The Agent Skills Directory

[COMMAND_EXECUTION]: Step 1f of the skill instructions executes a shell command using curl and awk to measure the Time to First Byte (TTFB). The command curl ... "{URL}" directly interpolates the user-provided URL argument. If the agent does not properly escape or sanitize the input, this can lead to shell command injection if the URL contains metacharacters like semicolons or ampersands.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection. It is designed to ingest and process large amounts of untrusted data from external sources including the target page text, JSON-LD schema, robots.txt, sitemap.xml, and llms.txt. These external sources could contain malicious instructions intended to manipulate the audit results or influence the agent's behavior.
Ingestion points: SKILL.md (Steps 1a, 1b, 1c, 1d, 1e) fetches data from external URLs.
Boundary markers: None. The skill does not use XML tags or specific delimiters to separate untrusted web content from its internal instructions.
Capability inventory: The skill has the capability to execute shell commands (bash in Step 1f) and write files to the local disk (Step 6).
Sanitization: None. There are no instructions to sanitize or escape the retrieved external content before it is processed by the LLM or written to the final HTML report.
[EXTERNAL_DOWNLOADS]: The skill initiates multiple automated network requests to various external files and domains based on the user's input URL to gather data for its analysis report.

landing-page-auditor