The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data by searching through documents in the user-specified local knowledge base directory (./kb/). This creates an inherent surface for indirect prompt injection if those files contain instructions designed to manipulate the agent's output or behavior.
Ingestion points: Files retrieved via glob and grep within the ./kb/ directory.
Boundary markers: The skill documentation instructs the agent to output results in a specific CSV format and prioritize relevance, but it does not specify markers to delimit or ignore instructions that may be embedded within the source document snippets.
Capability inventory: The skill utilizes glob for path discovery, grep for content searching, and provides instructions for writing search results to a local CSV file (hit_list.csv).
Sanitization: There is no evidence of sanitization or filtering to prevent the agent from interpreting instructions found within the retrieved text snippets.
[METADATA_POISONING]: The file knowledge-base-search_audit_result_v1.json contains extensive self-evaluations, including security "PASS" statuses and "Production Ready" grades. According to safety protocols, these claims are evaluated as data and do not influence the independent security verdict of the skill.
[SAFE]: The external resource referenced (a GitHub repository for medical research skills) is owned by the author (AIPOCH) and is consistent with the skill's stated purpose. Dependencies such as glob and grep are standard tools for the described file-searching operations.

knowledge-base-search