The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements procedural orchestration through the FastMCP instructions parameter, providing the model with strict execution orders and hard constraints. This approach minimizes the risk of indirect prompt injection by defining clear boundaries for how the model interacts with external resources. Every finding for this category includes the following evidence: Ingestion points are the MCP tool arguments and HTTP headers; Boundary markers are established in the 'Hard constraints' section of the instruction template; Capabilities include the Read, Write, and Edit tools declared in the skill metadata; Sanitization is guided by explicit field conventions and error handling logic.
[DATA_EXFILTRATION]: Authentication examples in examples/auth.py demonstrate the secure handling of headers and explicitly warn against logging sensitive request data. The use of OpenTelemetry attributes is correctly scoped to prevent telemetry data from leaking between tenants in multi-tenant environments.
[CREDENTIALS_UNSAFE]: The skill adheres to security best practices by avoiding hardcoded secrets and demonstrating how to dynamically retrieve credentials from request headers rather than storing them in code.
[SAFE]: The inclusion of a drift test in the example suite promotes reliability and security by ensuring that model instructions are always synchronized with the underlying code, reducing the potential for logic-based vulnerabilities resulting from outdated model guidance.

matematic-mcp-fastmcp-instructions-pl