The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-supplied documents to generate meeting briefings, creating a surface for indirect prompt injection.\n- Ingestion points: Processes user-provided "source material" as mentioned in the workflow of SKILL.md.\n- Boundary markers: The instructions include directives to "Preserve authorization boundaries" and "Keep core work local," but do not define specific delimiters for untrusted content.\n- Capability inventory: The skill leverages the sandbase_run tool for web scraping, external searches, and model inference.\n- Sanitization: The skill lacks explicit sanitization or escaping instructions for the input data, though it mandates sending the "minimum necessary data."\n- [DATA_EXFILTRATION]: The skill transmits data to external SandBase API endpoints for enrichment and inference. This behavior is documented and includes safeguards such as requiring user confirmation for costs and limiting data transmission to the minimum required.\n- [EXTERNAL_DOWNLOADS]: The skill uses sandbase_run and sandbase_discover to fetch external data and perform web scraping through the author's (sandbaseai) infrastructure.

meeting-briefing