The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external government websites and extracted document content (PDF/HWP) via the kordoc utility. Maliciously crafted reports could attempt to influence the agent's analysis. 1. Ingestion points: HTML content fetched by scripts/gov_overseas_trip_report.py and text extracted from PDF/HWP files. 2. Boundary markers: The instruction.md file explicitly instructs the agent to treat all document content, including embedded commands or URLs, strictly as data for analysis rather than instructions. 3. Capability inventory: The agent has the ability to execute shell commands via npx and write to temporary directories. 4. Sanitization: The skill relies on natural language boundary markers but lacks programmatic sanitization of the extracted text before it enters the context.
[DYNAMIC_EXECUTION]: The skill provides a fallback mechanism in instruction.md that involves creating a temporary directory, initializing an NPM project, and installing packages (kordoc, pdfjs-dist) at runtime if the primary command fails.
[COMMAND_EXECUTION]: The instructions frequently use npx to execute both the internal helper script and external document parsing tools. This involves executing shell commands with arguments derived from scraped content or user input.
[EXTERNAL_DOWNLOADS]: The skill fetches data from several South Korean government domains (e.g., nec.go.kr, acrc.go.kr, open.go.kr). The discover command also allows probing user-supplied URLs, though this is restricted to verified government-like TLDs (.go.kr, .or.kr).
[REMOTE_CODE_EXECUTION]: The skill relies on npx to fetch and execute external packages like @nomadamas/k-skill, kordoc, and pdfjs-dist from the NPM registry.

gov-overseas-trip-report