TypoPrompting: Lookalike AI Domains Deliver ClickFix and Browser Hijackers

Authors: Cloud Security Alliance AI Safety Initiative
Published: 2026-11-15

Categories: AI Threat Intelligence
Download PDF
Download Markdown

TypoPrompting: Lookalike AI Domains Deliver ClickFix and Browser Hijackers

Key Takeaways

Island researchers have documented 186 lookalike domains imitating 27 AI brands, a delivery pattern they named “TypoPrompting,” in which a mistyped address for a tool such as Claude, ChatGPT, DeepSeek, or Perplexity leads the user to a ClickFix lure, a browser-hijacking installer, or a search-hijacking browser extension [1][2]. ChatGPT drew the most lookalikes (25), followed by Claude (18) and DeepSeek (16), and some of the Claude typos, including cloude[.]ai and cluade[.]ai, have been registered since July 13, 2023 [1]. The count is an inventory of domains, not of confirmed infections, and no victim count or loss figure has been published for this activity.

The domains lead to three distinct routes. In the ClickFix route, a fingerprinting gate decides which visitors see a fake Cloudflare verification page, and that page instructs Windows users to paste a hidden-PowerShell command that addresses its download server as a decimal integer rather than a dotted IP [1]. In the installer route, eight AI typos lead to “Security Check” pages that deliver Net Guard, an app whose package Island found to be byte-for-byte identical to a Microsoft Store listing carrying Microsoft’s Marketplace signature, yet which can change the default search engine, run remote JavaScript in the browser, and install extensions [1][2]. In the extension route, a Perplexity typo leads to a Chrome extension from a family of 30 that override the default search provider; the remaining live listings declared roughly 972,000 users as of November 3, 2026 [1][2].

For enterprises, the extension route may be the most consequential, because of its reach (roughly 972,000 declared users) and because it exposes data without any permission prompt. These extensions request no permissions and execute no JavaScript, yet they silently redirect every address-bar search to a third party, which can expose internal hostnames, customer names, and project details typed by employees [1][2]. Store signatures and permission prompts did not surface this risk in the cases Island analyzed, so organizations need controls that inspect search-provider overrides, installer provenance, and the browser-to-shell handoff directly.

Security teams should take four steps now: publish managed bookmarks or launchers for sanctioned AI tools, block the published indicators, enforce a default search provider and an extension allowlist through browser policy, and hunt for browser activity followed by Run-dialog or Terminal execution.

Background

Typosquatting is a long-established technique in web-based fraud, but the population of brands worth squatting has shifted. Consumer and enterprise AI assistants have become daily destinations for large numbers of users, and their names are short, recently coined, and, as Island’s typo inventory illustrates, easy to misspell. Island’s research categorizes the observed typos into four familiar classes: adjacent-key substitution (clsude[.]ai), transposed letters (cluade[.]ai), omitted characters (caude[.]ai), and added characters (deeepseek[.]com) [1]. Island argues that this transforms the act of reaching a trusted AI tool into a prompt to install or run something else, which is the sense in which the activity is “TypoPrompting” [1][2].

The research, authored by Island’s Oleg Zaytsev and published on November 12, 2026, reflects a data snapshot taken on November 3, 2026 [1]. Targeted brands span the major assistants and many adjacent creative and developer tools, including Anthropic Claude, OpenAI ChatGPT, Google Gemini and NotebookLM, Microsoft Copilot, Grok, DeepSeek, Mistral, Perplexity, Midjourney, ElevenLabs, Lovable, Runway, Character.ai, and Cursor [1][2]. The per-brand distribution published by Island is summarized below. The seven brands shown account for 104 of the 186 domains; the remaining 82 domains span 20 other brands, for which Island did not publish individual counts.

AI brand Lookalike domains observed Example lookalikes (defanged)
ChatGPT 25 chatgpy[.]com, chtgpt[.]com, chatlgpt[.]com
Claude 18 cloude[.]ai, cluade[.]ai, clsude[.]ai, ckaude[.]ai
DeepSeek 16 deepseel[.]com, deeepseek[.]com, deepceek[.]com
Runway 12 Not specified in source
Perplexity 11 preplexity[.]ai, perplixity[.]ai
Lovable 11 Not specified in source
Character.ai 11 Not specified in source

Source: Island, November 2026 [1].

