SafetyDrift: Predicting When AI Agents Cross the Line Before They Actually Do Paper • 2603.27148 • Published 11 days ago • 1