Attackers can clone ML models or extract memorized data through API queries. This playbook details mechanisms, real-world research, and defenses such as output minimization, DP, and active monitoring.