Towards Safe Weakly Supervised Learning