AI models are trained to refuse harmful requests, but new research reveals their safety measures are alarmingly shallow