TypoPrompting campaign exploits 186 fake AI service domains

Cybersecurity company Island has identified a campaign it calls TypoPrompting, in which 186 lookalike domains impersonating 27 artificial-intelligence brands redirect users to ClickFix lures, browser-hijacking malware, or search-hijacking extensions. The targets include Anthropic Claude, OpenAI ChatGPT, Google Gemini, Gemini Notebook, Microsoft Copilot, SpaceXAI Grok, DeepSeek, Mistral, Perplexity, Midjourney, ElevenLabs, Lovable and Cursor.
The campaign turns a simple mistyped address into a different request: visitors may be asked to verify themselves, install software or execute a command. Island said some of the domains, including cloude[.]ai and cluade[.]ai, were registered as early as July 13, 2023.
Three delivery paths from AI lookalikes
One branch leads through parked domains and a redirect chain that fingerprints the browser and collects the visitor's public IP address. It then presents a fake Cloudflare verification page that serves a PowerShell command. Island found the fingerprinting gate across a broader set of 211 entry domains, including lookalikes for banks, insurers and enterprise software providers.
Those pages use several lure formats: fake Windows verification screens designed to make a user run a clipboard command, macOS terminal or security-update prompts that call for a shell command, and archive-download pages that instruct users to extract and execute a file. The technique fits the wider pattern of ClickFix chains turn routine prompts into execution steps where ClickFix chains turn a seemingly routine instruction into an execution step.
A second route uses eight typosquat domains for Claude, DeepSeek, Copilot, ElevenLabs and Voice AI to deliver fake “Security Check” pages. These ultimately install Net Guard, an application that claims to scan Wi-Fi and DNS safety and matches the MSIX hash of a legitimate application with the same name.
Search redirection can expose workplace queries
Island said the malicious Net Guard variant can run attacker-controlled JavaScript, alter the default search engine while the user is idle, edit browser profiles and install arbitrary Chrome extensions. That makes the apparent legitimacy of the installer particularly relevant: a matching hash for a legitimate app name does not describe the capabilities observed in the delivered application.
The third path begins with preplexity[.]ai, a Perplexity lookalike, and leads to a Chrome Web Store extension named Quick Search. The extension does not need executable JavaScript or declared permissions for its core action because its manifest replaces the browser's default search provider with find.flashysearches[.]com. Searches are then forwarded to a third party.
Quick Search belongs to a set of 16 extensions, including GPT for Chrome, which Island said had 80,000 users. The group of extensions had 972,000 declared users as of November 3, 2026. In a business setting, redirected searches may contain internal hostnames, customer names and project details.
Managed access reduces the typing risk
The campaign also illustrates how parked-domain traffic can be routed through third-party traffic distribution systems after visitors are profiled by host and device characteristics. The resulting chain can convert a near-match to a familiar AI service into malware delivery or browser search interception.
Businesses should provide approved AI services through managed bookmarks, portals or other controlled access paths, and should monitor for lookalike domains and unapproved browser extensions. Employees should treat any verification page that requests a terminal, shell or clipboard command as a warning sign rather than a routine access step.

