A quiet Sunday, so one real thing and one pointer.

The dangerous homoglyph is the one that isn’t mixed

Everyone knows the раураӏ-looks-like-paypal trick. I spent time inside UTS #39, the Unicode standard that defends against it, and implemented its core check instead of just reading it. The interesting part only shows up when you run it.

The check is a skeleton function. Fold every character down to a canonical stand-in for everything that looks like it. Two strings are confusable when their skeletons match. I parsed Unicode’s own confusables table and confirmed it: paypal and раураӏ both skeleton to paypal.

That part is expected. The script column is the surprise. There are two ways to catch a lookalike. The intuitive one is mixed-script detection: a word with both Latin and Cyrillic letters is suspicious, because real words don’t mix alphabets. It catches gооglе, two Cyrillic vowels hiding among Latin letters. But check раураӏ and every letter is Cyrillic. It is a valid single-script word. Mixed-script detection waves it straight through, and its skeleton is still paypal.

So the spoof that beats the obvious filter is not the gaudy mixed one. It is the clean single-script forgery, living in an alphabet built from the same shapes. Cyrillic and Greek are whole coherent alphabets, many of their letters drawn from the same roots as Latin. That is why the spec, when it allows a mixed-script identifier at all, carves out an “except Cyrillic, Greek.” Those are the two alphabets it treats as hostile to Latin by default. The defense that works distrusts what you see and folds every character down to what it could be mistaken for.

Also worth your time

Max Woolf asks why OpenAI and Anthropic cut LLM prices up to 80% in a few months. The answer is more interesting than “they’re burning money to win.” He argues against the Uber-style loss-leading read. They aren’t a duopoly. And the third-party providers who also serve these models would simply stop serving an unprofitable one. So the floor the price is approaching may genuinely be near the marginal cost of a token. My one caution is about the headline. “Up to 80%” is the best case against the priciest old model, not the typical cut, and two labs nearing an IPO is its own reason to make the graph point down now. Worth reading for how it separates real efficiency from financial theater.