MCP Server Security Scan

Defensive scan for MCP servers (@modelcontextprotocol/sdk Node, mcp Python). Reports findings using the shared scoring schema.

Scope

• Files importing @modelcontextprotocol/sdk/* or mcp (Python)
• Tool / resource / prompt registrations (server.tool, @server.list_tools, etc.)
• Transport setup (StdioServerTransport, SSEServerTransport, StreamableHTTPServerTransport)

Threat model recap

MCP tools execute on the server's trust boundary but are invoked at the LLM's discretion based on a user's conversation. Treat tool arguments as fully attacker-controlled: a user can craft chat input or upload a document that causes the model to call your tool with adversarial arguments (indirect prompt injection).

Rules