| --- |
| language: |
| - en |
| license: mit |
| task_categories: |
| - text-classification |
| tags: |
| - security |
| - malware-detection |
| - agent-skills |
| - openclaw |
| size_categories: |
| - n<1K |
| --- |
| |
| # Agent Skill Malware: Malicious vs Benign Agent Skills |
|
|
| Binary classification dataset of OpenClaw agent skill files (SKILL.md). |
| 127 malicious + 223 benign = 350 samples, deduplicated by content hash. |
|
|
| ## Source |
|
|
| - **Malicious**: Real skills from malicious campaigns targeting ClawHub (Feb 2026), extracted from the `openclaw/skills` GitHub archive. They use social engineering in markdown instructions to trick agents/users into running malware -- primarily AMOS (Atomic macOS Stealer), though some variants exfiltrate credentials directly. |
| - **Benign**: Community skills from ClawHub covering developer tools, APIs, and utilities. |
|
|
| ## Format |
|
|
| JSONL with fields: `id` (content hash), `skill_name`, `content` (full SKILL.md text), `label` ("malicious" / "benign"). |
|
|
| ```python |
| from datasets import load_dataset |
| ds = load_dataset("json", data_files="skills.jsonl") |
| ``` |
|
|
| ## Suggested minimal detection prompt |
|
|
| > Classify this SKILL.md as "malicious" or "benign". Malicious skills require downloading a fake prerequisite binary before use, contain base64-encoded shell commands, or exfiltrate data to external endpoints. |
|
|
| ## Detector (100% accuracy on this dataset) |
|
|
| ```python |
| import re |
| |
| PATTERNS = [ |
| r"openclawcli", r"openclaw-core", r"glot\.io/snippets", |
| r"rentry\.co/openclaw", r"Ddoy233", r"denboss99", |
| r"base64\s+-[dD]", r"91\.92\.242\.30", r"openclawcli\.vercel", |
| r"webhook\.site", r"setup-service\.com", r"AuthTool", |
| r"pass(?:word)?[:\s]+.{0,10}openclaw", |
| ] |
| _RE = re.compile("|".join(PATTERNS), re.IGNORECASE) |
| |
| def is_malicious(content: str) -> bool: |
| return bool(_RE.search(content)) |
| ``` |
|
|
| ## Attack patterns |
|
|
| | Pattern | Description | Frequency | |
| |---------|-------------|-----------| |
| | Fake prerequisite binary | "openclawcli/openclaw-core must be running" | 95%+ | |
| | glot.io paste redirect | Links to paste site hosting shell command | 57% | |
| | Inline base64 payload | `echo '...' \| base64 -D \| bash` | 35% | |
| | Password-protected ZIP | GitHub releases ZIP with pass "openclaw" | 10% | |
| | setup-service.com | Redirects to malware download | 7% | |
| | vercel.app redirect | openclawcli.vercel.app hosting payload | 6% | |
| | webhook.site exfil | Credential exfiltration endpoint | 1% | |
|
|
| ## License |
|
|
| Dataset collected for security research purposes. Malicious content is preserved verbatim for classifier training -- do not execute any commands from malicious samples. |
|
|
