Island finds 186 AI lookalike domains steering visitors to malware and search hijackers
The investigation traces three attack routes, including a Store-signed app with hidden browser controls and extensions that send searches to a third party.
Loading page…
The investigation traces three attack routes, including a Store-signed app with hidden browser controls and extensions that send searches to a third party.
Listen to this story
Island’s investigation traces three ways AI-brand typos can expose visitors to unwanted software or data collection: fake verification pages that trigger command execution, a signed safety app with built-in browser controls, and an extension that redirects searches. The domains imitate 27 AI brands, but the investigation does not identify the final payload delivered by the ClickFix route. For workplaces, the risk goes beyond malware: redirected searches may disclose internal hostnames, customer names, or project details.
TechRadar counted 25 ChatGPT lookalikes, 18 for Claude, and 16 for DeepSeek; some listed domains date back to July 2023.
In the ClickFix flow, Windows visitors are prompted to run a PowerShell command, but Island did not identify the final payload.
Island says Net Guard’s browser-control functions are present in the Microsoft Store-signed app and can change Chrome settings without prompting users.
A mistyped AI website address can turn into a request to run a command or install software instead of opening a chatbot. Security company Island identified 186 lookalike domains impersonating 27 AI brands, directing visitors toward fake verification screens, browser-controlling software and extensions that divert searches to a third party.
Island calls the pattern TypoPrompting. Its investigation, covered by The Hacker News on November 12, describes sites imitating services including ChatGPT, Claude, Perplexity, Midjourney and Cursor. These are lookalike addresses, not the AI services themselves. TechRadar’s November 16 coverage says ChatGPT accounted for 25 identified lookalikes, followed by Claude with 18 and DeepSeek with 16.
Some addresses were not newly registered: records cited in the investigation date cloude[.]ai and cluade[.]ai to July 13, 2023. The researchers’ finding concerns how these lookalike addresses deliver visitors to attack pages, rather than a newly created set of AI products.
The destination can change with a visitor’s location, operating system and browser. In the ClickFix route—a trick that persuades someone to run a command—the redirect chain first identifies the browser and collects the visitor’s public IP address. It then presents a fake Cloudflare verification page.
Windows visitors are prompted to copy and execute a PowerShell command. That command downloads and runs additional scripts from attacker-controlled infrastructure; Island did not identify the final payload. The same investigation links the browser-screening script to 211 entry domains, including lookalikes of banks, insurers and enterprise software providers, placing the AI addresses within a wider delivery system.
A second route uses eight lookalike domains for Claude, DeepSeek, Copilot, ElevenLabs and Voice AI. They lead to fake “Security Check” pages that distribute Net Guard, an app presented as a Wi-Fi and DNS safety scanner—a tool for checking network and website-address settings.
Island lead security researcher Oleg Zaytsev told The Hacker News that the download is byte-for-byte identical to Net Guard as Microsoft distributes it, with Microsoft’s Store signature. His finding was not that someone had modified the download after taking it from the Store. He said the browser-control functions ship in every copy, including copies installed directly from the Store.
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.
Oleg Zaytsev, lead security researcher at Island, speaking to The Hacker News
According to Zaytsev, Net Guard waits until the user stops using the computer before changing Chrome’s search settings. It can also edit browser profiles, install extensions and run JavaScript under instructions from the operator’s server. Those capabilities go beyond the network-safety role advertised to visitors.
The third route runs through preplexity[.]ai, a misspelling of Perplexity, to a Chrome extension called Quick Search. Island says the analyzed extension needs neither executable JavaScript nor declared permissions for its main action. Instead, its configuration replaces the default search provider with find.flashysearches[.]com, sending searches to a third party.
The Hacker News describes a family of 16 extensions with 972,000 declared users as of November 3, including GPT for Chrome with 80,000. TechRadar gives a broader tally of 30 identified extensions, with 13 removed and 17 still listed on that date. The accounts do not explain how those lists overlap; their figures should not be combined into a single user or extension total.
The workplace risk extends beyond an unwanted search engine. As Island notes, searches sent through the replacement provider can contain internal hostnames, customer names and project details. A visit intended to reach an AI tool can therefore create a separate path for sensitive work information to leave the browser.
Loading discussion...
Join the conversation
Explain what responsibility stores should take for hidden app behavior.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.