AgentSkillPrivacy/SkillLeakBench

---
title: SkillLeakBench
license: mit
---

# How Your Credentials Are Leaked by LLM Agent Skills: An Empirical Study

A credential leakage benchmark dataset for LLM agent skills. We collected **170,226 skills** from SkillsMP, sampled **17,022** for analysis, and identified **520 affected skills** containing **1,708 security issues** across **10 leakage patterns**.

## Purpose

The code and data in this repository are intended exclusively for:
- Academic research on AI agent security
- Developing defense mechanisms against credential leakage in agent skills
- Evaluating the robustness of AI agent platforms

## Files

| File | Rows | Description |
|------|------|-------------|
| `vulnerable_skills.csv` | 437 | Vulnerable skills with credential leakage patterns |
| `malicious_skills.csv` | 83 | Malicious skills with attack patterns |

## Columns

**vulnerable_skills.csv:** `source,skill_name,classification,patterns,issue_count,severity`

**malicious_skills.csv:** `source,skill_name,classification,patterns,issue_count,severity`

## Distribution

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

## 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

## Usage

```python
import pandas as pd

# Load datasets
vulnerable = pd.read_csv('vulnerable_skills.csv')
malicious = pd.read_csv('malicious_skills.csv')
```

## Citation

```bibtex
@inproceedings{skillleakbench2026,
title={How Your Credentials Are Leaked by LLM Agent Skills: An Empirical Study},
author={Anonymous},
year={2026},
}
```

## License

MIT License