The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface in the handle_ai_product_question function. Untrusted input from customer questions is interpolated directly into an AI prompt template.
Ingestion points: User-provided customer questions via the $_POST['question'] variable in SKILL.md.
Boundary markers: Absent. The input is concatenated directly into the prompt string without delimiters (such as XML tags or triple backticks) or explicit instructions to disregard potentially malicious commands embedded in the question.
Capability inventory: The skill registers powerful system-altering tools via the WordPress 7.0 Abilities API, including woocommerce/update-inventory and woocommerce/process-order.
Sanitization: While standard PHP sanitize_text_field is applied for database safety, no LLM-specific sanitization or validation of the question's content is performed before prompt construction.
[PROMPT_INJECTION]: The ai_validate_order function in SKILL.md presents an indirect injection surface where billing data (email, phone, address) is processed by an AI model to detect fraud.
Ingestion points: Billing email, phone, and address from checkout form fields.
Boundary markers: None; the data is directly formatted into a natural language prompt string.
Capability inventory: The result of this AI validation can trigger checkout errors, providing a mechanism that could be abused to influence order validation outcomes.
Sanitization: Employs standard WordPress sanitization functions (sanitize_email, sanitize_text_field) which do not mitigate prompt injection risks.

wordpress-woocommerce-development