Claude Code 安全優先開發
如何確保您的 Claude Code 工作流程產生安全的程式碼。從 security-auditor agent 到 PreToolUse hooks 和自動化審計。
2026 年 10 月 21 日更新: 補上 2026 年的權限模式預設值(互動式 session 自 v2.1.283 起以 auto 模式啟動、
claude -p自 v2.1.302 起以 Manual 模式啟動)、Bash sandbox 設定、--restricted與permissions.blockReadsOutsideWorkingDirectories;釐清 security-auditor 是自訂 subagent(內建的檢查指令是/security-review);依官方文件修正 prompt hook 與結束碼的寫法;並更新 managed settings 路徑與模型。
AI 輔助開發帶來一個重要問題:如何確保 Claude 編寫的程式碼是安全的?
答案不是避免使用 AI 輔助,而是使用 Claude Code 的官方工具將安全性融入您的工作流程:權限模式與規則、Bash sandbox、hooks、內建的 /security-review 指令,再加上自訂的 security-auditor subagent。
本指南涵蓋如何使用官方功能配置 Claude Code 進行安全意識開發。
官方安全工具
security-auditor Agent
security-auditor 並不是內建的 agent,而是你自己定義的自訂 subagent(Claude Code 內建的 subagent 是 Explore、Plan 與 general-purpose)。若想不做任何設定就立即檢查,可以執行內建的 /security-review 指令,它會審查目前分支的變更是否有漏洞。要建立這個 subagent,請新增如下檔案:
---
name: security-auditor
description: Reviews code changes for security vulnerabilities. Use for auth, payment, and admin changes.
tools: Read, Glob, Grep
model: sonnet
---
You are a security auditor. Check changes for OWASP Top 10 issues,
hardcoded secrets, injection, XSS, and authentication bypasses, and
report each finding with its severity.
把它存成 .claude/agents/security-auditor.md。之後 Claude 會透過 Agent 工具(v2.1.63 之前叫 Task)委派給它:
Agent({
subagent_type: "security-auditor",
model: "sonnet", // or "opus" for critical code
prompt: `
Audit the authentication module for security vulnerabilities.
Check for:
- OWASP Top 10 vulnerabilities
- Hardcoded secrets
- SQL injection patterns
- XSS vulnerabilities
- Authentication bypass risks
`
})
何時使用 security-auditor:
| 場景 | 模型 | 原因 |
|---|---|---|
| 認證/支付變更 | Sonnet/Opus | 關鍵程式碼 |
| 新 API 端點 | Sonnet | 輸入驗證 |
| 檔案上傳處理 | Sonnet | 清理處理 |
| 外部 API 整合 | Sonnet | 資料暴露 |
自動安全執行的 Hooks
根據官方 hooks 文件,hooks 可以在無需手動介入的情況下自動執行安全政策。
PreToolUse Hook - 阻擋危險操作
在 .claude/settings.json 中配置:
{
"hooks": {
"PreToolUse": [{
"matcher": "Bash",
"hooks": [{
"type": "prompt",
"prompt": "If the command contains destructive operations (rm -rf, DROP TABLE, git push --force), respond with {\"ok\": false, \"reason\": \"Destructive command: confirm with the user first\"}. Otherwise respond with {\"ok\": true}.",
"continueOnBlock": true
}]
}]
}
}
Prompt hook 只能回答 {"ok": true} 或 {"ok": false, "reason": "..."}。在 PreToolUse 上,ok: false 會拒絕該次呼叫,而 continueOnBlock: true 會把原因交還給 Claude,讓它調整做法。如果希望 Claude Code 在特定指令前先詢問你,請在 permissions.ask 下加入 "Bash(git push --force *)" 這類 ask 規則。
可用的 hook 事件:
| 事件 | 時機 | 安全用途 |
