README.md

metadata

license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - security
  - vulnerability
  - benchmark
size_categories:
  - n<1K
configs:
  - config_name: default
    default: true
    data_files:
      - split: test
        path: data/all/test.parquet
  - config_name: v8
    data_files:
      - split: test
        path: data/v8/test.parquet
  - config_name: sm
    data_files:
      - split: test
        path: data/sm/test.parquet

SEC-bench-Pro

SEC-bench-Pro is a benchmark dataset of real-world security vulnerabilities in JavaScript engines (V8 and SpiderMonkey). Each instance contains a verified vulnerability with its Docker-reproducible environment, detailed description, and ground-truth fix patch.

Dataset Summary

• Total instances: 183
• V8 (Chromium): 103 instances
• SpiderMonkey (Firefox): 80 instances
• Vulnerability types: 24 distinct categories (type confusion, use-after-free, sandbox bypass, OOB read/write, etc.)

Data Instances

Each row in the dataset represents a single vulnerability instance with the following fields:

Field	Type	Description
instance_id	string	Unique identifier in the format {project}__{bug_id} (e.g., v8__446124893, sm__1968423).
project	string	Target project: v8, sm (SpiderMonkey), etc.
bug_id	string	Bug tracker ID (Chromium issue ID for V8, Bugzilla ID for SpiderMonkey).
image_name	string	Docker image name for reproducing the vulnerable environment (e.g., hwiwonlee/v8.x86_64:446124893).
work_dir	string	Working directory inside the Docker container where the engine source is located (e.g., /src/v8, /src/gecko-dev).
verification_binary	string	Path to the built binary used to trigger the vulnerability (e.g., out/x64.asan/d8, /out/js).
command_options	string	Command-line flags needed to trigger the vulnerability (e.g., --allow-natives-syntax --experimental-wasm-exnref). May be empty.
target_source_files	list[string]	Source files that contain the vulnerability and need to be patched.
target_subdir	list[string]	Subdirectories within the engine source that are relevant to the vulnerability.
target_vulnerability_type	string	Classification of the vulnerability (e.g., Type confusion, Use-after-free, Out-of-bounds write).
error_type	string	How the vulnerability manifests when triggered: ASAN_CRASH, RUNTIME_CRASH, DCHECK, SANDBOX_VIOLATION, etc.
description	string	Detailed technical description of the vulnerability, including root cause analysis and exploitation potential.
vrp	string	Vulnerability Reward Program (bug bounty) value in USD, if applicable. Empty string or "none" if not awarded.
fix_patches	list[string]	Full text of the patch file(s) that fix the vulnerability, in unified diff format.

Usage

from datasets import load_dataset

# Load all instances (default config)
ds = load_dataset("SEC-bench/SEC-bench-Pro", split="test")

# Load only V8 instances
v8_ds = load_dataset("SEC-bench/SEC-bench-Pro", "v8", split="test")

# Load only SpiderMonkey instances
sm_ds = load_dataset("SEC-bench/SEC-bench-Pro", "sm", split="test")

# Filter by project
v8_only = ds.filter(lambda x: x["project"] == "v8")

# Filter by vulnerability type
type_confusion = ds.filter(lambda x: x["target_vulnerability_type"] == "Type confusion")

Citation

If you use this dataset, please cite:

@misc{lee2026secbenchprolanguagemodels,
      title={{SEC-bench Pro: Can Language Models Solve Long-Horizon Software Security Tasks?}}, 
      author={Hwiwon Lee and Jiawei Liu and Dongjun Kim and Ziqi Zhang and Chunqiu Steven Xia and Lingming Zhang},
      year={2026},
      eprint={2605.26548},
      archivePrefix={arXiv},
      primaryClass={cs.CR},
      url={https://arxiv.org/abs/2605.26548}, 
}