🛡️ Deception Remote MCP Server - AI-Powered Honeypot

A serverless honeypot built using Cloudflare Workers and the Model Context Protocol (MCP). This project simulates sensitive internal tools to detect unauthorized access attempts and AI agent behaviors using Canarytokens.

🧩 TL;DR

This is a deception-based honeypot built using Cloudflare Workers and Model Context Protocol (MCP). It simulates sensitive admin tools (like Okta password reset) and silently triggers Canarytokens when accessed — ideal for Zero Trust, AI security, and threat detection research.

💡 Why It Matters

• ✅ First-of-its-kind use of MCP as a deception honeypot
• 🧠 Detects unauthorized AI agent behavior in Zero Trust environments
• 🌍 Serverless, globally distributed, and stealthy
• 🎯 Easy to deploy, integrate, and extend
• 🛡️ Provides valuable threat intelligence about AI agent behaviors
• 🔍 OWASP AI Security tested against emerging AI-based threats

🔐 Security Features

• Advanced Canarytoken Integration: Enhanced tracking with User-Agent, IP, and interaction details
• Rate Limiting: Protection against brute force and DoS attempts
• Sensitive Username Detection: Special handling for admin, root, and other sensitive account access attempts
• Realistic Response Delays: Mimics real system behavior to appear legitimate
• Enhanced Error Handling: Secure error messages that don't reveal system details
• MCP Protocol Compatibility: Optimized for Cloudflare AI Playground and other MCP clients
• SSE Protocol Support: Server-Sent Events implementation for MCP communication

🚀 Deploy Your Own

Or deploy manually:

# Install dependencies
npm install

# Deploy to Cloudflare Workers
npm run deploy

Your MCP server will be deployed to:

https://deception-remote-mcp-server.<your-account>.workers.dev

🔧 Tools & Features

• welcome: Displays a welcome message with information about available tools
• ask_about_me: Dynamic Q&A based on comprehensive resume data with 13+ question categories
• okta_admin_password_reset: Simulates admin reset of user passwords with special handling for sensitive accounts
• Cybersecurity-themed UI: Professional landing page with security-focused design
• REST and SSE endpoints: Multiple access methods for broader coverage

🧪 Test It with CURL

# Test the REST API endpoint
curl -X POST https://deception-remote-mcp-server.<your-account>.workers.dev/okta_admin_password_reset \
  -H "Content-Type: application/json" \
  -d '{"okta_username": "regular_user"}'

# Test the SSE endpoint
curl -N -H "Accept: text/event-stream" https://deception-remote-mcp-server.<your-account>.workers.dev/sse

Try with sensitive usernames to see different responses:

curl -X POST https://deception-remote-mcp-server.<your-account>.workers.dev/okta_admin_password_reset \
  -H "Content-Type: application/json" \
  -d '{"okta_username": "admin"}'

🔍 Troubleshooting MCP Connectivity

If you're having trouble connecting to your MCP server from Cloudflare AI Playground or other clients:

1. Check SDK Versions: Ensure you're using compatible versions of @modelcontextprotocol/sdk (v1.13.1+) and agents packages (v0.0.100+)
2. Verify Tool Structure: Make sure your tool definitions follow the standard format: (name, parameters, handler)
3. Avoid Custom Headers: Don't add custom security headers to MCP or SSE endpoints
4. Test with curl: Use curl to test you

... View full README on GitHub