|---|---|---|
PreToolUse | 工具執行前 | 阻擋危險指令 |
PostToolUse | 工具執行後 | 審計檔案變更 |
Stop | Claude 完成回應時 | 驗證安全審查 |
SessionStart | 會話開始時 | 載入安全上下文 |
Stop Hook - 驗證安全審查
{
"hooks": {
"Stop": [{
"hooks": [{
"type": "agent",
"prompt": "Run git diff --name-only. If files in auth/, payment/, or admin/ changed and the transcript shows no security-auditor review, respond with {\"ok\": false, \"reason\": \"Run the security-auditor subagent on the changed files before finishing\"}. Otherwise respond with {\"ok\": true}. $ARGUMENTS"
}]
}]
}
}
Prompt hook 只看得到 hook 的輸入,所以這項檢查改用 agent hook:它能先執行指令、讀取對話紀錄再做判斷。Agent hook 目前仍屬實驗性功能。在 Stop 上,ok: false 會讓 Claude 以該原因作為下一步指示繼續工作。
Permission Settings
根據 Claude Code 文件,使用 settings.json 進行存取控制:
{
"permissions": {
"allow": [
"Read",
"Write",
"Edit",
"Bash(npm run *)",
"Bash(git *)"
],
"deny": [
"Read(.env)",
"Read(.env.*)",
"Read(secrets/**)",
"Read(**/*credentials*)",
"Bash(*rm -rf*)",
"Bash(*DROP*)"
]
}
}
權限模式語法:
Bash(npm run *)- 允許所有 npm run 指令(Bash(npm run:*)是等效的舊寫法)Read(.env)- 拒絕讀取 .env 檔案Read(secrets/**)- 拒絕讀取 secrets/ 中的任何內容Bash(*DROP*)- 拒絕包含 DROP 的指令
規則依 deny → ask → allow 的順序評估,而 deny 規則在每一種權限模式(包括 bypassPermissions)都有效。檔案路徑規則請寫成 Read(...) 或 Edit(...):Claude Code 從不檢查 Write(路徑) 規則。
權限模式與預設值(2026)
權限模式決定 Claude 行動前是否要先詢問,而預設值在 2026 年有了變化:
- 互動式 session 以 auto 模式啟動。 自 Claude Code v2.1.283 起,沒有設定權限模式的終端機與 VS Code session 會以 auto 模式啟動:由一個分類器模型代替你審查每個動作,而不是逐一詢問。明確的
ask規則仍會跳出詢問,組織也可以關閉 auto 模式。 - 非互動執行以 Manual 模式啟動。 自 v2.1.302 起,沒有設定模式的
claude -p與 Agent SDK 執行會以 Manual 模式(設定值為default)啟動,任何需要核准的呼叫都會被拒絕,除非有 allow 規則、hook 或--permission-prompt-tool核准。 - 受保護路徑依然受到保護。 對
.git、.claude、shell 啟動檔與.mcp.json等路徑的寫入,在 Manual 與 acceptEdits 模式會詢問你、在 auto 模式交給分類器判斷、在 dontAsk 模式直接拒絕;設定檔中的 allow 規則無法預先核准這些寫入。
對於敏感的儲存庫,請固定起始模式,並讓檔案工具留在專案內:
{
"permissions": {
"defaultMode": "default",
"blockReadsOutsideWorkingDirectories": true
}
}
"default" 就是 Manual 模式,會在編輯、執行 shell 指令與存取網路前詢問。專案的 .claude/settings.json 可以設定 auto 與 bypassPermissions 以外的任何起始模式。blockReadsOutsideWorkingDirectories(v2.1.257+)會讓 Read、Grep、Glob 與 LSP 在每一種模式(包括 bypassPermissions)下都拒絕讀取工作目錄以外的路徑;任何一個設定檔設為 true 就會生效,所以儲存庫可以替自己開啟,卻無法關掉你的設定。管理員可以在 managed settings 中把 permissions.disableAutoMode 與 permissions.disableBypassPermissionsMode 設為 "disable",移除 auto 模式或 bypass 模式。
Sandbox 與 Restricted 模式
權限規則管的是 Claude 的工具;Bash sandbox 則在作業系統層級限制 Claude 執行的 shell 指令(macOS、Linux 與 WSL2)。它預設關閉:執行 /sandbox,或在設定中開啟:
{
"sandbox": {
"enabled": true,
"failIfUnavailable": true,
"allowUnsandboxedCommands": false,
