Testing and assessment of AI model abilities to identify dangerous or unintended behaviors before deployment.