Claude Code Security全实战：AI驱动漏洞扫描与修复

graph TB
    subgraph "Claude Code Security 三层架构"
        A["第一层：终端即时扫描<br/>/security-review 命令"] --> D["核心推理引擎<br/>Claude LLM语义分析"]
        B["第二层：CI/CD自动审查<br/>GitHub Action"] --> D
        C["第三层：企业级持续扫描<br/>Claude Security Dashboard"] --> D
        D --> E["数据流追踪<br/>跨文件变量分析"]
        D --> F["逻辑推理<br/>授权语义理解"]
        D --> G["误报验证<br/>多阶段确认"]
        E --> H["发现报告<br/>置信度+严重度"]
        F --> H
        G --> H
        H --> I["补丁生成<br/>分支+PR"]
        H --> J["人工审查<br/>Human-in-the-loop"]
    end

    style D fill:#f9f,stroke:#333,stroke-width:2px
    style H fill:#bbf,stroke:#333,stroke-width:2px

> /security-review

# 扫描指定目录
> /security-review --scope src/auth/

# 扫描指定文件
> /security-review --scope src/api/users.ts

# 结合自然语言定位
> run a security review on the payment processing module

Claude Code发现 3 个安全问题：

┌─────────────────────────────────────────────────────┐
│ CRITICAL │ SQL Injection in src/api/users.ts:47     │
│ Confidence: HIGH                                    │
│                                                     │
│ 用户输入 req.query.name 直接拼接到SQL查询字符串中，  │
│ 攻击者可通过构造恶意输入执行任意SQL命令。           │
│                                                     │
│ 当前代码:                                           │
│   const query = `SELECT * FROM users                │
│     WHERE name = '${req.query.name}'`;              │
│                                                     │
│ 建议修复:                                           │
│   const query = `SELECT * FROM users                │
│     WHERE name = $1`;                               │
│   const result = await pool.query(query,            │
│     [req.query.name]);                              │
│                                                     │
│ 参考: OWASP A03:2021 - Injection                    │
└─────────────────────────────────────────────────────┘

> yes, apply the fix and run tests

✓ Applied fix to src/api/users.ts
✓ Running test suite... 47/47 tests passed
✓ No regressions detected

> /security-review src/auth/ and src/api/

# Security conventions
- Admin bypass routes in /src/admin/* intentionally skip auth checks
- Test utilities in /tests/helpers/ intentionally use hardcoded credentials
- The /api/internal/* endpoints are only accessible from internal network

name: Security Review

permissions:
  pull-requests: write   # 用于在PR上评论
  contents: read

on:
  pull_request:

jobs:
  security:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          ref: ${{ github.event.pull_request.head.sha || github.sha }}
          fetch-depth: 2

      - uses: anthropics/claude-code-security-review@main
        with:
          comment-pr: true
          claude-api-key: ${{ secrets.CLAUDE_API_KEY }}

- uses: anthropics/claude-code-security-review@main
  with:
    claude-api-key: ${{ secrets.CLAUDE_API_KEY }}
    comment-pr: true
    upload-results: true

    # 排除目录（vendor、测试fixture等）
    exclude-directories: "vendor/,tests/fixtures/,node_modules/"

    # 模型选择（平衡速度与深度）
    claude-model: "claude-opus-4-1-20250805"

    # 超时（大型monorepo可能需要更长）
    claudecode-timeout: 30

    # 每次提交都扫描（默认只扫描最新commit）
    run-every-commit: false

    # 自定义误报过滤指令
    false-positive-filtering-instructions: ".security/filter-rules.txt"

    # 自定义扫描指令
    custom-security-scan-instructions: ".security/custom-rules.txt"

Additional security checks for this project:

1. All API endpoints under /api/v2/ must validate Bearer tokens
2. Database queries must use parameterized statements (no string interpolation)
3. File uploads must validate MIME type AND file content, not just extension
4. Rate limiting is required on all public endpoints
5. PII fields (email, phone, SSN) must be encrypted at rest
6. Third-party OAuth tokens must never be logged

Known false positive patterns to exclude:

- Warnings about missing auth on /health and /metrics endpoints (intentional)
- Flags about hardcoded values in config/test.env (test environment only)
- Reports about CORS on localhost (development-only configuration)
- Missing CSRF warnings on API-only endpoints (using Bearer tokens instead)

sequenceDiagram
    participant Dev as 开发者
    participant PR as Pull Request
    participant CI as GitHub Actions
    participant CC as Claude Code Security
    participant Rev as 安全审查员

