Unicode has a block of 128 tag characters that mirror ASCII almost exactly: U+E0041 stands in for "A", U+E0061 for "a", U+E0020 for a space. They exist at the text-processing level, where machines read, and by design they are very nearly invisible to a human eye. The technique of hiding text in them is about two years old and it started on the other side of the fence, as a way of smuggling instructions to a language model that the person reading the same email would never see.
Microsoft Threat Intelligence says spammers have taken it over. On one day in early February the number of ASCII-smuggling signatures its Defender for Office filters raised went from roughly 21,000 a day to more than 1,300,000; four days later it was 2,500,000. The flood ran for months and then dropped away sharply in the middle of May. The words being hidden are the words of money: dollar amounts, "credit", "term", most of it from senders on finance-themed domains.
The mechanism is a splinter, not a disguise. Put an invisible character inside a word and the recipient still reads "funding", because nothing about the rendered line has changed; the filter reading the same line finds "fun" and "ding" instead. The real target, Microsoft says, is not the old business of matching strings but the machine-learning classifiers that have replaced it: a tokenizer that knew the shape of a lure word no longer recognises it, and emits rare or unknown pieces in its place. "The intent is inverted," the company wrote, "but the mechanism is similar, and a user's suspicions are not raised."
Two things undo it, and both amount to reading the message the way a person does. Normalisation strips the tag characters before anything else looks at the text, and "funding" comes back whole. Failing that, a filter can take a picture of the message and run OCR over it. Neither is exotic, and spammers have been using zero-width and non-breaking spaces for the same purpose for decades; the tags were probably picked up because some filters had not learned them yet. What the episode measures is not how clever the trick is but how far apart two readers of the same line can be.