The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content from product pages and merchant feeds, creating a vector for indirect injection. \n
Ingestion points: Data is read from rendered_dom, raw_html, and Merchant Center feed content within the audited environment. \n
Boundary markers: There are no defined delimiters or instructions to treat external data as purely informational, increasing the risk that the agent may interpret data as commands. \n
Capability inventory: The skill allows the use of Bash, WebFetch, and Read tools, providing a broad capability tier for potential post-injection exploitation. \n
Sanitization: No sanitization is specified for content extracted via evidence.observed before being utilized in findings or reproduction commands. \n
[COMMAND_EXECUTION]: The verification instructions include a Bash command template: node ${CLAUDE_SKILL_DIR}/../../scripts/validate-jsonld.mjs --url <u>. The use of the <u> placeholder for a URL audited by the agent represents a command injection risk if the URL is not sanitized. Additionally, the instruction for the agent to include 'runnable commands' in its findings based on page content (verification.reproduce) directly encourages the creation and execution of strings derived from untrusted external sources.

seo-ecommerce