The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to facilitate the creation of scripts that ingest data from potentially untrusted external sources, which creates a surface for indirect prompt injection.
Ingestion points: The skill explicitly supports and provides templates for reading local files (CSV, Excel, JSON via pandas, json, openpyxl) and scraping web content (via requests, BeautifulSoup, Selenium, Playwright, httpx).
Capability inventory: The scripts generated by this skill are intended to have capabilities including file system writes (to_csv, pathlib), network operations (requests.get), and system command execution (referenced via the subprocess keyword).
Boundary markers: The instructions do not mandate the use of delimiters or 'ignore embedded instructions' warnings when the agent-generated scripts process or output data derived from external sources.
Sanitization: Although the skill encourages 'safety awareness' for file paths, it lacks specific instructions for sanitizing the content of scraped web data or parsed files, which could contain malicious instructions designed to influence the agent's logic if that data is subsequently processed.

python-helper