One letter is all it takes. Type cloude.ai instead of claude.ai and the browser still loads a page, except it belongs to someone else. That page can end with a fake verification check that runs a hidden Windows command, a Microsoft Store app that takes over the browser, or an extension that quietly reroutes every search. Island, an enterprise-browser security company, traced 186 lookalike domains across 27 AI brands and gave the delivery pattern a name: TypoPrompting. The TypoPrompting fake AI domains turn an attempt to reach a familiar tool into a prompt to verify, install or run something else.
The 186 domains are not one scam. They feed three different routes, and Island says what happens next depends on who is visiting: a traffic distribution system picks a destination from the visitor’s location, operating system and browser. ChatGPT draws the most lookalikes at 25, followed by Claude at 18 and DeepSeek at 16, according to Island. Runway has 12, and Perplexity, Lovable and Character.ai have 11 each. The set reaches well past chat assistants, with lookalikes for Gemini, Grok, Mistral, NotebookLM, Manus, Midjourney, ElevenLabs, Cursor and Copilot, and some imitate the companies themselves, such as opnai[.]com for OpenAI and anthropiv[.]com for Anthropic. No one has published a count of people who actually ran the commands, installed the app or were infected, so this is a story about the infrastructure and what it can do, not a victim count.
TypoPrompting fake AI domains: three routes from one typo
The three routes share the same starting point, a mistyped AI address, but they end in different places. Island traced each one separately, and the comparison below is our own synthesis of its findings, not a table Island publishes.
| Route | Entry point | What the visitor is told to do | Payload | Platform | Scale in the report |
|---|---|---|---|---|---|
| ClickFix lure | cloude[.]ai (Claude typo) | Pass a fake Cloudflare check by pasting a command into Run | Hidden PowerShell script (W901.ps1) from 185.138.88[.]3 | Windows (macOS variant: fake Terminal update) | 211 gate entry domains; 550 scan records |
| Net Guard installer | Eight AI typos via deezyj[.]com | Pass a fake “Security Check” and install a Wi-Fi and DNS scanner | Store-signed Net Guard MSIX with hidden browser control | Windows | 191 related lander domains; 1,437 source domains |
| Search-hijack extension | preplexity[.]ai (Perplexity typo) | Install “Quick Search” from the Chrome Web Store | Extension overriding default search to find.flashysearches[.]com | Chrome | 30 extensions; 972,000 declared users |
Our own calculation: the top three impersonated brands, ChatGPT with 25 domains, Claude with 18 and DeepSeek with 16, account for 59 of the 186 domains, or 31.7% of the whole set, and that concentration is the point.
Route 1: a fake Cloudflare check that runs a Windows command
The ClickFix route starts with a Claude typo, cloude[.]ai. Island followed a scan of that address through its redirect chain: the parked page hands the browser to xml.securetrackingnode[.]com/click, which forwards it to biomimicrydevtech[.]buzz, and then to a random subdomain of blockchainfuturedev[.]buzz. The gate fingerprints the browser, makes it solve an MD5 proof-of-work challenge and looks up the public IP address before showing anything. Only selected visits get an answer, and the answer is a fake Cloudflare verification page in English, German, French, Spanish, Italian or Portuguese.
The page walks the visitor through opening the Windows Run dialog, pasting a command and pressing Enter. The recovered command uses curl.exe to download a file called W901.ps1 and then runs it in a hidden PowerShell window with an execution-policy bypass. The download address is written as a single decimal number, 3112851459, which is the IP address 185.138.88[.]3 with the dots removed, so the URL contains no dotted IP address for a pattern-based filter to catch. Island notes the traffic-distribution URL carried a bid parameter of 0.0065 to 0.009 across three scans, which it interprets as a per-click price in dollars; no ad platform has confirmed that interpretation.