    Dev->>PR: 创建PR
    PR->>CI: 触发 security.yml
    CI->>CC: checkout + diff分析
    CC->>CC: 语义推理 + 数据流追踪
    CC->>CC: 多阶段验证 + 误报过滤
    CC->>PR: 内联评论：发现 + 严重度
    CC->>CI: JSON结果文件
    Note over PR: Critical问题阻止合并<br/>Medium/Low作为建议
    PR->>Rev: 人工审查确认
    Rev->>PR: Approve
    PR->>Dev: 合并

01 Finding → 02 Validation → 03 Patching → 04 Review → 05 Ship

express-payments/
├── src/
│   ├── api/
│   │   ├── auth.ts          # 认证路由
│   │   ├── payments.ts      # 支付处理
│   │   └── users.ts         # 用户管理
│   ├── middleware/
│   │   └── auth.ts          # JWT验证
│   ├── db/
│   │   └── queries.ts       # 数据库查询
│   └── utils/
│       └── validators.ts    # 输入校验
├── tests/
├── .security/
│   ├── custom-rules.txt
│   └── filter-rules.txt
├── .github/
│   └── workflows/
│       └── security.yml
├── CLAUDE.md
└── package.json

# Security Policy

## Authentication
- All /api/* endpoints except /api/auth/login and /api/auth/register require Bearer token
- JWT tokens expire after 1 hour, refresh tokens after 7 days
- Admin endpoints under /api/admin/* require role=admin in token payload

## Data Handling
- PII (email, phone, address) must be encrypted with AES-256 at rest
- Payment card data MUST NEVER touch our servers — use Stripe tokenization
- Log files must not contain user IDs, tokens, or request bodies

## Database
- All queries MUST use parameterized statements ($1, $2 placeholders)
- Connection pooling via pg-pool, max 20 connections
- Migration files are immutable once merged

## Input Validation
- All user input validated with joi schemas before processing
- File uploads limited to 5MB, allowed types: jpg, png, pdf
- Rate limiting: 100 req/min per IP, 1000 req/min per API key

---
description: "Comprehensive security review with OWASP compliance"
---

Perform a thorough security review of the changed files with focus on:

1. **Injection vectors**: SQL, NoSQL, command injection, XSS
2. **Authentication bypass**: Token validation, session management
3. **Authorization flaws**: Missing role checks, IDOR
4. **Data exposure**: Secrets in logs, PII leakage, debug endpoints
5. **Payment security**: Ensure no card data touches our servers

For each finding:
- Rate severity: Critical/High/Medium/Low
- Provide confidence rating: High/Medium/Low
- Generate a fix that addresses the root cause
- Explain why this matters in our payment processing context

Known safe patterns to exclude:
- /api/health endpoint intentionally has no auth
- Test files use mock credentials
- CORS allows localhost in development

# 进入项目目录
cd express-payments

# 启动Claude Code
claude

# 运行安全审查
> /security-review

# 查看结果后，让Claude修复所有Critical和High问题
> fix all critical and high severity issues, run the test suite after each fix

# 提交修复
> commit the security fixes with descriptive messages referencing each CVE category

name: Security Pipeline

on:
  pull_request:
    paths:
      - 'src/**'
      - 'package.json'
      - 'package-lock.json'

jobs:
  security-scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          ref: ${{ github.event.pull_request.head.sha }}
          fetch-depth: 2

      # 第一步：Claude Code语义安全审查
      - uses: anthropics/claude-code-security-review@main
        id: claude-security
        with:
          claude-api-key: ${{ secrets.CLAUDE_API_KEY }}
          comment-pr: true
          upload-results: true
          exclude-directories: "node_modules/,tests/fixtures/"
          claude-model: "claude-opus-4-1-20250805"
          false-positive-filtering-instructions: ".security/filter-rules.txt"
          custom-security-scan-instructions: ".security/custom-rules.txt"

      # 第二步：传统依赖扫描（与Claude互补）
      - name: npm audit
        run: npm audit --audit-level=high

      # 第三步：输出扫描结果摘要
      - name: Security Summary
        if: always()
        run: |
          echo "### Security Scan Results" >> $GITHUB_STEP_SUMMARY
          echo "Claude findings: ${{ steps.claude-security.outputs.findings-count }}" >> $GITHUB_STEP_SUMMARY
          echo "Results file: ${{ steps.claude-security.outputs.results-file }}" >> $GITHUB_STEP_SUMMARY

flowchart TD
    A["开发者编写代码"] --> B["本地 /security-review"]
    B --> B1{发现漏洞?}
    B1 -->|是| B2["Claude生成修复补丁"]
    B2 --> B3["运行测试验证"]
    B3 --> B4["修复通过?"]
    B4 -->|是| C["Git commit + push"]
    B4 -->|否| B2
    B1 -->|否| C

    C --> D["创建Pull Request"]
    D --> E["GitHub Action触发"]