"filesystem": {
"denyRead": ["~/.aws/credentials"]
},
"network": {
"allowedDomains": ["github.com", "*.npmjs.org"]
}
}
}
failIfUnavailable 會在 sandbox 無法啟動時讓 Claude Code 直接結束,而不是在 sandbox 外執行指令;allowUnsandboxedCommands: false(嚴格 sandbox 模式)則禁止 Claude 把被擋下的指令改到 sandbox 外重試。Sandbox 只涵蓋 shell 指令;檔案工具、hooks 與本機 MCP 伺服器都在 sandbox 之外執行,並遵循權限規則。
在評估用的測試環境或共用機器上,請以 --restricted(或 CLAUDE_CODE_RESTRICTED=1,v2.1.248+)啟動 Claude Code。它會移除會執行指令或程式碼的內建工具與 WebFetch(除非你在 --tools 中指名),讓檔案工具只能存取工作目錄,忽略使用者、專案與本機設定,並拒絕 bypassPermissions。
安全挑戰
當 Claude 編寫程式碼時,存在幾個風險:
- 從訓練資料複製的不安全模式
- 缺少輸入驗證
- 洩露資訊的不當錯誤處理
- 硬編碼的憑證或密鑰
- SQL injection、XSS 和其他漏洞
解決方案:使用 agents 和 hooks 將您的安全需求程式碼化,讓 Claude 自動遵循。
建立安全優先的 CLAUDE.md
安全章節範本
# Security Requirements
## Non-Negotiables
These rules cannot be overridden for any reason:
1. **No Secrets in Code**
- Never hardcode credentials, API keys, or tokens
- Use environment variables for all secrets
- Reference secrets through secure config management
2. **Input Validation**
- All user inputs must be validated
- Use allowlist validation over blocklist
- Validate on both client and server
3. **Output Encoding**
- Encode all outputs for the context (HTML, URL, JS)
- Use framework-provided escaping mechanisms
- Never concatenate raw user input into responses
4. **Authentication & Authorization**
- All endpoints must verify authentication
- Check authorization before every action
- Use principle of least privilege
5. **Data Protection**
- Encrypt sensitive data at rest and in transit
- Minimize data collection and retention
- Log access to sensitive data
## Security Review Requirements
These types of changes require explicit security review:
- Authentication/authorization logic
- Payment processing
- Personal data handling
- API endpoint creation
- Database queries
- File uploads
- Email sending
- External API calls
輸入驗證標準
# Input Validation Standards
## Validation Library
Use Zod for all runtime validation:
\`\`\`typescript
import { z } from 'zod';
// Define schema
const UserSchema = z.object({
email: z.string().email(),
age: z.number().min(0).max(150),
name: z.string().min(1).max(100),
});
// Validate
const result = UserSchema.safeParse(input);
if (!result.success) {
throw new ValidationError(result.error);
}
\`\`\`
## Validation Rules
- String inputs: Define max length
- Numeric inputs: Define valid range
- Email: Use proper email validation
- URLs: Validate protocol and domain
- IDs: Validate format (UUID, numeric, etc.)
