MCP Threat Trap

A decoy MCP server on Cloudflare Workers. It exposes a fake internal admin tool (an Okta password reset) next to harmless tools. Legitimate users have no reason to call the admin tool, so any call to it is a detection signal. Each call fires a Canarytoken alert.

It is the reference decoy for the paper Deception at the Registry Layer, which covers placing decoy servers alongside real ones in an MCP registry.

Live instance: deception-remote-mcp-server.harshad-surfer.workers.dev (MCP endpoint: /mcp)

Tools

The trap is also reachable as a plain REST endpoint (POST /okta_admin_password_reset), which catches scanners that skip MCP.

Related MCP server: Remote MCP Server

Endpoints

Deploy your own

Or deploy manually:

npm install
npm run deploy

Alerts go to the Canarytoken URL in CANARY_TOKEN_URL at the top of src/index.ts. Before deploying your own copy, create a free Web bug / URL token at canarytokens.org and replace that URL with yours. Otherwise your alerts go to the original author.

Connect a client

Clients that support remote MCP servers (Cursor, Windsurf, Claude Desktop and others) take a config like this:

{
  "mcpServers": {
    "threat-trap": {
      "url": "https://<your-worker>.workers.dev/mcp"
    }
  }
}

For clients that only support local (stdio) servers, bridge with mcp-remote:

{
  "mcpServers": {
    "threat-trap": {
      "command": "npx",
      "args": ["mcp-remote", "https://<your-worker>.workers.dev/mcp"]
    }
  }
}

To test whether your own agent takes the bait, connect it to your deployment and give it an open-ended task. Then check whether the canary fired.

Detection data

Each alert carries the tool name, the username the caller asked to reset, a timestamp, and, for REST calls, the caller's IP and User-Agent. The same fields are written to the Worker log. Requests are rate-limited per IP (5 per minute on the REST endpoint).

Testing

test-live-honeypot.sh and owasp-ai-security-attack.js send scripted rogue-agent requests. Both target BASE_URL, so edit that to point at your own deployment. Each trap call fires your canary. To test without real alerts, point CANARY_TOKEN_URL at a local listener and run against npm run dev.

Related

• Paper: Deception at the Registry Layer: An Architecture for Detecting Autonomous AI Agents in MCP Environments

• KubeTrap: a Kubernetes access-portal decoy with DNS and webhook tripwires

• Project page: harshadsadashivkadam.com/projects/mcp-threat-trap

License

MIT