The Agent Skills Directory

[PROMPT_INJECTION]: The skill incorporates defensive measures against prompt injection. In references/audit-guardrails.md, the agent is specifically instructed to treat all external data—such as web page content, screenshots, and ad account exports—as information to be analyzed rather than instructions to be followed. It explicitly commands the agent to ignore directives like 'ignore previous instructions' if they appear within processed data.
[EXTERNAL_DOWNLOADS]: The skill references documentation and tool registries hosted on GitHub. These links point to the vendor's repository (github.com/mbzadev) and are used for reference rather than automated execution. It also provides standard tracking implementation scripts for well-known services (Google Tag Manager, Meta Pixel, LinkedIn Insight Tag), which are safe and expected in an advertising-focused skill.
[INDIRECT_PROMPT_INJECTION]: The skill identifies a potential attack surface where instructions could be embedded in landing pages or ad creative being audited. It successfully mitigates this risk through the 'Hard stops' and 'Untrusted data' sections in its reference playbooks, which define clear boundaries for the agent when processing external content.
[DATA_EXPOSURE]: The skill uses safe placeholders (e.g., AW-XXXXXXXXX, YOUR_PIXEL_ID) for tracking IDs and instructs users to manage sensitive credentials through standard practices rather than hardcoding them in the skill's instructions.

ads