There’s none, you always need to validate its output yourself. The models lack the capability to understand if something it produce is wrong or malicious.