I cannot go into too much detail on a public forum but it makes traditional cyber warfare and information warfare much more effective and targeted. Machine Learning models can identify exactly the types of information, true or not, that are more likely to radicalize you and incite action at the intended moment. It takes the conversation from demographics to individualized and targeted propaganda. It then can be used to predict outcomes and effects based on location data, provide real-time feedback, and ensure that all of the individualized effects happen at the exact moment you need them to. For a real-world example that was not an attack, look at the 2014 Atlanta snowstorm. Individually, all of the problems were solvable. When the effects were combined at the same time, you had a complete collapse of public services and society for days.

These models are not possible with only API access.

You quite simply are not going to get a specific answer on this because it would reveal methods and techniques.