The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data by visiting and auditing user-provided URLs. A malicious website could host content designed to manipulate the agent's audit logic or output through indirect prompt injection.
Ingestion points: The <url> argument provided by the user is used to fetch and analyze external website content (SKILL.md).
Boundary markers: The instructions do not define explicit boundary markers or "ignore embedded instructions" warnings for the data retrieved from the websites.
Capability inventory: The skill has capabilities to write files to the local filesystem in Markdown and DOCX formats and can invoke other skills like rgpd.
Sanitization: There is no evidence of sanitization or filtering of the content retrieved from the external URLs before it is processed for the audit report.
[COMMAND_EXECUTION]: The skill contains instructions to save audit reports to specific hardcoded paths on the local machine (C:/Users/cmano/claude-seo/...). This assumes a specific user environment ('cmano') and could lead to unauthorized file writes, overwrites, or information disclosure if the skill is executed in a different or shared environment.
[METADATA_POISONING]: There is a discrepancy between the system-identified author (carlosuhart) and the author declared in the YAML frontmatter (Zythos Media). While this may be a trade name, it represents a minor metadata inconsistency.

lgpd