Note that, in addition to normalizing domain components, the process also maps U+3002 IDEOGRAPHIC FULL STOP (γ) to U+002E FULL STOP (.) for use as a separator, a nice touch for domain names encoded in blackboard bold.
Also, the author's one-liner, updated for Python 3:
Not sure why so many things use the compatibility form of normalisation.
Some older RFCs even recommended using it for passwords, which needlessly reduces entropy (yes we need to normalise passwords because of combining characters, but we donβt need to collapse symbols that a user can choose from, like ΕΏ and s).
In my opinion, the compatibility form makes mostly sense for searching text, and even then itβs sometimes annoying (I want to search exactly for ss and not Γ, browser). If you know of other good use cases let me know.
It also OK to disallow registering two identifiers with the same compatibility normalisation form (edit: this does not prevent all homoglyph attacks BTW, that is another can of worms), but I donβt see the need to normalise the identifier when itβs used.
Tangentially related - 20+ years ago I managed to register a Ο domain under one of the primary TLDs.
As in, the registration form was accepted and the fee was charged.
Then, few days after they emailed to say that "due to technical difficulties" the registration can be registered and they have no choice but to unregister it.
My brother had a friend who tried to register fuck.com in the early days of the Internet, around 1994, and the registrar was confused and wanted to know why he wanted to register it. This was when the Internet was mostly universities and a .com cost $100 to register.
Maybe. That sounds like the story he told me. But now I'm wondering if I'm misremembering the friend part. Maybe he read the story when he was posted in 1997 and told me about it.
1994 they were definitely still free[1]. I had so many good ones, but when they switched to the paid model[0] I was a broke college student and let them all go.
[0] IIRC, it was Unilever who caused it -- they mailed in something like 19,000 domain registration applications, one for each of their current trademarks.
[1] OK, I looked it up, September 1995 was when the charging started.
Unilever later bought the domain substance.com from my friend. Remarkably, they apparently later let it lapse and someone else registered it. I'd assumed they were going to have a database of chemical substances that were used as ingredients, or something.
I managed to register a one letter domain below a CC-tld, despite that country having a rule against it. Their signup form had two stages and the check was only on the first page... so all it took was changing a hidden form field.
I had it for a while until they started charging a lot for it (it was Turkmenistan).
I managed to grab a lot of really good[0] emoji domains by bouncing between registrars until you find one that has some poor (good?) coding that lets them slip through.
Then you get stuck a year later when they've upgraded their system and now nothing you do will let you renew it or transfer it...!
> My favorite unicode character is zero width space. One of its funniest properties is that it's not a whitespace.
My editor is configured to never show any character (except the one and proper newline) at less than the full-width of a regular character (in the monospace font I'm using ofc).
And if it's not a visible ASCII char, not only it's always shown at least at full-width, it's also shown as bold magenta on a black background. As the non visibe zero-width is empty, it shows as full-width on a black background.
Doesn't work for me in Chrome either; it transforms the URL from the top comment into https://sp.m. I believe this is the result of browser-level defenses against IDN homograph attacks, some of which postdate this article.
Oh, it can. In fact there's an entire official tome of documentation describing how to carefully do Unicode-in-URLs to reduce the risk[0]. Browsers also have some mitigations to avoid the most common issues like homoglyph attacks where they show the domain in punycode if it doesn't match some local heuristic (like if you're in en-US, it's unlikely you're actually trying to visit a site with Cyrillic letter "a" instead of ASCII "a").
It's implemented because we have thousands of writing systems and we decided, for better or worse, we want to reflect those in URLs. Solving the subsequent security issues is likely easier than getting everyone to agree to read and write the same language.
As to me I'm running my own DNS resolver (I use unbound) and everything goes through port 53 UDP (and AIUI technically 53 TCP should be allowed, but nothing will happen if you just block it), unencrypted. What then unbound does on the Internet, that's up to you: you can encrypt those requests...
The nice thing with everything being in the clear, is that you can null route / NX DOMAIN every single domain that uses Unicode (punnycode) characters and you can even do that at the firewall level, by inspecting UDP packets.
