AI Skill Scanner

对本地 skill 文件夹运行 Python 扫描脚本，基于规则静态检测凭据外泄、可疑网络请求、混淆代码与提示注入等模式，按 CLEAN 到 DANGEROUS 分级并给出是否安装的建议，支持普通、verbose 与 JSON 输出，并用退出码 0-3 区分结果。

HugoSbl

热度 94.57

最近更新 2026-02-28 11:21:49

分类：安全依赖：代码执行 / 读写文件应用场景：代码审查 / 报告撰写 / 工作流

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add ai-skill-scanner

Metadata

name: skill-scanner
description: Scan OpenBot/Clawdbot skills for security vulnerabilities, malicious code, and suspicious patterns before installing them. Use when a user wants to audit a skill, check if a ClawHub skill is safe, scan for credential exfiltration, detect prompt injection, or review skill security. Triggers on security audit, skill safety check, malware scan, or trust verification.

Skill Security Scanner

Scan skills for malicious patterns before installation. Detects credential exfiltration, suspicious network calls, obfuscated code, prompt injection, and other red flags.

Quick Start

# Scan a local skill folder
python3 scripts/scan.py /path/to/skill

# Verbose output (show matched lines)
python3 scripts/scan.py /path/to/skill --verbose

# JSON output (for automation)
python3 scripts/scan.py /path/to/skill --json

Workflow: Scan Before Install

Download or locate the skill folder
Run python3 scripts/scan.py <skill-path> --verbose
Review findings by severity (CRITICAL/HIGH = do not install)
Report results to user with recommendation

Score Interpretation

Score	Meaning	Recommendation
CLEAN	No issues found	Safe to install
INFO	Minor notes only	Safe to install
REVIEW	Medium-severity findings	Review manually before installing
SUSPICIOUS	High-severity findings	Do NOT install without thorough manual review
DANGEROUS	Critical findings detected	Do NOT install — likely malicious

Exit Codes

0 = CLEAN/INFO
1 = REVIEW
2 = SUSPICIOUS
3 = DANGEROUS

Rules Reference

See references/rules.md for full list of detection rules, severity levels, and whitelisted domains.

Limitations

Pattern-based detection — cannot catch all obfuscation techniques
No runtime analysis — only static scanning
False positives possible for legitimate tools that access network/files
Always combine with manual review for HIGH/MEDIUM findings