README.md

---
license: mit
pretty_name: SkillLeakBench
language:
- en
tags:
- security
- ai-safety
- llm-agents
- agent-skills
- claude-code
- credential-leakage
- prompt-injection
task_categories:
- text-classification
size_categories:
- 1K<n<10K
configs:
- config_name: affected_skills
  data_files:
  - split: train
    path: skills_dataset.csv
- config_name: issues
  data_files:
  - split: train
    path: issues.csv
- config_name: remediation
  data_files:
  - split: train
    path: remediation_summary.csv
- config_name: popularity
  data_files:
  - split: train
    path: popularity_hardcoded_repos.csv
---

# SkillLeakBench

A credential-leakage benchmark for LLM agent skills, from the ASE 2026 paper [*How Your Credentials Are Leaked by LLM Agent Skills: An Empirical Study*](https://arxiv.org/abs/2604.03070).

- 📄 **Paper:** <https://arxiv.org/abs/2604.03070>
- 💻 **Code & detection pipeline:** <https://github.com/AgentSkillsPrivacy/SkillLeakBench>
- 📦 **Archive:** <https://doi.org/10.5281/zenodo.19367969>

## Dataset summary

We collected 170,226 skills from SkillsMP and analyzed a 17,022-skill sample with static secret extraction, dynamic sandbox testing, and manual review. This release contains the 520 affected skills (437 vulnerable + 83 malicious) and their 1,708 security issues, across 10 leakage patterns (4 vulnerability + 6 malicious), plus disclosure and popularity tables. The release is de-identified (developer usernames removed) and contains no live credential values.

| Classification | Skills | Issues |
|----------------|-------:|-------:|
| Vulnerable | 437 | 1,371 |
| Malicious | 83 | 337 |
| **Total** | **520** | **1,708** |

## Files & configs

### `skills_dataset.csv` — affected skills (520 rows)
Per-skill records. Columns: `source`, `skill_name`, `classification` (`vulnerable` / `malicious`), `patterns` (semicolon-separated), `issue_count`, `severity`.

### `issues.csv` — security issues (1,708 rows)
One row per issue. Columns: `skill_id`, `skill_name`, `classification`, `pattern_id`, `academic_code`, `pattern`, `severity`.

### `remediation_summary.csv` — disclosure outcomes (3 rows)
Columns: `classification`, `total`, `resolved`, `remaining`.

### `popularity_hardcoded_repos.csv` — repository popularity (37 rows)
Name-free popularity of hardcoded-credential repositories. Columns: `repo_status`, `stars`, `forks`.

**Leakage patterns.** Vulnerability (4): Information Exposure, Hardcoded Credentials, Insecure Storage, Artifact Leakage. Malicious (6): Remote Exploitation, Defense Evasion, Credential Compromise, Data Exfiltration, Resource Hijacking, Persistence.

### Loading

```python
from datasets import load_dataset

affected    = load_dataset("AgentSkillPrivacy/SkillLeakBench", "affected_skills")
issues      = load_dataset("AgentSkillPrivacy/SkillLeakBench", "issues")
remediation = load_dataset("AgentSkillPrivacy/SkillLeakBench", "remediation")
popularity  = load_dataset("AgentSkillPrivacy/SkillLeakBench", "popularity")
```

## Intended use & ethics

Released for defensive security research on agent skills. The data is de-identified and excludes live credential values; vulnerabilities were responsibly disclosed to the platform.

## Citation

```bibtex
@inproceedings{skillleakbench2026,
  title     = {How Your Credentials Are Leaked by {LLM} Agent Skills: An Empirical Study},
  author    = {Chen, Zhihao and Zhang, Ying and Liu, Yi and Deng, Gelei and Li, Yuekang and Zhang, Yanjun and Ning, Jianting and Zhang, Leo and Ma, Lei and Li, Zhiqiang},
  booktitle = {Proceedings of the 41st IEEE/ACM International Conference on Automated Software Engineering (ASE)},
  year      = {2026}
}
```

## License

MIT