Predictive Red Teaming: Breaking Policies Without Breaking Robots