The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The evaluation harness ('scripts/evaluation.py') reads test questions from XML files and interpolates them directly into the agent's system prompt, creating a surface for potentially malicious instructions to influence the agent. 1. Ingestion points: The 'parse_evaluation_file' function in 'scripts/evaluation.py' extracts content from 'question' tags in user-supplied XML files. 2. Boundary markers: The question text is not wrapped in protective delimiters or accompanied by instructions to ignore embedded commands. 3. Capability inventory: The evaluation agent has access to all tools registered by the MCP server being tested, which may include file system and network operations. 4. Sanitization: The script does not validate, escape, or filter the question content before sending it to the model.
[COMMAND_EXECUTION]: The skill provides the ability to execute arbitrary local commands via the 'stdio' transport in both 'scripts/evaluation.py' and 'scripts/connections.py'. This capability is intended for starting MCP servers during testing but allows subprocess execution of user-supplied strings.
[EXTERNAL_DOWNLOADS]: The documentation and skill instructions guide the agent to fetch SDK information and protocol specifications from the official Model Context Protocol GitHub repositories (e.g., github.com/modelcontextprotocol). These documentation fetches from the official source are documented neutrally and are not considered dangerous.

mcp-builder