Fake AI sites impersonating ChatGPT, Gemini, Claude and Perplexity are targeting advertising professionals with fraudulent account connection requests. The campaign uses convincing login windows to collect passwords and multi-factor authentication codes.
Researchers at browser security company Island identified pages aimed at agency employees, media buyers and advertising account administrators. These users often manage access to several clients, potentially extending the impact of a single account compromise.
The operation also exploits interest in Meta’s recently launched Muse AI assistant, presenting fraudulent advertising tools under familiar AI branding.
Advertising Tools Provide the Bait
The fake AI sites promise services such as finding buyers, generating advertising briefs and reviewing campaign spending. However, visitors must supposedly connect an existing account before accessing those features.
Clicking the connection button opens what appears to be a Google login window. Its imitation address bar displays accounts.google.com, making the request look legitimate.
In reality, the login interface forms part of the attacker-controlled webpage. Entering credentials sends them to the phishing operation.
Advertising accounts offer several opportunities for abuse. Attackers can use available balances for fraudulent campaigns or sell access to other criminals. Meanwhile, accounts linked to multiple clients may provide a route into additional advertising operations.
Fake Browser Windows Mimic Familiar Sign-In Screens
The campaign uses a technique known as browser-in-the-browser phishing. It recreates a browser login popup inside a webpage, including visual details that users associate with a trusted sign-in process.
According to Island, the phishing kit adjusts its appearance for Windows, macOS, iOS and Android. It also supports different browser styles and dark mode.
These adaptations help the fraudulent window match the victim’s device. However, the address bar inside that window is simply part of the imitation interface.
The platform supports sign-in workflows that impersonate Google, Meta, TikTok and Okta, giving operators several ways to approach their targets.
Human Operators Control Password and MFA Requests
Once a visitor enters the phishing process, a human operator can decide which prompts appear next.
The operator may request a password up to three times or ask for an SMS or authenticator code. Other options include displaying Google approval prompts, Okta push requests or QR codes.
Consequently, the attack does not stop at collecting a username and password. It also attempts to obtain the additional codes or approvals needed to complete a login.
Operators can reject submitted codes, leave victims on waiting screens or end the interaction. This control allows them to adjust the phishing process as it unfolds.
Shared Infrastructure Reveals a Wider Operation
Island linked the advertising pages to a broader campaign involving fake recruitment opportunities and refund offers.
The researchers identified shared software components, API endpoints and hosting patterns across the pages. Additionally, misconfigured public code repositories exposed older source code, helping them trace the operation back to March.
A Telegram control channel had received hundreds of victim submissions. However, that figure does not establish how many accounts the attackers successfully compromised.
The researchers identified dozens of URLs associated with the advertising, recruitment and refund campaigns.
The Imitation Window Remains Inside the Webpage
One useful warning sign concerns how the login window behaves. A genuine separate browser popup can move beyond the boundaries of the original browser window. An imitation embedded within a webpage cannot.
The apparent login address also requires care: a trusted domain drawn inside a page does not verify where submitted information goes.
Island found that this platform rebuilds provider login screens locally and collects credentials and authentication information through its own APIs. As a result, the visible traffic can resemble communication between an AI service and a separate application backend.
0 responses to “Fake AI Sites Target Advertising Accounts With Password and MFA Theft”