OpenSkillRisk: Benchmarking Agent Safety When Using Real-World Risky Third-Party Skills
Paper • 2607.20121 • Published
OpenSkillRisk Dataset
Dataset Description
OpenSkillRisk is a benchmark for evaluating whether CLI agents can safely handle risky third-party skills in open skill ecosystems.
This dataset contains the skill files and task specifications used by OpenSkillRisk. The code for building sandboxes and running evaluations is released separately at:
- Repository: Miaow-Lab/OpenSkillRisk
- Paper: OpenSkillRisk: Benchmarking Agent Safety When Using Real-World Risky Third-Party Skills
Dataset Structure
This dataset contains two top-level folders:
skills/contains the third-party skill files used in the benchmark.tasks/contains the task specification files that define benchmark instances.
Dataset Size
| Split | Number of skills and tasks |
|---|---|
| Obviously malicious | 139 |
| Contextually risky | 124 |
| Total | 263 |
| Risk category | Total number of skills and tasks |
|---|---|
| Control-Plane Hijacking | 13 |
| Authority Expansion | 38 |
| Data Harvesting | 27 |
| Execution Bootstrapping | 30 |
| Persistence Implantation | 24 |
| Outbound Exfiltration | 96 |
| External State Manipulation | 35 |
Usage
Please refer to the GitHub repository for instructions on downloading this dataset and running OpenSkillRisk.
Important Notice
The third-party skills in this repository contains high-risk skills collected for controlled safety research and benchmark evaluation. These skills should not be redistributed, copied into other agent environments, or used outside the intended benchmark context.
The benchmark is designed for sandboxed evaluation, but users should still run it only in controlled, security-conscious environments and avoid exposing sensitive credentials or personal data.
License
MIT License. See LICENSE.
Citation
If you use OpenSkillRisk in your research, please cite:
@article{liu2026openskillrisk,
title={OpenSkillRisk: Benchmarking Agent Safety When Using Real-World Risky Third-Party Skills},
author={Liu, Qiyuan and Hui, Tingfeng and Zhan, Kun and Zhang, Kaike and Miao, Ning},
journal={arXiv preprint arXiv:2607.20121},
year={2026}
}
Takedown and Removal Requests
If any skill, file, or derived content in this dataset belongs to you and you would like it removed, please provide a brief description of your ownership or rights claim and contact us. We will review valid requests and update the dataset as appropriate.
- Downloads last month
- -