The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes inbound user messages and conversation context to trigger bot responses, routing, and tagging. This creates an attack surface where a malicious user could craft messages designed to influence the agent's behavior, such as bypassing intent classification or manipulating priority levels.
Ingestion points: Untrusted data enters the agent context through the 'User Message (Inbound)' in the Conversation Flow and 'message_contains' checks in the Routing Rules (SKILL.md).
Boundary markers: There are no explicit boundary markers or instructions defined in the skill to delimit untrusted customer input from the agent's internal operational instructions.
Capability inventory: The skill utilizes several high-impact tools including intercom_message, intercom_user, intercom_conversation, and intercom_articles. It also describes workflows for creating Jira tickets and logging activities in Salesforce.
Sanitization: No evidence of input sanitization, filtering, or validation is present to mitigate instructions embedded within customer messages.

Intercom Automation