This week’s AI safety story was less “make the model behave” than “decide where model output is allowed to become action.”
The useful control point is no longer only model behavior. It is where AI output turns into action.