I suppose we're not supposed to ask where the information in these proprietary machine learning models came from.
https://cloud.google.com/blog/topics/threat-intelligence/distillation-experimentation-integration-ai-adversarial-use