    E --> F["Claude Code Security Review"]
    E --> G["npm audit依赖扫描"]
    E --> H["Semgrep模式匹配"]

    F --> I["语义安全报告"]
    G --> J["依赖漏洞报告"]
    H --> K["模式匹配报告"]

    I --> L{"合并检查"}
    J --> L
    K --> L

    L -->|Critical: 阻止| M["要求修复"]
    L -->|High: 警告| N["建议修复"]
    L -->|Low: 记录| O["允许合并"]

    M --> A
    N --> P["人工审查"]
    P --> Q{批准?}
    Q -->|是| R["合并到main"]
    Q -->|否| A
    O --> R

    R --> S["Claude Security定时扫描"]
    S --> T["持续安全态势监控"]

    style F fill:#f96,stroke:#333,stroke-width:2px
    style L fill:#bbf,stroke:#333,stroke-width:2px
    style R fill:#bfb,stroke:#333,stroke-width:2px

---
name: security-audit
description: OWASP Top 10 compliance audit for payment applications
---

# Security Audit Skill

Perform a comprehensive OWASP Top 10 (2025) audit:

## Checklist
1. **A01 - Broken Access Control**: Check all API routes for proper authorization
2. **A02 - Cryptographic Failures**: Verify TLS, encryption at rest, key rotation
3. **A03 - Injection**: Trace all user inputs to database queries and shell commands
4. **A04 - Insecure Design**: Review business logic for abuse scenarios
5. **A05 - Security Misconfiguration**: Check CORS, headers, debug modes
6. **A06 - Vulnerable Components**: Review package.json for known CVEs
7. **A07 - Auth Failures**: Test session management, brute force protection
8. **A08 - Data Integrity**: Verify input validation and deserialization safety
9. **A09 - Logging Failures**: Ensure security events are logged without PII
10. **A10 - SSRF**: Check all outbound HTTP requests for user-controlled URLs

For each category, read the relevant files, trace data flows, and report findings.
Generate a compliance score (0-100) for each category.

> /security-audit

---
name: security-reviewer
description: Dedicated security code reviewer
tools: Read, Grep, Glob, Bash
model: opus
---

You are a senior application security engineer with 15 years of experience
in penetration testing and secure code review.

Review code for these specific concerns:
- Injection vulnerabilities (SQL, XSS, command injection)
- Authentication and authorization flaws
- Secrets or credentials in code
- Insecure data handling and PII exposure
- Race conditions and TOCTOU issues
- Insecure deserialization
- SSRF and CSRF vulnerabilities

Provide specific line references with file paths.
For each finding, rate severity (Critical/High/Medium/Low)
and provide a concrete fix.

> use the security-reviewer subagent to review all changes in this PR

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Write|Edit",
        "command": "python3 .claude/hooks/check-secrets.py \"$TOOL_INPUT\""
      }
    ],
    "PostToolUse": [
      {
        "matcher": "Write|Edit",
        "command": "python3 .claude/hooks/check-security.py \"$TOOL_INPUT\""
      }
    ]
  }
}

#!/usr/bin/env python3
"""Hook: Block writes containing hardcoded secrets"""
import sys, json, re

input_data = json.loads(sys.argv[1])
content = input_data.get('content', '')
file_path = input_data.get('file_path', '')

# Skip test files
if '/test' in file_path or '/tests/' in file_path:
    sys.exit(0)

patterns = [
    (r'(?:password|passwd|pwd)\s*[=:]\s*["\'][^"\']{4,}', 'Hardcoded password'),
    (r'(?:api[_-]?key|apikey)\s*[=:]\s*["\'][^"\']{8,}', 'Hardcoded API key'),
    (r'(?:secret|token)\s*[=:]\s*["\']sk-[a-zA-Z0-9]{20,}', 'Hardcoded secret/token'),
    (r'-----BEGIN (?:RSA |EC )?PRIVATE KEY-----', 'Private key in code'),
]

for pattern, description in patterns:
    if re.search(pattern, content, re.IGNORECASE):
        print(f"SECURITY BLOCK: {description} detected in {file_path}")
        sys.exit(2)  # Block the tool call

sys.exit(0)  # Allow the tool call

jobs:
  security-layers:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      # 层1：快速模式匹配
      - name: Semgrep Scan
        uses: semgrep/semgrep-action@v1
        with:
          config: p/owasp-top-ten

      # 层2：语义安全推理
      - uses: anthropics/claude-code-security-review@main
        with:
          claude-api-key: ${{ secrets.CLAUDE_API_KEY }}
          comment-pr: true

      # 层3：依赖漏洞扫描
      - name: Dependency Review
        uses: actions/dependency-review-action@v4