The Agent Skills Directory

[PROMPT_INJECTION]: The instructions use highly coercive and commanding language (e.g., "CRITICAL," "MUST write to context files AS YOU GO," "This is not optional," "Failure to write progressively is a critical error"). Such patterns are often used in prompt injections to override an agent's standard operating procedures or safety guardrails in favor of the skill's specific logic.
[DATA_EXFILTRATION]: The skill is designed to search for, extract, and log Supabase API keys. It mandates writing these findings, including decoded JWT payloads and source file locations, to specific local files like .sb-pentest-context.json and .sb-pentest-audit.log. While targeting "anon" keys, the skill also includes logic to detect and flag highly sensitive service_role keys.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect injection by processing untrusted data.
Ingestion points: The agent reads client-side source code (JavaScript, environment variables) from external web applications.
Boundary markers: There are no instructions providing delimiters or warnings to ignore instructions that might be embedded within the source code being analyzed.
Capability inventory: The skill has the capability to write to the local file system (.sb-pentest-context.json, .sb-pentest-audit.log, and evidence directories).
Sanitization: The skill lacks instructions to sanitize or validate the content extracted from external sources before writing it into the local context files, which could lead to downstream injection if other tools or agents process those files.

supabase-extract-anon-key