Researchers at LayerX warn that custom fonts can fool AI web assistants into thinking phishing pages are benign, while the human user sees something completely different.