The Agent Skills Directory

[PROMPT_INJECTION]: The skill integrates data from untrusted local sources which provides a surface for indirect prompt injection. * Ingestion points: data/medications/medications.json, data/profile.json, and data/interactions/interaction-db.json mentioned in SKILL.md. * Boundary markers: Not present in instructions. * Capability inventory: File system Read and Write access via allowed tools. * Sanitization: None specified for the ingested content.
[DATA_EXFILTRATION]: The skill reads sensitive medical records from data/profile.json. While no network exfiltration tools are used, the combination of Read and Write permissions allows the agent to move or copy sensitive data within the file system. This access is consistent with the primary medical safety purpose of the skill.

interaction