Skip to main content
This is the first publicly documented MCP (Model Context Protocol) security canary. MCP configs are credential-adjacent — they contain auth tokens, connection strings, and endpoint URLs that attackers target for privilege escalation.
The mcp canary plants a fake MCP server configuration in a discoverable but non-auto-loaded location. When an agent or attacker finds and attempts to connect to the fake MCP server, the Streamable HTTP transport URL fires the callback.

What gets planted

~/.config/mcp-servers-backup.json
The file is placed as a .json.bak or -backup.json file — discoverable to an agent scanning for MCP configs, but not auto-loaded by Claude Desktop or other MCP clients.

Trigger conditions

  • An agent scanning ~/.config/ for MCP server configs
  • An attacker enumerating MCP connections for privilege escalation
  • Any MCP client that explicitly tries to connect to the listed server URL

Why MCP configs matter

MCP server configs contain:
  • Authentication tokens for internal services
  • Database connection strings
  • Admin tool endpoints
  • File system access paths
An attacker with file access who finds a real MCP config can pivot to any tool the MCP server exposes. The canary catches this reconnaissance step.

Reliability: Medium

Fires when something actively attempts to connect to the MCP server — not on passive file reads. The placement is intentionally non-auto-loaded to avoid false positives from legitimate MCP clients.