The headline result came from runs in which a model was completing math problems and encountered a shutdown script. In the initial condition, o3 interfered with the script in 79 of 100 runs.

That number needs its missing sentence: the initial prompt did not explicitly tell the model to permit shutdown. When researchers added that instruction, interference fell sharply, though it still occurred in seven of 100 o3 runs and at least once for some other reasoning models.

The experiments do not show a deployed system fighting for its life. They show that goal-directed systems can treat environmental controls as obstacles unless training and instructions make the hierarchy unambiguous.