AI typosquatting is catching people who mistype the address of ChatGPT, Claude, DeepSeek and other AI tools. Island researcher Oleg Zaytsev traced 186 lookalike domains across 27 AI brands, according to Island’s TypoPrompting research post, published November 12, 2026. Some of those typos led to fake verification pages, browser-hijacking apps and search-hijacking extensions.
Island calls the pattern TypoPrompting. It defines it as "AI-app lookalikes that turn an attempt to reach a familiar tool into a prompt to verify, install or run something else." GBHackers covered the research on November 13. Our readers’ staff open AI tools many times a day, often by typing the address. One wrong key is all it takes. Facts here are as of November 15, 2026.
The short version: give staff a managed bookmark for each AI tool your organization approves, so no one types the address. Block known lookalikes in Chrome or Edge. Tell staff that no website check ever needs them to paste a command into Windows or install a "security" app.
What Island found about AI typosquatting
Island’s numbers show which brands draw the most lookalikes:
- ChatGPT: 25, such as chatgpy[.]com and chtgpt[.]com.
- Claude: 18, such as cloude[.]ai, cluade[.]ai and caude[.]ai.
- DeepSeek: 16, such as deepseel[.]com and deeepseek[.]com.
- Runway: 12.
- Perplexity, Lovable and Character.ai: 11 each.
The list also covers Gemini, Grok, Mistral, NotebookLM, Midjourney, ElevenLabs, Cursor and more. Some lookalikes copy the company names, for people who type the company instead of the product.
The typos are ordinary. Island gives three Claude examples: hitting S instead of A, swapping two letters, and dropping the L. Each is a separate address, and none belongs to the real service.
Some of these AI typosquatting names are old. Island says cloude[.]ai, caude[.]ai and cluade[.]ai share a registration date of July 13, 2023.
How the trap decides who sees what
Most visitors to a typo domain see a parked page. That page hands the browser to a traffic distribution system, Island explains. The system picks a destination based on the visitor’s location, operating system and browser. "One visitor sees a parked page; another gets an attack," Island writes.
This makes AI typosquatting hard to spot. An IT person who checks a reported domain may see a harmless parking page. A staff member on Windows in another place may see a lure.
Island says the technique is not new. It credits Infoblox with documenting parked-domain traffic flowing through malicious redirect chains in December 2025. What is new is the focus on AI brands.
Three routes from a typo to malware
Island followed three AI typosquatting routes in detail.
| Route | Example typo | What the visitor got |
|---|---|---|
| ClickFix lure | cloude[.]ai | A fake Cloudflare check asking Windows users to paste a hidden PowerShell command |
| Fake “Security Check” | ckaude[.]ai, deepceek[.]com, copiot[.]com | Net Guard, a Microsoft Store-signed app with browser-hijacking code inside |
| Search hijacking | preplexity[.]ai | A Chrome extension that sends every address-bar search through a third party |
ClickFix. The Claude typo cloude[.]ai led, on Windows, to a fake Cloudflare verification page. The page walked visitors through opening the Windows Run dialog, pasting and pressing Enter. The command downloads a script and runs it in a hidden PowerShell window. It names its server as one long decimal number, which slips past filters that look for dotted IP addresses. The same gate served 211 entry domains, including lookalikes of banks, insurers, ServiceNow and Microsoft Azure.
Store-signed malware. Eight typos of Claude, DeepSeek, Copilot, ElevenLabs and Voice AI led to fake "Security Check" pages. They delivered Net Guard, a so-called Wi-Fi and DNS scanner. Island says the file matched the hash of the real Microsoft Store listing and carried Microsoft’s signature. Inside, Island found code that can change the default search engine while the user is idle, run remote scripts in the browser, and install extensions. As Island puts it, "a Store signature alone does not describe browser behavior."
Search hijacking. The Perplexity typo preplexity[.]ai led to a Chrome Web Store extension that sets the default search engine to a third-party site. It needs no declared permissions. Island tied it to 30 extensions built from one kit. Thirteen have been removed. The 17 still listed showed 972,000 users on November 3. At work, Island notes, those searches can include customer names and project details.
Why AI typosquatting matters to small organizations
Nonprofits and associations rarely block AI tools outright. Staff use them for drafts, research and member emails. Many type the address from memory or pick it from autocomplete. That habit is exactly what AI typosquatting targets.
