OpenCode 中文社区

Appearance

5.2c 权限与安全 ​

💡 一句话总结：配置 Agent 的权限控制，确保 AI 安全可控。

学完你能做什么 ​

  • 理解 Agent 权限控制的重要性
  • 能配置 Agent 的工具权限
  • 能设置访问控制策略
  • 能防止安全风险

🎒 开始前的准备 ​

确保你已经完成以下事项：

核心思路 ​

为什么需要权限控制 ​

权限控制可以：

  • 防止 Agent 执行危险操作
  • 限制敏感数据的访问
  • 确保操作可追溯
  • 满足合规要求

权限配置 ​

工具权限 ​

yaml
agents:
  safe-agent:
    name: "安全助手"
    description: "受限操作的助手"
    permissions:
      tools:
        allowed:
          - read        # 允许读取文件
          - write       # 允许写入文件
          - grep        # 允许搜索
          - glob        # 允许文件匹配
        blocked:
          - bash        # 阻止执行命令
          - delete      # 阻止删除操作
          - edit        # 阻止编辑（只读）
        read_only:
          - true        # 所有文件只读
        path_restrictions:
          allowed:
            - "src/**"
            - "tests/**"
          blocked:
            - "**/secrets/**"
            - "**/.env"
            - "**/*.pem"

操作权限 ​

yaml
agents:
  limited-agent:
    name: "受限助手"
    description: "操作受限的助手"
    permissions:
      operations:
        file_write:
          allowed: true
          require_confirmation: true
          max_file_size: "1MB"
        
        bash_execution:
          allowed: false
        
        network_requests:
          allowed: true
          whitelist:
            - "api.example.com"
            - "github.com"
          blacklist:
            - "localhost"
            - "127.0.0.1"
        
        database_write:
          allowed: false
        
        secret_access:
          allowed: false

数据权限 ​

yaml
agents:
  data-safe-agent:
    name: "数据安全助手"
    description: "保护敏感数据的助手"
    permissions:
      data:
        classification:
          public:
            - "*.md"
            - "*.txt"
          internal:
            - "src/**"
            - "docs/**"
          confidential:
            - "**/config/**"
            - "**/secrets/**"
        
        access_levels:
          - level: "public"
            can_read: true
            can_write: true
          
          - level: "internal"
            can_read: true
            can_write: false
          
          - level: "confidential"
            can_read: false
            can_write: false
        
        encryption:
          required: true
          algorithm: "aes-256"

安全策略 ​

危险操作确认 ​

yaml
agents:
  safe-coding-agent:
    name: "安全编码助手"
    description: "需要确认的助手"
    security:
      dangerous_operations:
        confirmation_required:
          - "删除文件"
          - "执行 rm 命令"
          - "修改系统配置"
          - "访问生产环境"
        
        auto_block:
          - "格式化磁盘"
          - "删除所有文件"
          - "修改权限为 777"
        
        require_approval:
          - "部署到生产"
          - "修改数据库"
          - "添加用户权限"
      
      audit:
        enabled: true
        log_all_operations: true
        sensitive_operations:
          - "文件删除"
          - "命令执行"
          - "网络请求"

输入验证 ​

yaml
agents:
  validated-agent:
    name: "输入验证助手"
    description: "验证所有输入的助手"
    security:
      input_validation:
        enabled: true
        
        file_paths:
          max_length: 255
          allowed_chars: "[a-zA-Z0-9/_.-]"
          block_patterns:
            - "../"
            - "..\\"
            - "/etc/passwd"
            - "C:\\Windows"
        
        commands:
          allowed_commands:
            - "ls"
            - "cat"
            - "grep"
            - "find"
          blocked_commands:
            - "rm"
            - "sudo"
            - "chmod"
            - "mkfs"
        
        sql_injection:
          enabled: true
          block_patterns:
            - "' OR '1'='1"
            - "; DROP TABLE"
            - "--"

跟我做 ​

实战：创建一个安全的开发助手 ​

目标：配置一个只能读、不能执行危险操作的 Agent

yaml
agents:
  safe-developer:
    name: "安全开发者"
    description: "只读模式的安全开发助手"
    role: |
      你是一个安全的开发助手，只能查看代码，不能执行任何命令或修改文件。
    
    permissions:
      tools:
        allowed:
          - read        # 只能读取
          - grep        # 只能搜索
          - glob        # 只能查找
          - lsp-*       # LSP 相关工具
        blocked:
          - write       # 阻止写入
          - bash        # 阻止命令执行
          - delete      # 阻止删除
          - edit        # 阻止编辑
      
      path_restrictions:
        allowed:
          - "src/**"
          - "tests/**"
          - "docs/**"
          - "*.md"
          - "*.json"
          - "*.yaml"
          - "*.yml"
        blocked:
          - "**/node_modules/**"
          - "**/.git/**"
          - "**/*.log"
          - ".env*"
          - "**/secrets/**"
          - "**/*.pem"
          - "**/*.key"
    
    security:
      input_validation:
        enabled: true
        
        file_paths:
          max_depth: 10
        
        commands:
          auto_block: true
    
    audit:
      enabled: true
      log_operations:
        - "read"
        - "grep"

📋 权限配置模板 ​

模板 1：只读助手 ​

yaml
agents:
  reader-agent:
    name: "只读助手"
    description: "只能查看代码的助手"
    permissions:
      tools:
        allowed: [read, grep, glob, lsp-*]
        blocked: [write, bash, delete, edit]

模板 2：安全编码助手 ​

yaml
agents:
  safe-coder:
    name: "安全编码助手"
    description: "需要确认的危险操作助手"
    permissions:
      tools:
        allowed: [read, write, grep, glob]
        blocked: [bash]
      
      operations:
        file_write:
          require_confirmation: true
        
        dangerous_commands:
          auto_block: true

模板 3：生产环境助手 ​

yaml
agents:
  production-agent:
    name: "生产环境助手"
    description: "最高安全级别的助手"
    permissions:
      tools:
        allowed: [read, grep]
        blocked: [write, bash, delete, edit, exec]
      
      path_restrictions:
        blocked:
          - "**/*"
          - "!**/docs/**"
      
      operations:
        all_require_approval: true
        audit: true

检查点 ✅ ​

全部通过才能继续

  • [ ] 理解权限控制的重要性
  • [ ] 能配置工具权限
  • [ ] 能设置操作权限
  • [ ] 能配置数据权限

本课小结 ​

你学会了：

  1. 权限控制的必要性
  2. 工具权限配置
  3. 操作权限配置
  4. 数据权限配置
  5. 安全策略配置
  6. 输入验证

下一课预告 ​

下一课我们将学习 Agent 高级技巧，掌握更多优化方法。

📚 更多完整模板：Prompt 模板库

微信群