command-guardian 安全扫描报告 — 可信 | ClawSafe

5 /100

command-guardian

Preflight safety guard for shell and infrastructure commands

Command Guardian is a legitimate preflight safety analyzer for shell commands. It performs read-only risk analysis and does not execute analyzed commands or exhibit any malicious behavior.

技能名称command-guardian

分析耗时41.0s

引擎pi

✓

可以安装

This skill is safe for use. The dangerous shell patterns (curl|sh, wget|bash) referenced in documentation are documented as dangerous patterns TO FLAG, not patterns the skill performs itself.

资源类型	声明权限	推断权限	状态	证据
文件系统	`READ`	`READ`	✓ 一致	agents/openai.yaml declares filesystem:READ
命令执行	`WRITE`	`WRITE`	✓ 一致	guardlib.py:294-297 uses subprocess for git commands only
环境变量	`NONE`	`READ`	✓ 一致	guardlib.py:76 uses os.path.expandenv() for path resolution only
网络访问	`NONE`	`NONE`	—	No network calls found
技能调用	`NONE`	`NONE`	—	No skill invocation found
剪贴板	`NONE`	`NONE`	—	No clipboard access found
浏览器	`NONE`	`NONE`	—	No browser access found
数据库	`NONE`	`NONE`	—	No database access found

2 严重 2 项发现

💀

严重危险命令危险 Shell 命令

curl | sh

SKILL.md:54

💀

严重危险命令危险 Shell 命令

wget | bash

references/tool-patterns.md:30

目录结构

11 文件 · 45.5 KB · 1208 行

Python 7f · 1007L Markdown 3f · 193L YAML 1f · 8L

├─ ▾ 📁 agents

│ └─ 📋 openai.yaml YAML 8L · 300 B

├─ ▾ 📁 references

│ ├─ 📝 risk-rules.md Markdown 41L · 1.3 KB

│ └─ 📝 tool-patterns.md Markdown 42L · 1.4 KB

├─ ▾ 📁 scripts

│ ├─ 🐍 classify_command.py Python 15L · 307 B

│ ├─ 🐍 guardlib.py Python 824L · 31.8 KB

│ ├─ 🐍 path_guard.py Python 17L · 466 B

│ ├─ 🐍 preflight.py Python 60L · 1.9 KB

│ ├─ 🐍 rollback_hints.py Python 15L · 317 B

│ └─ 🔑 secret_guard.py ⚠ Python 15L · 315 B

├─ ▾ 📁 tests

│ └─ 🐍 test_powershell_risk.py Python 61L · 3.0 KB

└─ 📝 SKILL.md Markdown 110L · 4.4 KB

安全亮点

✓ Skill performs read-only command analysis without executing analyzed commands

✓ Subprocess usage limited to git context (git rev-parse, git status) - legitimate for security tooling

✓ No credential harvesting or exfiltration - only detects inline secrets to warn about them

✓ No network requests or data exfiltration channels

✓ No obfuscation techniques (no base64, no eval, no atob)

✓ No access to sensitive paths (~/.ssh, ~/.aws, .env)

✓ No persistence mechanisms (no cron, no startup hooks, no backdoors)

✓ Capabilities properly declared in agents/openai.yaml

✓ Comprehensive test suite validates expected security behaviors

✓ Dangerous patterns (curl|sh, wget|bash) are documented AS PATTERNS TO FLAG, not as behaviors exhibited