Machine learning models are immune to blandishments, threats, flattery and so - so why not set them against social engineers?