AI researchers discover AI models deliberately reject instructions