THREAT INTELLIGENCE | November 17, 2026 — A familiar-looking address can be the beginning of an attack before a user submits any AI prompt. In research published November 12, browser-security company Island documented 186 lookalike domains spanning 27 AI brands and named the delivery pattern TypoPrompting. This is newly added IPCRIME analysis of earlier research, not a fresh breach announcement.
What the researchers followed
Island traced routes from mistyped addresses to deceptive verification pages, browser-hijacking software and search-changing extensions. One route used a ClickFix lure that asked the visitor to run a Windows command. Another offered a Microsoft Store-signed package that the researchers found could hijack browser behaviour.
The researchers also describe traffic routing based on characteristics such as the visitor’s operating system or location. Not everyone visiting a lookalike address necessarily receives the same destination. The inventory is a set of domains, not 186 verified victim organizations, and it does not establish that the impersonated AI providers were breached.
IPCRIME assessment: the unexpected instruction is the warning
The defining moment is a website changing the task. A person wanted to open a service, but the page instead demands a software installation, browser modification or command execution. Familiar branding should not make that new instruction trustworthy.
Our practical recommendation is to reach regularly used services through verified bookmarks and pause whenever a page asks for system-level actions merely to demonstrate that you are human. Check the full address rather than the logo, and obtain applications through the developer’s independently verified distribution channels. A signature identifies a signing relationship; it is not a substitute for evaluating what an application does.
Respond to what actually happened
If a user only encountered a suspicious page, distinguish that from executing its instructions or entering credentials. Where a command was run or software installed, preserve relevant details and involve the device administrator or incident-response team. Where credentials were entered, review the affected account through a trusted route. Do not assume closing the tab reverses changes already made elsewhere.
Primary source
Island’s original TypoPrompting investigation, November 12, 2026
Findings are attributed to Island; IPCRIME has not executed the payloads or independently verified every domain. Featured artwork is illustrative, not an attack screenshot.