- Arrays: Validate length and item types
- Objects: Validate required fields
安全編碼模式
# Secure Coding Patterns
## Database Queries
ALWAYS use parameterized queries:
\`\`\`typescript
// NEVER do this
const query = \`SELECT * FROM users WHERE email = '\${email}'\`;
// ALWAYS do this
const user = await db.user.findUnique({
where: { email }, // Prisma handles parameterization
});
\`\`\`
## Error Handling
Never expose internal details:
\`\`\`typescript
// NEVER do this
catch (error) {
return res.status(500).json({ error: error.message });
}
// ALWAYS do this
catch (error) {
console.error('Internal error:', error); // Log full error
return res.status(500).json({ error: 'An error occurred' }); // Generic message
}
\`\`\`
## Authentication Checks
Always verify auth at the start:
\`\`\`typescript
export async function handler(req, res) {
// First: Authenticate
const user = await getAuthenticatedUser(req);
if (!user) {
return res.status(401).json({ error: 'Unauthorized' });
}
// Second: Authorize
if (!user.canAccessResource(resourceId)) {
return res.status(403).json({ error: 'Forbidden' });
}
// Then: Proceed with logic
// ...
}
\`\`\`
安全的 Agent 委派
配置 Claude 對敏感變更自動進行安全審查:
# Agent Delegation Rules
## Automatic Security Review
These changes automatically trigger security-auditor agent:
- Files in: auth/, payment/, admin/
- Functions containing: password, token, secret, key, credential
- Operations: user creation, permission changes, data deletion
- External calls: API integrations, webhooks, email
## Security Agent Configuration
When security-auditor runs, it checks for:
- OWASP Top 10 vulnerabilities
- Hardcoded secrets
- Missing authentication
- SQL injection patterns
- XSS vulnerabilities
- Insecure direct object references
- Missing rate limiting
- Insufficient logging
使用 Agent Tool 實作
// Automatic security review for auth changes
Agent({
subagent_type: "security-auditor",
model: "sonnet",
prompt: `
Review changes to auth/ directory.
Check for OWASP Top 10 vulnerabilities:
1. Injection
2. Broken authentication
3. Sensitive data exposure
4. XML external entities
5. Broken access control
6. Security misconfiguration
7. Cross-site scripting
8. Insecure deserialization
9. Using components with known vulnerabilities
10. Insufficient logging
Report findings with severity levels.
`
})
安全工作流程模式
模式 1:實作前審查
在編寫安全敏感程式碼之前:
Agent({
subagent_type: "security-auditor",
model: "sonnet",
prompt: `
Planning to implement password reset functionality.
Review security considerations:
1. What vulnerabilities should I address?
2. What OWASP vulnerabilities are relevant?
3. What's the secure implementation pattern?
4. What tests should I write?
`
})
模式 2:帶有防護機制的實作
使用自動 hooks 進行實作:
{
"hooks": {
"PostToolUse": [{
"matcher": "Edit|Write",
"hooks": [{
"type": "prompt",
"prompt": "If tool_input.file_path is in an auth/, payment/, or admin/ directory, respond with {\"ok\": false, \"reason\": \"Security-sensitive file changed: run the security-auditor subagent on it\"}. Otherwise respond with {\"ok\": true}.",
"continueOnBlock": true
}]
}]
}
}
模式 3:實作後審計
實作完成後:
Agent({
subagent_type: "security-auditor",
model: "opus", // Use Opus for critical security review
prompt: `
Comprehensive security audit of password reset feature.
