Foveate, Attribute, and Rationalize: Towards Physically Safe and Trustworthy AI