Beyond static AI evaluations: advancing human interaction evaluations for LLM harms and risks