Check for:
- Token generation entropy
- Timing attacks in token comparison
- Rate limiting bypass
- User enumeration
- Session fixation after reset
- Proper invalidation of old sessions
Provide findings categorized by severity:
- Critical
- High
- Medium
- Low
`
})
常見漏洞與預防
SQL Injection
有漏洞的:
const query = `SELECT * FROM users WHERE id = ${userId}`;
安全的:
const user = await prisma.user.findUnique({ where: { id: userId } });
Hook 保護:
{
"hooks": {
"PreToolUse": [{
"matcher": "Edit|Write",
"hooks": [{
"type": "prompt",
"prompt": "If the new code builds SQL by concatenating strings with SQL keywords (SELECT, INSERT, UPDATE, DELETE), respond with {\"ok\": false, \"reason\": \"Possible SQL injection: use parameterized queries\"}. Otherwise respond with {\"ok\": true}.",
"continueOnBlock": true
}]
}]
}
}
Cross-Site Scripting (XSS)
有漏洞的:
<div dangerouslySetInnerHTML={{ __html: userContent }} />
安全的:
<div>{userContent}</div> // React auto-escapes
CLAUDE.md 規則:
## XSS Prevention
- Never use dangerouslySetInnerHTML without sanitization
- Use DOMPurify for HTML content that must be rendered
- Escape all dynamic content in non-React contexts
認證繞過
有漏洞的:
if (user.role === 'admin') {
// Allow action
}
安全的:
if (await authService.hasPermission(user.id, 'admin:action')) {
// Allow action
}
敏感資料暴露
有漏洞的:
return res.json({ user }); // Includes password hash!
安全的:
return res.json({ user: sanitizeUser(user) });
持續安全的 Hooks
預提交安全檢查
{
"hooks": {
"Stop": [{
"hooks": [{
"type": "command",
"command": "npm run security-scan >&2 || exit 2"
}]
}]
}
}
Stop hook 以結束碼 2 結束時,會把 hook 的 stderr 當成原因,讓 Claude 回去繼續工作。onFailure: "block" 欄位(v2.1.295+)對 Stop hook 沒有作用,所以掃描失敗時必須由指令本身回傳結束碼 2。
自動密鑰檢測
{
"hooks": {
"PreToolUse": [{
"matcher": "Edit|Write",
"hooks": [{
"type": "prompt",
"prompt": "Scan the new content for potential secrets: API keys, passwords, tokens, private keys. If found, respond with {\"ok\": false, \"reason\": \"Possible secret on lines ...\"}. Otherwise respond with {\"ok\": true}.",
"continueOnBlock": true
}]
}]
}
}
檔案存取控制
{
"permissions": {
"deny": [
"Read(.env*)",
"Read(**/secrets/**)",
"Read(**/*credential*)",
"Read(**/*.pem)",
"Read(**/*.key)",
"Edit(.env*)",
"Edit(**/secrets/**)"
]
}
}
使用 test-runner 進行安全測試
和 security-auditor 一樣,test-runner 也是你在 .claude/agents/ 中自訂的 subagent。
自動化安全測試套件
// Launch parallel security tests
Agent({
subagent_type: "test-runner",
model: "haiku",
prompt: "Run all security-related tests in tests/security/"
})
Agent({
subagent_type: "security-auditor",
model: "sonnet",
prompt: "Verify test coverage for authentication edge cases"
})
安全測試範例
describe('Password Reset Security', () => {
it('rejects invalid tokens', async () => {
const response = await request(app)
.post('/api/auth/reset-password')
.send({ token: 'invalid', password: 'newpassword' });
expect(response.status).toBe(400);
expect(response.body).not.toContain('token');
});
it('prevents timing attacks on token validation', async () => {
// Use constant-time comparison
const times = [];
for (let i = 0; i < 100; i++) {
const start = performance.now();
await validateToken(`token${i}`);
times.push(performance.now() - start);
}
// Variance should be minimal
expect(standardDeviation(times)).toBeLessThan(5);
});
