The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external sources (ServiceNow records), which creates a potential surface for indirect prompt injection if the ingested data contains malicious instructions.
Ingestion points: Data is ingested from the incident, problem, problem_task, and known_error tables in ServiceNow via the SN-Query-Table MCP tool and REST API endpoints (SKILL.md).
Boundary markers: The instructions employ markdown headers and specific delimiter patterns (e.g., === INCIDENT PATTERN ANALYSIS ===, === 5-WHYS ROOT CAUSE ANALYSIS ===) to structure the data, which provides basic context separation but does not fully prevent instruction override.
Capability inventory: The skill has access to tools that can modify the system state, including creating records (SN-Create-Record), updating records (SN-Update-Record), adding work notes (SN-Add-Work-Notes), and closing problems (SN-Close-Problem). It also has access to Bash for network operations.
Sanitization: There are no explicit instructions or automated steps for the agent to sanitize or validate the content of ServiceNow fields (like short_description or resolution_notes) before they are processed or used to generate new record content.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to perform network operations.
Evidence: The procedure sections (Phase 1, 3, and 6) provide curl commands to interact with ServiceNow REST API endpoints (e.g., GET /api/now/table/incident). These commands are limited to legitimate API interactions within the intended ITSM workflow.

problem-analysis