how are you handling safeguards if a model hallucinates an action (e.g., deleting data) or tries to access an unsupported service? | discoverkit | discoverkit