The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from numerous external, third-party sources such as global news monitoring (GDELT), web archives (Wayback Machine), and Wikipedia summaries. This text is processed and stored, potentially containing adversarial instructions that could influence the agent's behavior during analysis.
Ingestion points: Multiple fetcher scripts in scripts/ (e.g., fetch_gdelt.py, fetch_wikipedia.py, fetch_courtlistener.py) retrieve JSON/CSV data from external APIs.
Capability inventory: Terminal tool usage to execute scripts with file system and network access for evidence aggregation.
Boundary markers: No explicit boundary delimiters are used to isolate gathered evidence when the agent interprets findings.json or intermediate CSVs.
Sanitization: The skill performs basic name normalization in scripts/_normalize.py but lacks sanitization for large narrative text fields gathered from the web.
[SAFE]: Fetches data from official government domains (SEC, USAspending, Treasury, Senate, NYC Open Data) and well-known public interest services (ICIJ, OpenCorporates, CourtListener, Internet Archive, Wikimedia, GDELT). All network activity and script behavior align with the stated primary purpose of the skill.

osint-investigation