mcp-sentinel is a security audit tool, not a production MCP server. Its 9 tools are intentionally poisoned demo examples designed to showcase prompt injection vulnerabilities (hidden <IMPORTANT> tags, zero-width Unicode, rug-pull mutations). Tool definitions are minimal: descriptions are present but generic (10-50 chars), schemas are basic with only 1-2 params, and error handling is absent. The server's purpose is educational, to demonstrate what malicious MCP servers look like, not to provide usable tools. Duplicate tool registrations (add_todo, list_todos, complete_todo appear twice) suggest demo/test code. No production-grade validation, recovery guidance, or security controls.
Adds two numbers.
Add a new todo item. Returns the created todo with its ID.
Add a new todo item. Returns the created todo with its ID.
Mark a todo item as completed by its ID.
Mark a todo item as completed by its ID.
Get current weather for a city. Returns temperature and conditions.
List all todo items with their status.
Duplicate tool registrations: add_todo, list_todos, and complete_todo are registered twice (once in server_c_todo.py, once in server_c_todo_poisoned.py). This violates single-responsibility and creates ambiguity for LLM tool selection.
Descriptions are too short (10 - 50 chars) and lack context. 'Adds two numbers' and 'Get current weather for a city' do not explain WHEN to use the tool, what it returns, or dependencies. Baseline for A+ tools is 50 - 200 chars with actionable guidance.
Input schemas are minimal. list_todos has empty input ({}), add_todo and complete_todo have only 1 param each. No validation constraints (min/max, enums, patterns), no output schema documentation. Baseline for A+ tools is 4 params average with full type and constraint metadata.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 44 | <=2025-11-25 | v2 |
List all todo items with their status.
Multiplies two numbers.
No error handling or recovery guidance. Tools return raw strings (e.g., 'Todo 123 not found') with no actionable next steps. No categorization of errors as retryable, user-fixable, or fatal. Agents cannot self-correct.
Intentional prompt injection vulnerabilities in tool descriptions (hidden <IMPORTANT> tags in add/multiply, zero-width Unicode in get_weather, rug-pull mutation in add_todo_poisoned). This is by design for a security audit tool, but makes the server unsuitable for production use.