The gate behind cloude[.]ai is wider than one Claude typo. Pivoting on the gate script hash turned up 550 matching scan records across 211 entry domains between April 2025 and November 2026, including lookalikes of Scotiabank, Charles Schwab, Blue Cross Blue Shield, Verizon, ServiceNow and Microsoft Azure. The kit serves three lure types: fake verification pages on Windows designed to run a clipboard command, Terminal or security updates on macOS designed to run a shell command, and archive download pages instructing users to extract and run an executable. One macOS lure even plays a how-to video hosted on the Claude lookalike maccl0ude[.]com. That is about 2.6 scan records per entry domain on our own arithmetic, and the matching script hash appears across all 211 entry domains.
Other AI lookalikes have delivered malware directly, without the ClickFix step. In October and November 2026, chatlgpt[.]com offered a file called Installer_ChatGPT.zip, containing an executable that filescan.io classifies as malicious and that 9 of 16 MetaDefender engines detect. In 2023, midj0urney[.]org served a Windows payload that AlienVault OTX attributes to the Aurora stealer. In early 2025, deepsek[.]icu served an Android package that K7 Labs attributes to OctoV2. In each case, the lookalike domain is the route the malware took to the visitor; the payload changes with the year.
Route 2: a Microsoft Store app that hijacks the browser
The second route ends somewhere that looks far more trustworthy. Eight typosquat domains for Claude, DeepSeek, Copilot, ElevenLabs and Voice AI, among them ckaude[.]ai, deepceek[.]com, copiot[.]com, elevenlebs[.]io and vocie[.]ai, lead through deezyj[.]com to fake “Security Check” pages that deliver Net Guard, an app sold as a Wi-Fi and DNS safety scanner. The downloaded MSIX was byte-for-byte identical to the Microsoft Store-distributed Net Guard app and carried Microsoft’s Marketplace signature. The browser-control engine came with it.
Island lead researcher Oleg Zaytsev told The Hacker News: “The download from the ‘Security Check’ pages is the Microsoft Store app itself, unmodified. It is byte-for-byte identical to Net Guard as Microsoft distributes it, and it carries Microsoft’s own Store signature.” He added: “The hidden browser control comes with every copy, whether installed from the Store or from a typo site, and it switches itself on without asking. It waits until the user stops using the computer, then takes over Chrome’s search settings and can run scripts inside the browser, following instructions from the operator’s server.”
Net Guard is a Tauri application built on a module called buntai-sdk and driven by remote configuration. Its native code can change the default search engine while the user is idle, using UI Automation and synthetic input in Chrome, with a matching Firefox flow; run remotely supplied JavaScript in the browser through DevTools Runtime.evaluate; and edit browser profiles and install extensions through Chrome profile files, PowerShell and the registry. These features are all enabled by default. The welcome screens count down 30 seconds and then submit browser synchronisation as enabled, without mentioning the browser changes. The app runs at startup with full trust and checks for a debugger and for enterprise browser management.
Island also analysed Net Guard 1.0.9.0, Privacy Keeper 1.0.10.0 and MyFlixly Browser 1.0.10.0. All three share the buntai-sdk code, and Island’s static analysis classifies each as a malicious browser hijacker. That classification is Island’s own assessment; Microsoft has not commented in any source read for this article. The Net Guard package is byte-for-byte the Store-distributed app, and Island’s classification rests on the app’s hidden browser-control behaviour, not on any modification to the download. The AI-typo installer funnel spans 191 related lander domains, about 45 product names and 1,437 source domains across AI tools, banks, insurers and telecoms. Net Guard-branded installer pages appeared on netguardsite[.]com in May 2026, and records for netguardprotect[.]com begin in August 2026.
Route 3: an extension that quietly changes your search engine
The third route starts at the Perplexity typo preplexity[.]ai. A related-search page passes the browser to quickflashtool[.]com, an installation page for a Chrome Web Store extension called “Quick Search”. The extension needs no executable JavaScript and no declared permissions to do its job. Its listing promises Bing search, but its manifest sets the default search provider to find.flashysearches[.]com, so every address-bar search passes through a third party first. At work, those searches can include internal hostnames, customer names and project details, according to Island.
