The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (customer support requests, emails, and ticket subjects) to drive automated workflows. An attacker could embed instructions within a support ticket to manipulate the agent's logic.
Ingestion points: SKILL.md defines ingestion of 'Support requests', 'Ticket data', and 'Customer information' via tools like zendesk_search and zendesk_update_ticket.
Boundary markers: None identified in the provided templates. The templates interpolate ticket data directly (e.g., {{ticket.requester.name}}, {{ticket.subject}}) without explicit delimiters or warnings to the LLM to ignore instructions within the data.
Capability inventory: The skill possesses significant capabilities including ticket creation, updating, routing, and integration with external platforms like Slack and JIRA. Specifically, it can set_group, set_priority, notify external channels, and create_jira issues.
Sanitization: No evidence of sanitization or filtering of customer-provided text before it is processed by the AI for sentiment analysis, intent detection, or auto-response generation.

Zendesk Automation