SafetyDrift: Predicting When AI Agents Cross the Line Before They Actually Do Paper • 2603.27148 • Published 12 days ago • 1