SafetyDrift: Predicting When AI Agents Cross the Line Before They Actually Do Paper • 2603.27148 • Published Mar 28 • 1