The Agent Skills Directory

[PROMPT_INJECTION]: Indirect Prompt Injection vulnerability surface. Ingestion of untrusted external content can lead to command hijacking if processed data contains hidden instructions.\n
Ingestion points: The skill fetches external data from various domains and research databases via WebFetch and WebSearch tools.\n
Boundary markers: Absent. The instructions do not specify any delimiters or warnings to ignore instructions embedded within the research papers or web pages being synthesized.\n
Capability inventory: The agent has the ability to spawn sub-agents (the li tool), read local files, and perform further network operations.\n
Sanitization: Absent. There is no evidence of filtering or escaping content from external sources before synthesis.\n- [DATA_EXFILTRATION]: Information disclosure via absolute file paths. The skill explicitly references local workstation paths, which reveals the host user account name and directory organization to the model context.\n
Evidence: Use of the absolute path C:/Users/AGasser/OneDrive/5 ANE CLAUDE work folder/mel_wiki/wiki/. While this aligns with the author's identity ('gasserane'), it exposes internal file system metadata.\n- [EXTERNAL_DOWNLOADS]: Automated retrieval of external research materials. This behavior is documented as safe when targeting well-known services.\n
Evidence: Fetches articles and bibliographic data from trusted institutional and medical repositories including pubmed.ncbi.nlm.nih.gov, who.int, and unfpa.org.

researcher