The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external websites via curl commands in references/checks.md. This represents a surface for indirect prompt injection if the fetched content (HTML, sitemaps, or JSON-LD) contains instructions designed to manipulate the agent's behavior.
Ingestion points: curl -s "$URL" and curl -s "$SITE/sitemap.xml" are used to fetch content from user-defined or externally-sourced URLs.
Boundary markers: The skill does not explicitly use delimiters or instruction-ignore warnings when processing the fetched content.
Capability inventory: The skill uses grep, sed, strings, and python3 to process data. It does not perform file-writes or network operations using the extracted data, limiting the impact of potential injections.
Sanitization: No specific sanitization or escaping of external content is performed before processing.
[DYNAMIC_EXECUTION]: The skill uses python3 -c in references/checks.md to execute a small, embedded script for parsing and filtering ImageObject schema.
Evidence: A multi-line Python script is passed to python3 -c to process standard input using json.loads() and recursion to find image keys. This is a low-risk pattern as the code is static and locally defined within the skill's reference files, rather than being generated from untrusted input.

image-audit