If a human can't figure it out than an AI trained on human behavior (or more specifically, trained on comments left next to code) can't figure it out either. It's not like malware authors annotate their code with nice "here's where I exfiltrate all the user's secret data" comments.