MCP Security Threat Demo

A self-contained CLI demo of 8 real-world MCP (Model Context Protocol) security threats, built for interview/presentation use.

Architecture

main.py  →  ThreatScenario  →  DemoAgent  →  MCPHost  →  MCP Server (subprocess)
                                                               ↕ stdio JSON-RPC
                                                          SQLite demo.db

Components:

• AI Agent — scripted demo mode (default) or real Claude claude-sonnet-4-6 via Anthropic API
• MCP Host — manages MCP server subprocess lifecycle via stdio transport
• MCP Server — FastMCP server with 18 intentionally vulnerable tools
• Database — SQLite with 15 fake employees, tokens, secrets, and audit logs

Quick Start

pip install -r requirements.txt
python main.py

Real Claude AI Mode

export ANTHROPIC_API_KEY=sk-ant-...
export DEMO_MODE=false
python main.py

8 Threat Scenarios

#	Threat	Severity
1	Prompt Injection / SSRF	CRITICAL
2	Auth & Access Abuse	HIGH
3	Data Exfiltration	HIGH
4	Privilege Escalation	HIGH
5	Resource Exhaustion	MEDIUM
6	Supply Chain Risk	MEDIUM
7	Business Logic Abuse	HIGH
8	Shadow Agent Activity	MEDIUM

Each scenario shows: Setup → Attack Steps → Impact Analysis → Why Dangerous → Mitigations

Requirements

• Python 3.11+
• See requirements.txt