Of course you gotta be sure to really block DoH / DoT internally but it's all doable.
I set that up years ago and I've been running like that since then without any issue.
Just as entire countries's IP blocks do not exist in my world, domains using Unicode characters do not exist. And I really don't care about your feelings about it: I'm just explaining that technically it can be done to block them all and that the Internet keeps working just fine.
Unicode really is just too complex to ever be secure. And those who don't understand that are part of the problem.
Looks like HN broke the URL after submitting. Let's try again:
Here is the backup link:
https://shkspr.mobi/blog/2018/11/domain-hacks-with-unusual-u...
That shows up as βα΅.β, so a bunch of characters are missing.
HN seems to encode it as punycode when it shouldnβt; it skipped normalization.
Python 3 encodes the domain to xn--hkr-zdyy2ssn00g.mobi.
Not that that works.
Note that, in addition to normalizing domain components, the process also maps U+3002 IDEOGRAPHIC FULL STOP (γ) to U+002E FULL STOP (.) for use as a separator, a nice touch for domain names encoded in blackboard bold.
Also, the author's one-liner, updated for Python 3:
Not sure why so many things use the compatibility form of normalisation.
Some older RFCs even recommended using it for passwords, which needlessly reduces entropy (yes we need to normalise passwords because of combining characters, but we donβt need to collapse symbols that a user can choose from, like ΕΏ and s).
In my opinion, the compatibility form makes mostly sense for searching text, and even then itβs sometimes annoying (I want to search exactly for ss and not Γ, browser). If you know of other good use cases let me know.
It also OK to disallow registering two identifiers with the same compatibility normalisation form (edit: this does not prevent all homoglyph attacks BTW, that is another can of worms), but I donβt see the need to normalise the identifier when itβs used.
Tangentially related - 20+ years ago I managed to register a Ο domain under one of the primary TLDs.
As in, the registration form was accepted and the fee was charged.
Then, few days after they emailed to say that "due to technical difficulties" the registration can be registered and they have no choice but to unregister it.
I was so close, damn amateurs.
My brother had a friend who tried to register fuck.com in the early days of the Internet, around 1994, and the registrar was confused and wanted to know why he wanted to register it. This was when the Internet was mostly universities and a .com cost $100 to register.
What was his response? Did they refuse to allow it?
Was that friend Justin Hall of Wired?
There was a story about Jon Postel (pre-ICANN) and fuck.com, this might be it:
https://www.links.net/webpub/fuck.com.html
Maybe. That sounds like the story he told me. But now I'm wondering if I'm misremembering the friend part. Maybe he read the story when he was posted in 1997 and told me about it.
1994 they were definitely still free[1]. I had so many good ones, but when they switched to the paid model[0] I was a broke college student and let them all go.
[0] IIRC, it was Unilever who caused it -- they mailed in something like 19,000 domain registration applications, one for each of their current trademarks.
[1] OK, I looked it up, September 1995 was when the charging started.
Unilever later bought the domain substance.com from my friend. Remarkably, they apparently later let it lapse and someone else registered it. I'd assumed they were going to have a database of chemical substances that were used as ingredients, or something.
I managed to register a one letter domain below a CC-tld, despite that country having a rule against it. Their signup form had two stages and the check was only on the first page... so all it took was changing a hidden form field.
I had it for a while until they started charging a lot for it (it was Turkmenistan).
I managed to grab a lot of really good[0] emoji domains by bouncing between registrars until you find one that has some poor (good?) coding that lets them slip through.
Then you get stuck a year later when they've upgraded their system and now nothing you do will let you renew it or transfer it...!
[0] "good" is subjective in this context
My favorite unicode character is zero width space. One of its funniest properties is that it's not a whitespace.
Also, some programming languages treat it as a valid identifier.
> My favorite unicode character is zero width space. One of its funniest properties is that it's not a whitespace.
My editor is configured to never show any character (except the one and proper newline) at less than the full-width of a regular character (in the monospace font I'm using ofc).
