The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes text patterns like "Ignore previous instructions" within its educational content. These are clearly marked as examples of malicious input (Mistake 3) to help developers recognize and defend against prompt injection attacks.
[INDIRECT_PROMPT_INJECTION]: The skill identifies the attack surface for indirect prompt injection in WordPress applications.
Ingestion points: Identifies untrusted sources including comments, reviews, post content, order notes, and uploaded files.
Boundary markers: Advises that delimiters help models distinguish content but are not a primary security control.
Capability inventory: Reviews risks associated with tools for trashing posts, publishing content, and sending emails.
Sanitization: Recommends consistent use of esc_html(), wp_kses_post(), and sanitize_textarea_field() for all data originating from models.
[EXTERNAL_DOWNLOADS]: Contains references to official WordPress repositories on GitHub, including the MCP Adapter. These references are provided as legitimate supporting documentation for building AI features.

ai-llm-integration-security