The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches and processes HTML content from external URLs to perform SEO analysis, which inherently creates a surface for indirect prompt injection if the fetched content contains instructions designed to influence the agent.
Ingestion points: External HTML content is fetched from user-provided URLs via the fetch_page.py script.
Boundary markers: The skill structures data into JSON snapshots and computes SHA-256 hashes of the HTML and schema content, which helps in identifying changes but does not sanitize the input text for LLM instructions.
Capability inventory: The skill uses network fetching (with SSRF protection), writes to a local SQLite database, and generates HTML reports.
Sanitization: The skill performs URL validation through google_auth.validate_url() to prevent SSRF but does not explicitly detail the sanitization of parsed HTML elements before they are presented to the agent.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill stores snapshot data in a local cache directory, which is a standard practice for maintaining application state and does not constitute a security risk.
Evidence: Data is stored in ~/.cache/claude-seo/drift/baselines.db using parameterized queries to prevent SQL injection.

seo-drift