it('rate limits reset requests', async () => {
const email = 'test@example.com';
// First 3 requests succeed
for (let i = 0; i < 3; i++) {
const response = await request(app)
.post('/api/auth/forgot-password')
.send({ email });
expect(response.status).toBe(200);
}
// 4th request is rate limited
const response = await request(app)
.post('/api/auth/forgot-password')
.send({ email });
expect(response.status).toBe(429);
});
});
安全審計檢查清單
在任何部署之前:
## Pre-Deployment Security Checklist
### Authentication
- [ ] All endpoints require authentication (except public ones)
- [ ] Tokens expire appropriately
- [ ] Password requirements enforced
- [ ] Account lockout after failed attempts
### Authorization
- [ ] Permission checks on all protected actions
- [ ] No privilege escalation paths
- [ ] Admin functions properly protected
### Data Protection
- [ ] Sensitive data encrypted at rest
- [ ] HTTPS enforced
- [ ] No secrets in code or logs
- [ ] PII handling compliant
### Input/Output
- [ ] All inputs validated
- [ ] All outputs properly encoded
- [ ] File uploads validated and sanitized
- [ ] No SQL injection vulnerabilities
### Logging & Monitoring
- [ ] Security events logged
- [ ] Logs don't contain sensitive data
- [ ] Alerting configured for anomalies
企業安全功能
Claude Code 透過設定層級支援企業安全需求:managed settings 的優先序高於命令列、local、專案與使用者設定。
企業層(Managed Settings)
// /etc/claude-code/managed-settings.json(Linux 與 WSL;由 IT 管理)
{
"permissions": {
"deny": [
"Bash(*curl*)",
"Bash(*wget*)",
"Read(//etc/**)",
"Edit(//etc/**)"
],
"disableBypassPermissionsMode": "disable"
},
"allowedMcpServers": [
{ "serverName": "filesystem" },
{ "serverName": "memory" }
]
}
在 macOS 上此檔案位於 /Library/Application Support/ClaudeCode/managed-settings.json,Windows 則是 C:\Program Files\ClaudeCode\managed-settings.json。在 Read 與 Edit 規則中,絕對路徑要用 // 開頭。
審計日誌配置
## Audit Requirements
Log these events to secure audit system:
- Authentication attempts (success/failure)
- Authorization decisions
- Data access (especially sensitive data)
- Configuration changes
- API calls to external services
合規整合
## Compliance Requirements
- PCI DSS: No card data in logs, encrypt at rest
- GDPR: Data minimization, right to deletion
- HIPAA: PHI access controls and audit trails
- SOC 2: Change management documentation
安全的模型選擇
| 任務 | 推薦模型 | 原因 |
|---|---|---|
| 快速安全掃描 | Haiku 5.5 | 快速模式匹配 |
| 程式碼審查 | Sonnet 5.5 | 平衡的分析 |
| 安全審計 | Opus 5.5 | 徹底審查;官方文件建議尋找漏洞時使用 max effort |
| 關鍵安全 | Fable 5.1 | 最高智能 |
修正自己程式碼中的 bug 對這些模型而言是日常工作,但以安全為主題的請求可能會碰到安全防護機制。Anthropic 的 Opus 5.5 發表文章寫道:「Users will be able to identify and fix bugs in their code as part of the routine software development lifecycle, but most cybersecurity tasks will be re-routed to Opus 4.8.」(來源)被 Fable 模型安全分類器標記的請求(最常見於資安領域)會觸發自動模型 fallback,滲透測試這類攻擊性安全工作更是經常觸發。對於這類工作,Anthropic 建議組織洽詢其受信任存取計畫,例如 Cyber Verification Program。
快速模型切換(v2.0.65+):在提示時按 Option+P(macOS)或 Alt+P(Windows/Linux)。
開始行動
今天:
- 在 CLAUDE.md 中添加 security-auditor 委派規則
- 為敏感檔案配置權限拒絕
- 為危險指令添加一個 PreToolUse hook
- 對目前分支執行
/security-review
本週:
- 實作用於安全驗證的 Stop hook
- 添加安全測試套件
- 對關鍵模組執行 security-auditor
本月:
- 部署預提交安全 hooks
- 在 CI/CD 中整合安全掃描
- 培訓團隊進行安全優先開發
安全是一段旅程,而非終點。透過 Claude Code 的 agents 和 hooks,您可以自動化安全最佳實踐。
資料來源:Claude Code 文件、安全性、權限模式、Sandboxing、Hooks 文件、Claude Code GitHub