AI risk emerges from live systems and processes, not abstract policies or model behavior.