Quick Search is one of 30 extensions linked by a shared landing kit. They use the same minimal search-override template under separate developer accounts, and every listing Island checked is unlisted, so the only way in is an installation page. Since the kit’s landing pages were first observed, 13 of the 30 extensions have been removed from the Chrome Web Store, and the 17 still listed displayed 972,000 users in total on November 3, 2026. That is 43% removed and 57% still listed on our own arithmetic, with an average of about 57,200 declared users per still-listed extension. “GPT for Chrome” alone has 80,000 users, about 8.2% of the 972,000 total. The 972,000 figure counts declared extension installs, not confirmed victims, and no source in this article names a count of people whose searches were actually hijacked. Socket documented the same search-hijacking technique at a larger scale in its “Searchjacked” research on October 29, 2026, and lists 22 of the 30 extensions among its malicious indicators.
One detail to read carefully: Island’s report text says 17 extensions are still listed, while its own Figure 8 caption says 16 live listings, and The Hacker News reports 16 live listings with the same 972,000-user total. This article uses Island’s 17, from the report text, and does not state 16 as fact.
This infrastructure is more than three years old
The earliest lookalike domains were registered on July 13, 2023, including caude[.]ai, cloude[.]ai, cluade[.]ai and the Perplexity lookalikes perplecity[.]ai and perplixity[.]ai. Video-tool lookalikes joined between November 2023 and November 2024. The earliest matching records in the wider gate-kit corpus appear in April 2025. A central redirector, a11[.]click, was registered on March 6, 2026. The research snapshot connecting cloude[.]ai to the Windows ClickFix command and eight AI typos to the installer funnel was taken on November 3, 2026. That is about three years and four months from the first registrations to disclosure on our own arithmetic.
An 81-domain portfolio, including about 23 AI lookalikes, is tied by shared publisher keys and affiliate IDs to a11[.]click. Its captured redirects rotate between Adsterra smartlinks and Monica or Pollo affiliate referrals, and the same Pollo affiliate ID appears behind the video-tool typos hailou[.]ai, hygen[.]ai and viggie[.]ai. Three Adsterra keys replaced one another between March and October 2026 while the source domains stayed the same. No threat actor or operator is named in any source read for this article; Island attributes the delivery pattern, not a group.
How do I check if a typo caught me?
No single check proves whether a typo caught you, so look for the signs in three places: your Chrome extensions for a changed default search provider, your installed Windows apps for Net Guard, Privacy Keeper or MyFlixly Browser, and any “verify you are human” page that asked you to paste a command.
- Open Chrome settings, Extensions, and remove any extension you do not recognise, especially one that changed your default search provider.
- Check the default search setting itself, and set it back to the provider you chose.
- Open Windows Settings, Apps, and uninstall Net Guard, Privacy Keeper or MyFlixly Browser if any of them is installed.
- Never paste a command from a “verify you are human” page into Run, Terminal or PowerShell.
- Bookmark the AI tools you use every day instead of typing their addresses.
- If anything unfamiliar turns up, remove it, and change the passwords you typed while it was present.
These steps are general precautions, not instructions from Island. Island’s own advice, as reported by The Hacker News, is that reaching the right AI service “should be a managed path, not a test of typing accuracy”.
What to watch next
Three things are worth watching. First, whether the 17 still-listed extensions stay up or come down, and whether Google or Microsoft comment on the Store-signed Net Guard package, since the app is byte-for-byte the Store-distributed version. Second, whether future observations show the a11[.]click portfolio still rotating Adsterra keys and affiliate referrals, or whether the redirector itself is taken down. Third, whether the TypoPrompting fake AI domains spread beyond the 186 Island traced. Island already found lookalikes across 27 AI brands.
The typo is the oldest trick in the book, and it works because the destination looks right. A page that says “verify you are human” and asks you to paste a command is not a verification page. An app that calls itself a Wi-Fi scanner and then changes your search engine while you are idle is not a scanner. And an extension that promises Bing search while sending every query to a third party is not a search tool. A mistyped address can lead to a very different destination, and that is the whole trick: the page looks right either way, and one letter is the difference.


