The Agent Skills Directory

[PROMPT_INJECTION]: The skill facilitates the ingestion and processing of content from inbound emails, which are untrusted external sources. This creates a potential surface for indirect prompt injection where malicious instructions hidden in an email could influence the agent's behavior.
Ingestion points: The process methods in SupportMailbox (SKILL.md) and RepliesMailbox (reference.md) ingest content from mail.subject and mail.decoded.
Boundary markers: The provided examples do not use delimiters or instructions to ignore embedded commands within the email data.
Capability inventory: The implementation includes database writes (SupportTicket.create!), file attachments via Active Storage (record.files.attach), and background job enqueuing (SupportNotificationJob.perform_later).
Sanitization: While reference.md suggests using ActionController::Base.helpers.strip_tags for HTML processing, the primary implementation examples in SKILL.md perform direct interpolation of email content into the database without explicit sanitization.

action-mailbox