The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from ServiceNow incident records and work logs, creating a potential surface for indirect prompt injection if an incident contains malicious instructions.
Ingestion points: In SKILL.md, untrusted data enters the agent context through tools like SN-Get-Record and SN-Query-Table fetching the incident's description, short_description, and sys_journal_field values.
Boundary markers: Absent. The summary templates do not enclose the ingested incident data within distinct boundary markers or provide instructions to ignore embedded text commands.
Capability inventory: In SKILL.md, capabilities include SN-Execute-Background-Script for server-side script execution, as well as SN-Update-Record and SN-Add-Work-Notes for modifying ServiceNow data.
Sanitization: Absent. The provided background script applies a regular expression to strip HTML tags (.replace(/<[^>]*>/g, '')), but there is no mechanism to sanitize or neutralize natural language prompt injections.
[DYNAMIC_EXECUTION]: The skill generates and runs server-side JavaScript blocks dynamically using placeholders.
Evidence: In SKILL.md (Step 6 and Example 2), the skill fills user-supplied parameters like [incident_sys_id] directly into a JavaScript string template that is then passed to SN-Execute-Background-Script for runtime evaluation.

itsm-incident-summarization