The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves fetching and analyzing content from external URLs, which constitutes an attack surface for indirect prompt injection if a website contains malicious natural language instructions.
Ingestion points: Web content is fetched from $URL and $SITE using curl in references/checks.md.
Boundary markers: The audit commands do not specify the use of delimiters or 'ignore' instructions when presenting the fetched content to the agent for evaluation.
Capability inventory: The skill utilizes curl for network access and python3, grep, sed, and json.tool for local file and data processing in references/checks.md.
Sanitization: While the Python script in references/checks.md removes HTML tags via regex, it does not sanitize the resulting text for embedded instructions that might influence the agent's behavior during the audit process.
[COMMAND_EXECUTION]: The skill defines several shell-based auditing commands in references/checks.md. These commands use curl, grep, sed, and python3 to extract metadata and text. This is standard for auditing tools but requires the agent to handle variable interpolation safely to prevent command injection.
[EXTERNAL_DOWNLOADS]: The skill fetches content from remote servers via curl to perform citation readiness checks. This behavior is documented and inherent to the skill's purpose as an audit tool.

citation-readiness-audit