The example URL nobody reserved
Documentation is full of stand-in web addresses. Most of them are example.com, which the Internet Assigned Numbers Authority holds permanently so that nobody can ever register it. Two of the most common alternatives are not held by anyone.
yoursite.com and your-domain.com appear in roughly 350,000 files on GitHub and are cited by 349 AI agent skills, according to research published by Manifold Security. Both are registerable, both have been registered, and both now send visitors to scams.
The original documentation did not change. The link in it did not change. What changed is who answers when something follows it.
What the domains serve
your-domain.com returns scareware dressed as macOS security software, including fake virus alerts offering a discounted antivirus renewal, and counterfeit news pages built to look like BBC articles promoting investment schemes.
yoursite.com routes macOS visitors to fabricated news articles — the researchers single out pages made to resemble the German broadcaster ZDF — advertising investment fraud. Visitors on other systems get an ordinary parked page and see nothing wrong.
That selectivity is the reason this went unnoticed. So is the mechanism: the redirect to the scam fires only after the page’s JavaScript runs. A static check that fetches the URL and reads the response sees a parking page. Only a real browser session sees where it goes.
Why agents are the exposed party
A human reading a tutorial glances at a placeholder link and moves on. An agent executing a skill does not. It follows the citations in its instructions because following them is what it was built to do, and it has no instinct that a URL called yoursite.com in a code sample was never meant to resolve to anything.
Manifold’s point is not that these particular domains are catastrophic. It is that a class of links which every security model treats as inert — a literal, a placeholder, a thing in a comment — becomes live the moment somebody buys the name, and nothing in the repository that cites it has to be touched for that to happen.
A third domain, third-party.com, has already been through this. It is referenced across more than 1,700 repositories and has served ClickFix malware.
What to do about it
The fix is dull and complete: cite only the addresses IANA reserves. example.com, example.org and example.net cannot be registered by anyone, which is the entire reason they exist.
For anyone shipping agent skills, the practical step is an audit of every domain that appears in a skill’s documentation, not just the ones in its code. And for anyone running agents, the broader lesson is that the trust boundary is wider than the prompt. Everything an agent reads is something it may act on, including a link written years ago by someone who assumed nobody would ever click it.
