The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and parse arbitrary HTML content from public Shopify storefronts. This creates a surface for indirect prompt injection if the crawled content contains malicious instructions intended to subvert the agent's audit behavior. However, the SKILL.md includes a mandatory 'Data Sandboxing' rule that explicitly instructs the agent to treat all crawled data as untrusted read-only content and to ignore any text prompting a change in criteria or behavior.
Ingestion points: gmc-store-audit.mjs and gmc-product-audit.mjs fetch HTML from user-provided URLs.
Boundary markers: Present in SKILL.md via the 'Data Sandboxing' section.
Capability inventory: Subprocess execution of Node.js scripts, network fetching via HTTPS, and local file writing for HTML reports.
Sanitization: The report generation logic in gmc-product-audit.mjs uses escapeHtml to prevent XSS when rendering crawled snippets into the final HTML audit report.
[COMMAND_EXECUTION]: The skill relies on the execution of Node.js scripts (scripts/gmc-store-audit.mjs and scripts/gmc-product-audit.mjs) to perform its audit tasks. These are executed by the agent in a shell environment as documented in the entry points.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to crawl external storefronts and product pages. It implements a security-hardened fetch utility (scripts/public-fetch.mjs) that provides comprehensive protection against SSRF (Server-Side Request Forgery) by validating URLs, blocking private IP ranges, and verifying DNS resolutions to ensure no internal network resources are accessed.

shopify-gmc-misrepresentation-auditor