Researchers have observed AI models faking compliance with training and choosing to blackmail in controlled situations