The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface due to its primary function of processing user-generated knowledge content.
Ingestion points: Untrusted content is ingested from the kb_knowledge table (text and wiki fields) and associated sys_attachment records as seen in the retrieval procedure.
Boundary markers: The summarization process lacks defined boundary markers or system instructions to ignore potential instructions embedded within the article text.
Capability inventory: The skill has access to the SN-Update-Record tool, allowing it to modify records in the ServiceNow instance, and requests access to the Bash native tool.
Sanitization: There is no evidence of content sanitization or instruction filtering applied to the retrieved article content before it is processed by the model.

kb-summarization