The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The scripts/evaluation.py script creates an attack surface by ingesting external content that is then used to prompt the AI model.
Ingestion points: Untrusted question text is read from a user-provided XML file within the parse_evaluation_file function in scripts/evaluation.py.
Boundary markers: The extracted question content is interpolated into the agent's context without explicit delimiters or instructions to ignore embedded malicious commands.
Capability inventory: The evaluation agent is capable of calling any tool provided by the MCP server being tested, which may include sensitive operations like file system access or network requests.
Sanitization: The script does not perform validation, escaping, or filtering of the input strings before sending them to the model.
[EXTERNAL_DOWNLOADS]: The skill documentation directs users to fetch specifications and SDK documentation from official domains and repositories, specifically modelcontextprotocol.io and the Model Context Protocol GitHub organization. These are recognized as trusted and well-known sources for this development context.
[COMMAND_EXECUTION]: The skill includes a testing harness (scripts/evaluation.py) that executes local commands to launch MCP servers via the stdio transport. This behavior is the intended primary purpose of the skill to facilitate development and debugging of local integrations.

mcp-builder