And if it's not a visible ASCII char, not only it's always shown at least at full-width, it's also shown as bold magenta on a black background. As the non visibe zero-width is empty, it shows as full-width on a black background.
Sticks out like a sore thumb.
Ha! Yes I discovered that this is true in C as part of my IOCCC entry:
https://github.com/ioccc-src/winner/blob/master/2025/ncw3/pr...
Firefox doesn't even let me open the link, wow.
Same here on Firefox mobile, link is not clickable.
Same, can't open on Firefox
Doesn't work for me in Chrome either; it transforms the URL from the top comment into https://sp.m. I believe this is the result of browser-level defenses against IDN homograph attacks, some of which postdate this article.
On Firefox, hover-over of the link doesn't show the URL in the bottom left of the browser.
I click on the link on Android Firefox and nothing happens.
This has given me a lot of trouble with emails.
I have a domain that contains Nordic characters and it is a crapshoot if the email client will accept and send.
Meanwhile I have not encountered problems with http access.
Looks like that it could be exploited by some means for nefarious purpose. I wonder why it's even implemented in the first place.
Oh, it can. In fact there's an entire official tome of documentation describing how to carefully do Unicode-in-URLs to reduce the risk[0]. Browsers also have some mitigations to avoid the most common issues like homoglyph attacks where they show the domain in punycode if it doesn't match some local heuristic (like if you're in en-US, it's unlikely you're actually trying to visit a site with Cyrillic letter "a" instead of ASCII "a").
It's implemented because we have thousands of writing systems and we decided, for better or worse, we want to reflect those in URLs. Solving the subsequent security issues is likely easier than getting everyone to agree to read and write the same language.
[0]: https://www.unicode.org/reports/tr36/tr36-10.html
Here is the URL encoded in base64 so HN does not break it:
aHR0cHM6Ly/wn4WC8J2WjfCdkKTigpvhtZbwnZKTLuKTnPCdlaDwnZKD8J2Tsi/wn4aG8J+Gg/CfhbU=
There have been some additions since 2018 when this article was written; here is the current list, from what I can tell:
Interesting, thanks. How did you generate this list? And do you know if the lowercasing rules have ever changed between Unicode version?
Via <https://data.iana.org/TLD/tlds-alpha-by-domain.txt>, <https://www.unicode.org/Public/UNIDATA/UnicodeData.txt>, some Python and some regexps.
(I donβt know anything about Unicode case conversions.)
Oh, some good spots there. Thanks!
Is it too late to say most Unicode characters should not be permitted in domains?
My browser refuses to open this link. It couldnβt find the server.
Which of course Bruce Schneier predicted in 2000:
https://www.schneier.com/crypto-gram/archives/2000/0715.html...
"Unicode is just too complex to eve be secure"
As to me I'm running my own DNS resolver (I use unbound) and everything goes through port 53 UDP (and AIUI technically 53 TCP should be allowed, but nothing will happen if you just block it), unencrypted. What then unbound does on the Internet, that's up to you: you can encrypt those requests...
The nice thing with everything being in the clear, is that you can null route / NX DOMAIN every single domain that uses Unicode (punnycode) characters and you can even do that at the firewall level, by inspecting UDP packets.
Of course you gotta be sure to really block DoH / DoT internally but it's all doable.
I set that up years ago and I've been running like that since then without any issue.
Just as entire countries's IP blocks do not exist in my world, domains using Unicode characters do not exist. And I really don't care about your feelings about it: I'm just explaining that technically it can be done to block them all and that the Internet keeps working just fine.
Unicode really is just too complex to ever be secure. And those who don't understand that are part of the problem.
Does this also allow you to work with DNS in a fully private/anonymized way?
Now, please present a list of these to your grandmother, and ask her which ones are phishing, and which are legitimately from the government.
Now, present the same list to an I.T. professional with the same questions. Perhaps make sure they are not carrying any weapons before you try.
This is just one reason why visual inspection of URLs and "hover over this link" haven't been useful defenses for a long time.
The author of the article isn't advocating for using these domains for trusted services.