The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the open web using search and extraction tools.
Ingestion points: The research workflow in SKILL.md utilizes web_search and web_extract to gather external information on entities, creating a surface for indirect prompt injection.
Boundary markers: To mitigate risks, the instructions mandate date grounding, quality filtering of results, and explicit source attribution for every claim.
Capability inventory: The skill possesses capabilities to fetch network data (public sanctions lists) and write to a local cache directory (.sanctions_cache) via scripts/entity_research.py.
Sanitization: The skill incorporates source weighting, name normalization, and a structured dossier format that distinguishes between allegations and outcomes, ensuring the final report is reviewed by a human.
[EXTERNAL_DOWNLOADS]: The helper script scripts/entity_research.py fetches public sanctions lists from official government and international organizations.
Evidence: The script fetches data from treasury.gov (US Department of the Treasury), un.org (United Nations), and ofsistorage.blob.core.windows.net (UK Office of Financial Sanctions Implementation). These are well-known and official sources for the required compliance data.

entity-research