Machine-learning models vulnerable to undetectable backdoors: new claim – The Register