The infrastructure behind these domains is not new, and it rests on a monetization ecosystem that defenders have tracked for some time. Most of the lookalikes resolve to parked pages, which pass visitors to a traffic distribution system that selects a destination by country, operating system, and browser [1]. In December 2025, Infoblox reported that parked domains had shifted from largely benign to predominantly malicious, with over 90% of visits to parked domains in its study leading to scams or malware (compared with under 5% in a 2014 study), and described redirect chains that profile visitors using geolocation, device fingerprinting, and cookies before choosing a destination [3][4]. TypoPrompting can be understood as that parking-and-redirect economy aimed specifically at AI-brand traffic. The underlying actors also appear to be opportunistic rather than AI-focused: the same fingerprinting gate appeared on 211 entry domains, including lookalikes for Scotiabank, Charles Schwab, Blue Cross Blue Shield, Verizon, ServiceNow, and Microsoft Azure [1][2].

ClickFix, the social engineering technique at the end of the first route, has also become a widely used initial-access method. Microsoft Threat Intelligence reported in August 2025 that ClickFix campaigns were targeting thousands of devices daily and noted that, because the technique relies on the user to launch the command, it can bypass conventional automated defenses [5]. CSA’s own July 2026 research note on ACR Stealer documented how ClickFix lures were used to steal browser session tokens and pivot into Microsoft 365 data [6]. What TypoPrompting appears to add is an inexpensive source of traffic from users who arrive already expecting to interact with an AI product, although the volume of visitors who reach the lures has not been published.

Security Analysis

Route 1: Gated ClickFix Lures

Island traced the Claude typo cloude[.]ai through a parked page and a redirect via xml.securetrackingnode[.]com to biomimicrydevtech[.]buzz, with URL parameters that tagged the source domain and a Windows-specific flow identifier [1]. A randomized subdomain of blockchainfuturedev[.]buzz then collected a FingerprintJS profile, required the browser to solve an MD5 proof-of-work challenge, and looked up the visitor’s public IP before deciding whether to serve the lure at all [1]. This gating has practical consequences for defenders. Sandboxes, URL scanners, and analysts who visit the domain from cloud infrastructure may see only benign parking content, which can lead reputation services to classify the domain as harmless while real users are still being targeted.

Visitors who passed the gate received an obfuscated script, decoded through Base64, a per-visit XOR key, and string reversal, that rendered a fake Cloudflare verification page in six languages [1]. The page placed a command on the clipboard and instructed the user to paste it into the Windows Run dialog. That command used curl.exe to save a script named W901.ps1 to the temporary directory and then executed it in PowerShell with a hidden window, no profile, and an execution-policy bypass [1]. The download server, 185.138.88[.]3, was written as the decimal integer 3112851459, a choice that likely evades detection rules written to match dotted-quad IP addresses [1]. Island’s wider corpus also included macOS variants that persuade users to run a curl … | zsh pipeline in Terminal and archive-download lures, which means a Windows-only control strategy would leave gaps [1].

Route 2: Store-Signed Browser Hijacker

The second route raises a different problem for enterprise controls, because the installer carries a legitimate Microsoft Store signature. Eight typosquats for Claude, DeepSeek, Copilot, ElevenLabs, and Voice AI passed through deezyj[.]com to fake “Security Check” pages offering Net Guard, which presents itself as a Wi-Fi and DNS safety scanner [1][2]. Island reported that the delivered MSIX package matched the hash of the Net Guard listing in the Microsoft Store, published under the name “Myflixly App,” and carried Microsoft’s Marketplace signature [1]. According to Island, the hidden browser-control features are present in every copy, including those installed directly from the Store, and they activate without asking the user [2]. Neither Island nor The Hacker News reported a response from Microsoft or the publisher, or a change in the listings’ availability, as of November 12, 2026, so defenders should not assume the Store packages have been withdrawn [1][2].

Net Guard is a Tauri application built on a remotely configured module that Island identifies as buntai-sdk [1]. Island’s analysis found that it can change the default search engine while the user is idle by driving the browser UI through UI Automation and synthetic input, execute remotely supplied JavaScript through the DevTools Runtime.evaluate interface, and modify browser profiles or install extensions through profile files, PowerShell, and the registry [1]. These features are enabled by default, the welcome screen counts down 30 seconds before enabling “browser sync” without describing the browser changes, and the application checks for a debugger and for enterprise browser management at startup [1]. Two sibling packages, Privacy Keeper and MyFlixly Browser, share the same SDK code, and Island tied the wider installer funnel to 191 lander domains, about 45 product names, and 1,437 source domains [1].

The significance of this route is that a vendor-signed, store-distributed package does not establish benign browser behavior. Allowlisting rules that trust Microsoft Store signatures, or that treat MSIX packages as lower risk than unsigned executables, would likely have permitted this installer. Remote JavaScript execution through DevTools also means the observed behavior (search redirection) may not bound the application’s actual capability, because the operator can change the payload through configuration at any time. That assessment is an inference from the documented capabilities rather than an observed escalation.

