The safety failures we are not instrumenting: a perspective on hidden safety-critical challenges in modern AI systems