code-security-audit（代码安全审计）

Code Security Audit

AI-powered security audit for code changes with false positive filtering. Based on claude-code-security-review.

Bundled Resources

claude-code-security-review/
├── SKILL.md                          # This file — workflow and instructions
├── action.yml                        # GitHub Actions composite action definition
├── claudecode/                       # Python package (imports: from claudecode.*)
│   ├── __init__.py                   # Package init — re-exports main entry points
│   ├── audit.py                      # Entry point — wraps github_action_audit
│   ├── github_action_audit.py        # GitHub Action audit runner (PR fetch, Claude runner, filter pipeline)
│   ├── prompts.py                    # Security audit prompt templates
│   ├── findings_filter.py            # Hard exclusion rules + Claude API false positive filtering
│   ├── claude_api_client.py          # Claude API client for single-finding analysis
│   ├── json_parser.py               # Robust JSON extraction from text (code blocks, nested braces)
│   ├── constants.py                  # Config: model name, timeouts, token limits, exit codes
│   ├── logger.py                     # Stderr logging with GitHub context prefix
│   ├── requirements.txt              # Python deps: anthropic, requests, PyGithub
│   └── evals/                        # Evaluation framework
│       ├── run_eval.py               # CLI: python -m claudecode.evals.run_eval owner/repo#123
│       └── eval_engine.py            # Git worktree management + SAST runner
├── scripts/                          # Standalone scripts (non-Python-package)
│   └── comment-pr-findings.js        # Node.js script to post findings as PR review comments
├── references/                       # Knowledge loaded into context as needed
│   ├── false-positive-filtering.md   # 20 hard exclusions, signal quality criteria, 12 precedents
│   ├── custom-scan-instructions.md   # Industry templates (compliance, finserv, e-commerce, GraphQL)
│   ├── custom-false-positive-filtering.txt   # Example custom FP filtering rules
│   └── custom-security-scan-instructions.txt # Example custom scan categories
└── assets/                           # Files used in output
    └── security-review-command.md    # Claude Code /security-review slash command template

Manual Audit Workflow

Use this when auditing code changes directly (without CI scripts).

1. Gather Changes

git diff --merge-base origin/main        # Branch diff
git diff --cached                         # Staged changes
git diff HEAD~N                           # Last N commits
git diff --name-only origin/main...       # List modified files

2. Three-Phase Analysis

Phase 1 — Context Research: Identify security frameworks, ORMs, auth libraries, sanitization patterns, and trust boundaries in the codebase.

Phase 2 — Comparative Analysis: Compare new code against established secure patterns. Flag deviations, inconsistencies, and new attack surfaces.

Phase 3 — Vulnerability Assessment: Check each modified file for:

• Input Validation: SQL injection, command injection, XXE, template injection, NoSQL injection, path traversal
• Auth & Authz: Auth bypass, privilege escalation, session flaws, JWT vulnerabilities
• Crypto & Secrets: Hardcoded keys/tokens, weak algorithms, improper key storage
• Code Execution: RCE via deserialization, pickle/YAML injection, eval injection, XSS
• Data Exposure: Sensitive data logging, PII violations, API leakage, debug exposure

Trace data flow from user inputs to sensitive operations. Look for privilege boundary crossings.

3. Filter False Positives

Load references/false-positive-filtering.md and apply all rules. Assign confidence 1-10 per finding; only keep findings with confidence ≥ 8.

For domain-specific categories, load references/custom-scan-instructions.md.

4. Output Format

# Vuln N: [Category]: `file.ts:42`

* Severity: HIGH | MEDIUM
* Confidence: 8/10
* Description: [What the vulnerability is]
* Exploit Scenario: [Concrete attack path]
* Recommendation: [Specific fix]

GitHub Action Integration

Run automated security audits on PRs via CI. Requires ANTHROPIC_API_KEY and GITHUB_TOKEN.

Setup

pip install -r claudecode/requirements.txt

Environment Variables

Run

python claudecode/audit.py

Output is JSON with findings, analysis_summary, and filtering_summary.

GitHub Actions Workflow

name: Security Review
permissions:
  pull-requests: write
  contents: read
on:
  pull_request:
jobs:
  security:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          ref: ${{ github.event.pull_request.head.sha }}
          fetch-depth: 2
      - uses: anthropics/claude-code-security-review@main
        with:
          comment-pr: true
          claude-api-key: ${{ secrets.CLAUDE_API_KEY }}

PR Commenting

Post findings as inline review comments:

node scripts/comment-pr-findings.js

Reads findings.json from cwd. Requires GITHUB_TOKEN and GITHUB_EVENT_PATH.

Example Customization Files

See references/ for sample customization files:

• references/custom-false-positive-filtering.txt — Template for custom FP filtering rules
• references/custom-security-scan-instructions.txt — Template for custom scan categories

Evaluation Framework

Test the audit against any public PR:

export ANTHROPIC_API_KEY=sk-...
python -m claudecode.evals.run_eval owner/repo#123 --verbose

Results saved to ./eval_results/ as JSON with findings, runtime, and success status.

Slash Command

Copy assets/security-review-command.md to .claude/commands/security-review.md in any project to enable /security-review in Claude Code.

Key Principles

1. Minimize false positives — only flag issues with >80% confidence of exploitability
2. Skip noise — no theoretical issues, style concerns, or low-impact findings
3. Focus on impact — prioritize unauthorized access, data breaches, system compromise
4. Only new issues — do not comment on pre-existing security concerns
5. Better to miss theoretical issues than flood with false positives

Severity Guidelines

• HIGH: Directly exploitable → RCE, data breach, auth bypass
• MEDIUM: Requires specific conditions but significant impact
• Do NOT report LOW severity findings