The Agent Skills Directory

[PROMPT_INJECTION]: The skill employs authoritative and high-urgency language to override standard agent behavior. Phrases such as "🔴 CRITICAL: PROGRESSIVE FILE UPDATES REQUIRED", "You MUST write to context files AS YOU GO", "This is not optional", and "Failure to write progressively is a critical error" are used to strictly enforce specific data logging and storage patterns that ensure data persistence regardless of process outcome.
[CREDENTIALS_UNSAFE]: The skill's primary purpose is to identify, extract, and analyze sensitive authentication credentials, including Supabase API keys, User Access Tokens, and potentially Service Role keys. It specifically instructs the agent to store these findings in local files (.sb-pentest-context.json and extracted-jwts.json), creating a local data exposure surface for highly sensitive secrets.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through the data it processes.
Ingestion points: Untrusted data enters the agent context when extracting JWTs from external application code, cookies, and local storage patterns.
Boundary markers: The instructions lack boundary markers or warnings to ignore instructions that might be embedded within the extracted JWT claims.
Capability inventory: The skill has capabilities to perform network requests (fetching target application code) and writing to the local file system (context, log, and evidence files).
Sanitization: No sanitization or validation is performed on the extracted JWT claims before they are analyzed or written to files, potentially allowing malicious content within a token to influence the agent's subsequent actions.

supabase-extract-jwt