Route 3: Zero-Permission Search-Hijacking Extensions

The third route begins with preplexity[.]ai, which redirects to quickflashtool[.]com, an installation page for a Chrome Web Store extension called Quick Search [1][2]. The extension requests no permissions and contains no executable JavaScript; instead, its manifest sets the default search provider to find.flashysearches[.]com, even though the store listing advertises Bing search [1]. Island found 30 extensions built from the same template under separate developer accounts, all unlisted, of which 13 had been removed by November 3 [1]. The remaining listings, including one named “GPT for Chrome” with 80,000 users, declared roughly 972,000 users combined [1][2]. Island’s text cites 17 remaining extensions while a figure in its post and The Hacker News coverage cite 16; this note treats the exact number of live listings as unresolved.

This family is part of a larger pattern. Socket’s “Searchjacked” research, published October 29, 2026, identified 319 Chrome extensions that change default search engines and route queries through operator-controlled servers, with an estimated 13.4 million installs; the extensions were disguised as PDF converters, manual finders, safe-search tools, and AI search tools [7][8]. Island reported that Socket’s list flags 22 of the 30 extensions in this kit as malicious [1]. Because search-provider overrides are declared in the manifest rather than requested as permissions, users see no runtime permission prompt that would signal the change. For an enterprise, every query an employee types into the address bar, including internal system names, customer names, incident details, and project codenames, becomes visible to a third party [2]. The operator can also change how those queries are handled on the server side without shipping an extension update, so a store review of the extension code would not detect the change [7][8].

Assessment

Taken together, the three routes illustrate why AI-brand typosquatting deserves attention distinct from generic phishing. AI tools are used frequently, and in many organizations employees likely reach them by typing the address rather than through managed links. Many AI products, particularly developer tools, legitimately onboard users through browser extensions, desktop installers, or terminal commands, which may make similar requests from a lure seem routine and lower the friction for each lure. The table below summarizes how each route maps to defensive controls.

Route Entry example Delivered artifact Primary enterprise risk Controls likely to interrupt it
ClickFix cloude[.]ai Clipboard command; hidden PowerShell fetching W901.ps1 Arbitrary code execution, follow-on stealer or loader Run-dialog restrictions, PowerShell logging, detection of browser-to-shell chains
Store-signed installer ckaude[.]ai, deepceek[.]com Net Guard / Privacy Keeper / MyFlixly MSIX Search hijack, remote JavaScript in browser, extension installation Installer provenance policy by package family and hash, not signature alone
Search extension preplexity[.]ai Quick Search and sibling extensions Disclosure of every address-bar query to a third party Enforced default search provider, extension allowlist

Source: CSA analysis of Island research [1][2].

CSA’s earlier work on “phantom squatting” showed that attackers register brand domains that language models hallucinate, so users who trust the model’s output end up on attacker infrastructure [9]. TypoPrompting is the human-error counterpart: rather than relying on a model to produce the wrong address, it relies on the user’s fingers. Both patterns exploit the same gap: in organizations that do not provide managed entry points, the path an employee takes to reach an AI service is itself an attack surface that warrants governance alongside the service.

Recommendations

Immediate Actions

Organizations should first publish authoritative entry points for every sanctioned AI service, delivered as managed bookmarks, enterprise browser launchers, or single sign-on portal tiles, so that employees no longer type these addresses by hand. Security teams should then import the defanged indicators published by Island into DNS filtering, secure web gateways, and EDR blocklists. These include the AI lookalikes listed in this note, the ClickFix infrastructure (xml.securetrackingnode[.]com, biomimicrydevtech[.]buzz, blockchainfuturedev[.]buzz, 185.138.88[.]3), the installer hosts (deezyj[.]com, netguardsite[.]com, netguardprotect[.]com), the Store package identifiers 9MX3CH3BQ44Q, 9N1B4VPCS2S2, and 9PG42QDLFW8T, and the extension infrastructure quickflashtool[.]com and find.flashysearches[.]com [1].

In parallel, teams should run a fleet query for the Net Guard package family MyflixlyApp.NetGuard_k6q889pad57pt and its sibling packages, and for the extension IDs lihnhjjcdlfdomjliljlommppbfacmbj and nlmpdomhofmjdidcghodlgneobojdjfm [1]. Any endpoint where the browser’s default search provider is not the organization’s approved provider should be treated as a potential hijack and investigated, because the Net Guard route changes the search engine without a visible prompt.

Short-Term Mitigations

Over the next several weeks, organizations should use browser management policy to set the default search provider and prevent users from changing it, and they should move Chrome and Edge to an extension allowlist model. The Quick Search family shows that “requests no permissions” is not a meaningful safety signal, so extension review processes should explicitly check manifests for chrome_settings_overrides and search-provider entries [1][7]. On the endpoint, Windows Store and MSIX installation from web downloads should be governed by package family, publisher, and hash rather than by the presence of a Microsoft signature [1].

