The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted data through file upload patterns in controllers and views (Step 4, Step 7 in SKILL.md).
Ingestion points: External files are ingested via params in controller actions such as user.avatar.attach(params[:avatar]) and through multi-file form submissions.
Boundary markers: The skill emphasizes the use of signed_id (Step 7) to securely reference attachments, providing a boundary against unauthorized blob access.
Capability inventory: The skill environment includes capabilities for shell execution (Bash for rails and database tasks), file writing, and editing.
Sanitization: The skill provides a dedicated 'Content Type Validation' section in reference.md, detailing how to implement manual validation for content_type and byte_size or use the active_storage_validations gem to mitigate risks from malicious file uploads.

active-storage