mcp-security
具有5层防御架构的多代理和MCP管道安全。在构建MCP服务器、多代理系统或处理用户输入以防止提示注入并确保适当授权的任何管道时使用。
📖 中文导读:具有5层防御架构的多代理和MCP管道安全。在构建MCP服务器、多代理系统或处理用户输入以防止提示注入并确保适当授权的任何管道时使用。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。
name: mcp-security description: Multi-agent and MCP pipeline security with 5-layer defense architecture. Use when building MCP servers, multi-agent systems, or any pipeline that handles user input to prevent prompt injection and ensure proper authorization.
MCP Security Skill
This skill enforces security best practices for MCP servers and multi-agent pipelines.
5-Layer Defense Architecture
- Input Validation - Sanitize all user inputs
- Prompt Injection Prevention - Detect and block injection attempts
- SQL/NoSQL Validation - Prevent query injection
- User Context Propagation - Maintain identity through pipeline
- Authorization (RBAC/ABAC) - Enforce access controls
Prompt Injection Prevention
# Always validate and sanitize inputs
def sanitize_input(user_input: str) -> str:
# Remove potential injection patterns
# Escape special characters
# Limit length
pass
# Never directly concatenate user input into prompts
# ❌ Bad
prompt = f"Process this: {user_input}"
# ✅ Good
prompt = sanitize_input(user_input)
validated_prompt = validate_against_schema(prompt)
User Context Propagation
@dataclass
class UserContext:
user_id: str
roles: list[str]
permissions: list[str]
tenant_id: str
# Pass context through all pipeline stages
async def process_request(context: UserContext, request: Request):
# Validate permissions at each step
if not has_permission(context, "read:data"):
raise AuthorizationError()
Authorization Patterns
RBAC (Role-Based Access Control)
ROLE_PERMISSIONS = {
"admin": ["read", "write", "delete", "admin"],
"editor": ["read", "write"],
"viewer": ["read"],
}
ABAC (Attribute-Based Access Control)
def can_access(user: User, resource: Resource) -> bool:
return (
user.department == resource.department
and user.clearance >= resource.sensitivity
)
Security Checklist
- All user inputs validated and sanitized
- Prompt injection patterns detected
- SQL queries parameterized
- User context propagated through pipeline
- Authorization checked at each step
- Sensitive data encrypted
- Audit logging enabled