The Agent Skills Directory

[COMMAND_EXECUTION] (HIGH): Multiple bash scripts are vulnerable to SQL/Command Injection due to unsafe variable interpolation.
Evidence (scripts/create-indexes.sh): The script executes psql "$DB_URL" -t -c "SELECT COUNT(*) FROM information_schema.tables WHERE table_name = '$TABLE_NAME';". If the $TABLE_NAME variable contains malicious SQL (e.g., docs'; DROP TABLE users; --), the database will execute the injected command.
Evidence (scripts/setup-hybrid-search.sh): Similar patterns exist where ${TABLE_NAME} is interpolated into complex SQL blocks (e.g., ALTER TABLE ${TABLE_NAME} ADD COLUMN fts tsvector).
Evidence (scripts/test-vector-search.sh): The validation script also interpolates the table name directly into psql calls for row counts and index checks.
[DATA_EXPOSURE] (LOW): Sensitive database credentials are handled via environment variables and command-line arguments.
Evidence: The scripts rely on SUPABASE_DB_URL which contains the database password in plain text. This is a common but sub-optimal pattern that can expose credentials in process lists (ps -aux) or shell history.
[INDIRECT_PROMPT_INJECTION] (MEDIUM): The skill creates a vulnerability surface by exposing setup scripts to the agent that lack input validation.
Evidence: If an agent is directed to use this skill by an untrusted source (e.g., a malicious README in a different repo) and provides a crafted table name, the agent will inadvertently compromise the user's database via the HIGH severity injection mentioned above.

pgvector-setup