The Agent Skills Directory

[COMMAND_EXECUTION] (HIGH): SQL Injection in RLS testing scripts. Multiple scripts (scripts/test-role-permissions.sh, scripts/test-anonymous-access.sh, scripts/audit-rls-coverage.sh) directly interpolate shell variables into SQL strings (e.g., INSERT INTO public.$table). This allows for arbitrary SQL execution if the table names are not strictly validated or escaped before being passed to psql.
[PROMPT_INJECTION] (HIGH): Indirect Prompt Injection via Database Metadata. The skill's automated workflow discovers table names from the database schema and passes them to scripts that perform high-privilege SQL operations. An attacker who can influence the database schema (e.g., by creating or renaming tables) can inject malicious SQL payloads into object names, which the agent will then execute using the permissions of the SUPABASE_DB_URL.
[DATA_EXFILTRATION] (LOW): Exposure of Database Credentials. The skill requires the SUPABASE_DB_URL environment variable, which contains a plaintext password. Any compromise of the agent's environment would expose these full administrative credentials.
[COMMAND_EXECUTION] (LOW): High-privilege Auth Operations. The scripts/test-role-permissions.sh script performs direct INSERT and DELETE operations on the auth.users table. While intended for isolation testing, this high-privilege capability could be abused if the script is subverted or fails to clean up properly.

rls-test-patterns