File size: 3,331 Bytes
c12142a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 | ---
license: mit
pretty_name: PatchBench
size_categories:
- n<1K
task_categories:
- text-generation
language:
- en
tags:
- code
- security
- vulnerability-repair
- program-repair
- agents
- benchmark
- fuzzing
configs:
- config_name: default
data_files:
- split: test
path: data/test-*.parquet
---
# PatchBench
[](https://arxiv.org/abs/2609.04075)
[](https://github.com/ai-sec-lab/PatchBench)
[](https://github.com/ai-sec-lab/PatchBench/blob/main/LICENSE)
PatchBench is a benchmark for evaluating AI agents on realistic vulnerability patching tasks: **213 tasks drawn from 32 popular GitHub C/C++ projects**. It selects vulnerabilities whose ground-truth fixes lie *outside* the crash stack, and uses vulnerability transplant plus code mutation to mitigate surface-level fixes and patch memorization.
This repository holds the **task metadata**, one row per task to identify the project, the exact repository state, the crash, and the reference patch.
## Dataset Structure
Each example contains:
- **id**: Unique task identifier, also naming the task's image (Same as the one in [ARVO](https://github.com/n132/ARVO-Meta))
- **image**: Docker Hub image for the task, `b4drequest/vulpatch:{id}-vul`
- **project**: Real-world repository name
- **repo_addr**: Repository URL
- **base_commit**: Historical commit at which the vulnerability is still present, upstream of the developer's fixing commit
- **inverted_commit**: Newer upstream commit the vulnerability was transplanted onto
- **fuzzer**: Fuzzing engine that produced the crash, one of `libfuzzer`, `honggfuzz`, `afl`, `aflpp`
- **sanitizer_type**: The error detection tool type
- **CWE_ID**: The weakness the task's vulnerability manifests, derived from the sanitizer report and confirmed against the ground-truth patch
- **secondary_CWE_ID**: The underlying defect behind it where the evidence pins one down, and `None` where the primary already states the weakness
- **command**: Command that reproduces the crash inside the image, with the PoC at `/tmp/poc`
- **sanitizer_report**: Sanitizer output for the original PoC, including the crash stack
- **patch**: Reference patch against the task repository
`base_commit`, `inverted_commit` and `patch` are **ground truth** and are not visible to the agent during inference.
## Dataset Usage
```python
from datasets import load_dataset
dataset = load_dataset("ai-sec-lab/PatchBench", split="test")
```
Filtering works the usual way:
```python
libxml2 = dataset.filter(lambda r: r["project"] == "libxml2")
ubsan = dataset.filter(lambda r: r["sanitizer_type"] == "ubsan")
```
## Evaluation
See our repo for details: [ai-sec-lab/PatchBench](https://github.com/ai-sec-lab/PatchBench).
## Citation
```bibtex
@misc{shen2026PatchBench,
title = {{{PatchBench}}: Evaluating {{AI}} Agents for Vulnerability Patching},
author = {Shen, Chihao and Li, Jiacheng and Mahajan, Aastha and Tian, Jeffery Siyuan and Kwon, Yonghwi and Chen, Yizheng},
year = 2026,
number = {arXiv:2609.04075},
eprint = {2609.04075}
}
```
|