Internal Meta testing shows chatbot failed to protect minors from exploitation nearly 70% of time, disclosed in New Mexico lawsuit documents.