The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill integrates with external webhook events that contain user-controllable data fields, creating a potential attack surface if processed by an agent without further validation.
Ingestion points: Webhook payloads are processed in WEBHOOKS.md and WORKFLOWS.md, specifically accessing event.object.metadata and event.object.custom_fields which can contain arbitrary strings provided by end-users during checkout.
Boundary markers: The implementation examples use standard JSON.parse(rawBody) for data ingestion without explicit instructions for the agent to use delimiters or ignore potential commands embedded within user-supplied fields.
Capability inventory: The skill patterns include capabilities to write to databases (db.user.update, db.subscription.upsert, db.purchase.create) and send automated emails (sendWelcomeEmail, sendDeliveryEmail, sendCancellationEmail).
Sanitization: Although the skill correctly mandates cryptographic signature verification using HMAC-SHA256 to ensure data authenticity, it does not explicitly guide the developer on sanitizing the natural language content of these fields before they are presented back to the LLM or stored in a way that affects agent state.

creem-api