“Existing AI safety frameworks, such as responsible scaling policies and preparedness frameworks, do not adequately account for the variable of test-time compute.”