Detection engineering should focus on the browser-to-shell handoff that ClickFix depends on. Useful signals include Run-dialog (RunMRU) entries or Terminal activity that follow closely after browser activity, PowerShell launched with hidden-window and execution-policy-bypass flags, curl.exe invoked from explorer.exe, and command lines containing a long decimal integer in a URL host position [1][5]. Where the Run dialog is not needed, Microsoft recommends restricting it through policy [5]. Organizations should also ask their URL-reputation and secure web gateway vendors how they handle cloaked or fingerprint-gated infrastructure, since such infrastructure may present only benign content to scanners [1].

Strategic Considerations

Over the longer term, organizations should treat access paths to AI tools as part of their AI governance program rather than as general web hygiene. This includes maintaining an inventory of sanctioned AI services, monitoring newly registered lookalikes of those services, and deciding whether AI tool access should be routed through an enterprise browser or a managed launcher. User awareness programs should be updated to explain that legitimate AI products do not require pasting commands into the Run dialog or Terminal to pass a “verification,” and that a browser extension or desktop installer offered on an unfamiliar domain is a warning sign even when it is store-hosted or signed.

AI vendors and domain ecosystem participants also have a role. Vendors whose brands appear in Island’s list may wish to expand defensive registrations for common typos and to publish canonical domain lists that enterprises can import. Because defensive registration cannot cover every permutation, canonical domain lists and managed entry points remain the primary control. Island’s findings, along with Infoblox’s observation that the parking ecosystem lacks a clear path for reporting abuse, suggest that registrars, parking providers, and browser extension stores have room to improve enforcement against monetized redirect chains [1][3]. Finally, platform operators should consider whether store review processes need to account for remotely configured capabilities, such as DevTools-based script execution, that do not appear in static analysis.

CSA Resource Alignment

CSA’s most directly relevant prior work is its July 2026 research note Phantom Squatting: AI Hallucinated Domains as Phishing Infrastructure, which analyzed how attackers register domains that language models consistently hallucinate for real brands [9]. TypoPrompting extends the same threat model from model error to user error, and the recommendations in that note on monitoring brand-adjacent registrations and providing authoritative domain sources apply directly here. Organizations that have already implemented phantom-squatting monitoring can extend the same watchlists to keyboard-distance and character-omission variants of their sanctioned AI tools.

CSA’s research note ACR Stealer’s ClickFix Campaigns Drain M365 Tokens documents how ClickFix lures lead to browser session-token theft and pivoting into Microsoft 365 data [6]. Although Island did not identify the final payload behind W901.ps1, that note illustrates the class of follow-on activity that a successful ClickFix execution can enable, and its detection guidance for clipboard-to-shell execution chains is directly applicable to the first TypoPrompting route.

For control mapping, the CSA AI Controls Matrix (AICM) v1.1 provides the governing framework [10]. Its Threat and Vulnerability Management domain covers threat-intelligence ingestion and indicator blocking, its Universal Endpoint Management domain covers application allowlisting, software provenance, and browser and extension configuration, and its Human Resources domain covers security awareness training that should now include AI-brand impersonation and ClickFix lures. Mapping the controls in this note to AICM gives organizations a consistent way to report on remediation and to validate that AI tool access paths are governed alongside the AI services themselves.

References

[1] Zaytsev, O. / Island. “TypoPrompting: From AI Typos to Malware.” Island Blog, November 12, 2026.

[2] Lakshmanan, R. “186 Fake AI Domains Redirect Users to ClickFix Lures, Browser Malware, and Search Hijackers.” The Hacker News, November 12, 2026.

[3] Infoblox Threat Intel. “Parked Domains Become Weapons with Direct Search Advertising.” Infoblox Blog, December 16, 2025.

[4] Krebs, B. “Most Parked Domains Now Serving Malicious Content.” Krebs on Security, December 16, 2025.

[5] Microsoft Threat Intelligence. “Think before you Click(Fix): Analyzing the ClickFix social engineering technique.” Microsoft Security Blog, August 21, 2025.

[6] Cloud Security Alliance. “ACR Stealer’s ClickFix Campaigns Drain M365 Tokens.” CSA AI Safety Initiative, July 17, 2026.

[7] Tuckner, J. / Socket. “Searchjacked.” Socket Blog, October 29, 2026.

[8] Toulas, B. “Chrome extensions with 13.4 million installs caught hijacking searches.” BleepingComputer, October 29, 2026.

[9] Cloud Security Alliance. “Phantom Squatting: AI Hallucinated Domains as Phishing Infrastructure.” CSA AI Safety Initiative, July 1, 2026.

[10] Cloud Security Alliance. “AI Controls Matrix (AICM) v1.1.” Cloud Security Alliance, 2026.

← Back to Research Index