name	string	Short vulnerability title
description	string	Detailed technical description
category	string	Vulnerability category
additional_info	string	CWE classification (e.g., "CWE-89: SQL Injection")
cvss	float	CVSS 4.0 base score
cvss_vector	string	Full CVSS 4.0 vector string
subset_name	string	Human-readable target name
target_id	string	Machine-readable target identifier

Usage

from datasets import load_dataset

# Load all targets
dataset = load_dataset("ethiack/ethibench-gt")

# Load a specific target
paygoat = load_dataset("ethiack/ethibench-gt", "paygoat")
vulnbank = load_dataset("ethiack/ethibench-gt", "vulnbank")
xben = load_dataset("ethiack/ethibench-gt", "xben")

# Example: inspect entries
for entry in paygoat["test"]:
    print(f"{entry['name']} (CVSS {entry['cvss']}): {entry['additional_info']}")

Use with EthiBench

This ground truth is designed to be used with the EthiBench evaluation framework. Install and run:

pip install ethibench
ethibench evaluate <experiment_dir> --dataset data/dataset.yaml --gt-dir data/

See the full documentation for details on the evaluation pipeline, matching algorithm, and metrics.

Citation

If you use this dataset, please cite:

@misc{conde2026controlledwildevaluationpentesting,
      title={From Controlled to the Wild: Evaluation of Pentesting Agents for the Real-World},
      author={Pedro Conde and Henrique Branquinho and Valerio Mazzone and Bruno Mendes and André Baptista and Nuno Moniz},
      year={2026},
      eprint={2605.10834},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2605.10834},
}

License

This dataset is released under the MIT License.

---

Disclaimer: This content is intended for educational purposes and authorized security testing only. Users are responsible for ensuring compliance with applicable laws and regulations.