Skill Security Auditor
Use this skill to produce a very short safety report for non-technical users.
This skill is inspired by the conservative checklist style of skill-vetter, but the final answer should read like office software guidance, not a code audit memo.
Workflow
- Run the bundled static scanner from this skill directory:
node .\scripts\audit_skill.js "<path-to-skill-folder>"
- Read the JSON output. Prioritize:
simple_reportrisk_levelsummary
Only use findings when you need to explain why a rating is high.
- Convert the JSON into a concise Chinese report. Avoid technical jargon unless it is absolutely necessary.
Report Format
Use this structure:
# Skill 安全审查报告
## 1. 总体结论
- Skill 名称：...
- 审查对象：...
- 总体风险：低 / 中 / 高 / 阻止使用
- 使用建议：可以直接使用 / 先看一下再用 / 谨慎使用 / 不建议使用
## 2. 分项安全等级
- 脚本执行安全：低 / 中 / 高 / 阻止
- 联网访问安全：低 / 中 / 高 / 阻止
- 本地读取安全：低 / 中 / 高 / 阻止
- 本地改动安全：低 / 中 / 高 / 阻止
- 隐私与密钥安全：低 / 中 / 高 / 阻止
- 指令与内容安全：低 / 中 / 高 / 阻止
- 上传整洁度：低 / 中 / 高 / 阻止
## 3. 一句话说明
用不超过 3 句大白话说明这个 skill 大概会做什么，以及为什么会得到这个风险等级。
## 4. 最终建议
只给一个直接建议，不要展开讲修改方案，不要让用户自己去逐行看代码，除非用户明确要求。
Risk Rules
- Low: only instructions/docs, or scripts that only read user-provided files.
- Medium: scripts, declared dependencies, network access with clear purpose, or local file reads.
- High: shell execution, file writes/deletes, broad directory scans, environment variable reads, external APIs, or network + shell together.
- Block: credential path reads, unknown exfiltration endpoints, obfuscated code, download-and-execute patterns, safety bypass instructions, destructive commands, or clear prompt-injection instructions.
What To Look For
- Scripts:
.js,.py,.sh,.ps1,.bat, executable package scripts. - Network:
fetch,axios,request,http,https,curl,wget, websocket usage. - File access:
fs.readFile,fs.writeFile,rm,unlink, recursive scans, absolute user paths. - Shell:
child_process,exec,spawn,powershell,cmd.exe,bash. - External APIs/secrets:
API_KEY,TOKEN,SECRET,PASSWORD,COOKIE,Bearer,Authorization,.env,process.env. - Sensitive paths:
.ssh,.aws,.npmrc,.git-credentials,C:\Users,/Users,/home. - Prompt injection:
ignore previous instructions,disable safety,bypass,system prompt,developer message. - Upload hygiene:
node_modules, caches, logs, temporary files, sample private documents, lock files, hidden files.
Rules
- Prefer
simple_reportover raw findings. - Keep the answer short.
- Do not dump long evidence lists unless the user asks for details.
- Treat the user as a beginner who mainly uses office software and does not understand code terms.
- Replace engineering terms with plain descriptions whenever possible.
- Default to outcome-first wording such as "会不会联网" "会不会动本地文件" "能不能放心用".
- If the scanner finds something suspicious and the user asks why, then quote the file path and line number.
- If the scanner misses context, inspect the relevant file manually before the final report.
