An ML paper on data stealing provides a construction for "gradient hacking" — LessWrong