Unveiling Covert Threats: Towards Physically Safe And Transparent Ai Systems