The researchers are concerned that AI models are being deployed faster than we are learning how to make them safe.