Skip to content
AI Security Wire

Published

- 4 min read

By

TypoPrompting: 186 AI Lookalike Domains, Some Serving ClickFix and Malware

img of TypoPrompting: 186 AI Lookalike Domains, Some Serving ClickFix and Malware

A wrong keystroke on an AI tool’s address can now cost more than a parked page. On November 12, browser-security vendor Island published research on what it calls TypoPrompting: 186 lookalike domains covering 27 AI brands. Island traced specific typos to a ClickFix lure, Store-signed malware and a search-hijacking extension; the rest of the pool is parked or was not traced end to end.

This is a single-vendor report, and everything below about the infrastructure and malware comes from Island unless stated otherwise. We opened Island’s report and Socket’s related research. The Hacker News coverage was used only as a pointer.

What Island found

The domain set covers Claude, ChatGPT, Gemini, NotebookLM, Copilot, Grok, DeepSeek, Mistral, Perplexity, Midjourney, ElevenLabs, Lovable, Cursor and others. ChatGPT draws the most lookalikes (25), followed by Claude (18) and DeepSeek (16). The slips are the usual ones: an adjacent key (clsude[.]ai), swapped letters (cluade[.]ai), a dropped letter (caude[.]ai).

Some of the names are old. Island says cloude[.]ai, cluade[.]ai and caude[.]ai share a registration date of July 13, 2023. Many sit in a monetised portfolio tied to the redirector a11[.]click through shared publisher keys and affiliate IDs.

The parked page does not show the same thing to everyone. It passes the browser to a traffic distribution system that picks a destination by country, operating system and browser. So one visitor sees a parking page and another gets an attack.

Three routes from a typo

ClickFix. In a scan Island traced, cloude[.]ai ended at a fake Cloudflare verification page. It fingerprints the browser, makes it solve a proof-of-work challenge, then shows a lure telling the visitor to open the Windows Run dialog and paste a command. The command downloads a script with curl.exe and runs it in a hidden PowerShell window with execution policy bypassed. The download address is written as a single decimal number, 3112851459, which Island says is 185.138.88[.]3, so a filter looking for dotted IPs would miss it.

The gate behind this route is bigger than the AI angle. Island matched 550 scan records across 211 entry domains between April 2025 and November 2026, including lookalikes of banks and insurers. Windows visitors get the clipboard command, macOS visitors get a fake Terminal update.

Store-signed malware. Eight typos of Claude, DeepSeek, Copilot, ElevenLabs and Voice AI led through a redirect hop to fake “Security Check” pages. These served Net Guard, an app presented as a Wi-Fi and DNS safety scanner. Island says the MSIX file matches the hash of the Microsoft Store listing, published by “Myflixly App”, and carries Microsoft’s Marketplace signature. Inside, Island’s static analysis found code that can change the default search engine while the user is idle, run remotely supplied JavaScript through DevTools, and edit browser profiles and install extensions. Island classes Net Guard, Privacy Keeper and MyFlixly Browser, which share the same code, as malicious browser hijackers. As of 2026-11-13 we did not open the Microsoft Store or Chrome Web Store listings, so their current status is unverified.

Search hijacking. A Perplexity typo, preplexity[.]ai, led to an install page for a Chrome extension that sets the default search provider to a third party. It needs no permissions and no executable code, since the manifest does the job. Island links it to a kit of 30 extensions, of which 13 have since been removed. About 17 listings were still live (Island’s text says 17, one figure caption says 16) and showed 972,000 declared users on November 3. Socket’s October 29 “Searchjacked” research describes the same technique at much larger scale: 319 extensions and roughly 13.4 million users, by Socket’s count.

Why it matters

None of this needs an AI vulnerability. The AI brand is just the bait. Island’s examples of the victim are an engineer, an analyst or a marketer reaching for an assistant. The practical hazard is the handoff from a web page to a command, an installer or an extension.

Island’s advice is to publish managed bookmarks for approved AI services, hunt for browser activity followed by hidden PowerShell with bypass flags or decimal IPs, treat Store-signed MSIX packages as something to check by publisher and hash rather than trust, and set the search provider and extension list centrally. Island’s own products are named in parts of that guidance, so weigh it accordingly. The first and last items work with any browser tooling.

References

Frequently Asked Questions

What is TypoPrompting?
It is the name Island's researchers gave to a delivery pattern in which a typo of an AI tool's address lands on a lookalike domain that prompts the visitor to verify, install or run something else. Island catalogued 186 lookalike domains across 27 AI brands, including Claude, ChatGPT, DeepSeek, Perplexity, Midjourney, ElevenLabs and Cursor. It traced specific ones to a ClickFix lure, Store-signed malware and a search-hijacking extension; many in the pool are parked pages.
What happens when someone lands on one of these domains?
It depends on the visitor. According to Island, a parked page hands the browser to a traffic distribution system that picks a destination from location, operating system and browser. Island observed three outcomes: a ClickFix lure that tells Windows users to paste a PowerShell command, a fake Security Check page that installs Net Guard, and a funnel to an install page for a search-hijacking Chrome extension. Some visits stop earlier or never receive a lure.
How can organisations reduce the risk?
Island's guidance is to give staff direct routes to approved AI services such as managed bookmarks, and to block confirmed harmful lookalikes. It also suggests flagging hidden PowerShell with execution-policy bypass flags or decimal IP addresses launched shortly after browsing, treating Microsoft Store-signed MSIX files as policy decisions rather than trusted by default, and centrally managing the default search provider and extensions.