The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted data from external sources, creating a vulnerability where embedded instructions could influence agent behavior.
Ingestion points: Tools like fetch, fetch_batch, read_doc, research, and download ingest content from arbitrary URLs, including research papers, web pages, and datasets.
Boundary markers: There are no instructions for the agent to wrap external content in delimiters or use "ignore embedded instructions" warnings when processing the data.
Capability inventory: The skill includes a download tool that writes files to a local filesystem ($HOME/.cache/search-mcp/downloads), providing a persistent side-effect for potential injection attacks.
Sanitization: No explicit sanitization or validation of the retrieved content is mentioned before it is processed or stored.
[EXTERNAL_DOWNLOADS]: The skill utilizes external dependencies and provides tools for downloading remote content.
Evidence: The installation instructions recommend using uvx free-search-mcp, which downloads and runs a Python package from a remote registry.
Evidence: The download(url) tool facilitates the acquisition of arbitrary remote files into a local staging directory.

mathodology-evidence-search