The payloads also aim past email filters. ClickFix asks the person to run the command. A signed Store app may not trip basic defenses. A search extension with no permissions looks harmless in a review. Our phishing attacks explainer covers the older email forms of the same trick: get the person to take the action for the attacker.
If your team already allows or blocks AI browser extensions, the same review applies here. Our post on the ChatGPT Chrome extension walks through what IT should allow and block.
Step 1: Publish the real addresses as managed bookmarks
The fix Island lists first is simple: "Provide a direct route to approved AI services." That means managed bookmarks or launcher entries, plus watching for lookalikes of the tools you approve.
Start with the real addresses, taken from each vendor:
- ChatGPT: chatgpt.com. OpenAI’s own ChatGPT overview page now redirects to chatgpt.com/overview.
- Claude: claude.ai. Anthropic’s help article What is Claude? points web users to claude.ai.
- DeepSeek: chat.deepseek.com. The DeepSeek homepage links its web chat there.
Only add the tools your organization has approved.
In Chrome. Google’s help page on managing bookmarks shows the Admin console path: Menu > Devices > Chrome > Settings, then User experience > Managed bookmarks. With Chrome Enterprise Core, the path starts at Chrome browser > Settings. The bookmarks appear in a folder on users’ bookmarks bar. Without the Admin console, you can set the ManagedBookmarks policy through Group Policy or your device manager. Google says users cannot change the managed folder, though they can hide it.
In Edge. Microsoft’s ManagedFavorites policy does the same job. Microsoft says the favorites go in a folder users cannot modify, though they can hide it from the favorites bar. Extensions cannot modify managed favorites either.
Step 2: Block known lookalikes
Bookmarks help only if staff use them. Add a block list for the AI typosquatting domains you know about.
In Chrome, Google’s page on allowing or blocking websites uses Devices > Chrome > Settings, then Content > URL Blocking. In Edge, use the URLBlocklist policy. Microsoft notes the block and allow lists are limited to 1,000 entries each.
Island’s post ends with a list of indicators, including the AI lookalikes above. A block list will never catch every new typo. Treat it as a second layer behind the bookmarks.
Step 3: Tell staff about AI typosquatting in plain words
Short, specific rules work better than a long training. Here is a version you can send this week:
- Open AI tools from the bookmark folder. Do not type the address.
- No website needs you to paste a command. If a page asks you to press keys and paste something into Windows, close it and report it.
- Do not install a "security check" app from a website. Even one from the Microsoft Store.
- Watch for search changes. If your search engine suddenly changes, report it.
- Report, do not fix. Tell IT which address you typed, if you remember.
Our security awareness training post covers how to make short messages like this stick, and our social engineering explainer covers lures beyond email.
Island also lists steps for IT teams with more tools: watch for browser activity followed by Run dialog or Terminal use, hidden PowerShell and decimal IP addresses, check web-delivered app installs by publisher and hash, and set the search provider and extension policy centrally.
Frequently Asked Questions
What is AI typosquatting?
It is registering misspelled versions of AI tool addresses, such as cloude[.]ai for claude.ai, to catch people who mistype them. Island found 186 such domains across 27 AI brands.
What is TypoPrompting?
It is Island’s name for this pattern: AI lookalike sites that turn a visit into a prompt to verify, install or run something else. Island researcher Oleg Zaytsev named it.
Which AI tools have the most lookalikes?
In Island’s set, ChatGPT has 25, Claude 18 and DeepSeek 16. Runway has 12, and Perplexity, Lovable and Character.ai have 11 each.
Does every typo domain serve malware?
No. A traffic distribution system picks what each visitor sees by location, operating system and browser. Many visitors see only a parked page.
What is ClickFix?
It is a lure that asks the visitor to paste and run a command, often dressed up as a human check. In Island’s case it was a fake Cloudflare page that ran a hidden PowerShell script.
Is a Microsoft Store app always safe?
Not always. Island found Net Guard matched the real Store listing and carried Microsoft’s signature, yet contained browser-hijacking code.
What are the official addresses?
ChatGPT is at chatgpt.com, Claude at claude.ai and DeepSeek’s web chat at chat.deepseek.com, per each vendor’s own pages.
Do managed bookmarks need special software?
You need a way to push browser policy. Google’s Admin console or the ManagedBookmarks policy covers Chrome. The ManagedFavorites policy covers Edge through Group Policy or your device manager.