The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The evaluation harness in scripts/evaluation.py processes external XML files containing test questions which are then fed to the LLM. If these files are sourced from untrusted locations, they could contain instructions designed to influence the agent's behavior during the evaluation.
Ingestion points: Test questions are loaded from XML files via the parse_evaluation_file function in scripts/evaluation.py.
Boundary markers: The system prompt defines specific XML tags for structuring output (<summary>, <feedback>, <response>), but the input questions are not encapsulated within strict delimiters to isolate them from instructions.
Capability inventory: The evaluation loop can execute any tool exposed by the connected MCP server using the call_tool method in scripts/connections.py.
Sanitization: There is no explicit sanitization or instruction-filtering performed on the question text before it is processed by the Anthropic API.
[COMMAND_EXECUTION]: The scripts/connections.py and scripts/evaluation.py scripts allow for the execution of local MCP servers. The stdio_client is used to spawn server processes based on commands and arguments provided via command-line flags. This is the intended primary functionality for local integration and testing.

mcp-builder