胖叔 Skill 安全审查

该 Skill 是安装前的安全审查工具，通过 scripts/vetter.py 对目标 Skill 目录做静态扫描：匹配危险 Bash 命令、敏感文件与凭据访问、网络外传、混淆代码等模式，按严重级别（Critical/Warning/Info）汇总发现并输出拦截或放行判定，可手动运行或由宿主安装钩子自动触发。

hjshysst-dot

热度 39.15

最近更新 2026-05-11 17:40:29

分类：安全依赖：代码执行 / 读写文件应用场景：代码审查 / 自动化脚本

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 pangshu-skill-vetter

Metadata

name: skill-vetter
description: Security vetting for agent skills before installation. Scans skill code for dangerous Bash commands, sensitive file access, network exfiltration, obfuscated code, and other security risks. Triggers automatically before any skill install, update, or when user asks to "vet" or "check" a skill's security.

Skill Vetter

Security scanner that analyzes skill code before installation.

When to Use

Before any skill installation via ClawHub or manual
Before any skill update (even trusted skills)
When user asks: "vet this skill", "check security", "scan before install"
When installing skills from unknown sources

Security Checks (Priority Order)

🔴 Critical (Block Install)

Check	Pattern	Action
Disk wipe	`rm -rf /`, `rm -rf ~`, `dd if=.*of=/dev/sdX`	BLOCK
Fork bomb	`:(){ :\|:& };:`, `fork()` loop	BLOCK
Format	`mkfs`, `newfs`, `umount -f`	BLOCK
SSH key deletion	`rm.\.ssh/`, `ssh-keygen.-D`	BLOCK
System takeover	`chmod 777.*shadow`, `/etc/passwd` edit	BLOCK

🟡 Medium (Warn + Allow)

Check	Pattern	Action
Dangerous rm	`rm -rf [a-z]+\` (recursive without targets)	WARN
Network exfil	base64 remote exfil, `curl.\|.sh`	WARN
Credential access	`.env`, `~/.aws/`, API key patterns	WARN
Suspicious encoding	Obfuscated JavaScript, encoded commands	WARN
High-privilege	`sudo`, `chmod 777`, `setfacl`	WARN
Unknown network	Non-standard ports, suspicious domains	WARN

🔵 Info (Log Only)

Check	Pattern
File write	Writes outside skill directory
Permission change	Any chmod
New file creation	File creation in system paths

Usage

Automatic Vetting (via Hook)

When OpenClaw hooks are configured, the vetter runs automatically:

# Configure in OpenClaw settings
skill_vetter:
  enabled: true
  auto_block: true   # Block critical issues
  warn_only: false    # false = block criticals

Manual Vetting

python3 ~/.openclaw/skills/skill-vetter/scripts/vetter.py \
    scan ~/.openclaw/skills/my-skill

Output Format

╔══════════════════════════════════════════════════╗
║ 🛡️ Skill Security Vetting Report                  ║
╠══════════════════════════════════════════════════╣
║ Skill: my-skill                                   ║
║ Scan Time: 2026-04-06 07:21:00                   ║
║                                                          ║
║ 🔴 CRITICAL: 0  🟡 WARNING: 2  🔵 INFO: 1            ║
╠══════════════════════════════════════════════════╣
║ Findings:                                              ║
║   (see actual scan output for details)                 ║
╠══════════════════════════════════════════════════╣
║ Verdict: ⚠️ INSTALL WITH CAUTION                       ║
╚══════════════════════════════════════════════════╝

Integration

For automatic pre-install vetting, this skill should be invoked by the OpenClaw skill installation hook. The hook configuration:

# In OpenClaw config
hooks:
  pre_skill_install:
    - name: skill-vetter
      action: scan_then_block
      block_on_critical: true

Verdict Logic

Severity	Count	Verdict
🔴 Critical	> 0	🚫 BLOCK INSTALL
🟡 Warning	> 3	⚠️ WARN + CONFIRM
🟡 Warning	≤ 3	✅ INSTALL WITH CAUTION
Only 🔵 Info	Any	✅ CLEAR TO INSTALL

Implementation

The scanner is in scripts/vetter.py. Key functions:

scan_skill(skill_path) - Main entry point
check_dangerous_commands(content) - Bash pattern matching
check_sensitive_access(content) - File/credential patterns
check_network_activity(content) - Exfil indicators
check_obfuscation(content) - Obfuscated code detection
generate_report(findings) - Formatted output