The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation instructs the user to download a remote configuration file (CLAUDE.md) from a specific GitHub repository (yoanbernabeu/supabase-pentest-skills). Since files like CLAUDE.md often serve as behavioral instructions for AI agents, this mechanism allows the agent's logic to be updated from a remote, untrusted source.
[COMMAND_EXECUTION]: The skill frequently references shell commands and curl operations for environment setup, reproduction of findings (via curl-commands.sh), and package installation via npx. It encourages the agent to perform active network and file system operations based on user-supplied target URLs.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its core functionality.
Ingestion points: The skill fetches and analyzes external web content (HTML and JavaScript) from target URLs to detect Supabase configurations and extract API keys.
Boundary markers: There are no explicit markers or instructions to ignore embedded prompts within the data fetched from the target websites.
Capability inventory: The skill has the capability to write to the local file system (evidence logs, context files) and perform network requests.
Sanitization: No input sanitization or content filtering is defined for the data retrieved from external URLs.
[DATA_EXFILTRATION]: The skill is designed to search for and extract high-sensitivity information, including service role keys (which bypass security policies), JWTs, and database connection strings. While this is the intended purpose of a pentesting tool, the extraction and storage of these credentials in local files (.sb-pentest-context.json) represents a significant data handling risk.

supabase-pentest-skills