The short version
- The technique. Each fake portal is built around a single "Connect" button that opens a browser-in-the-browser window with a forged address bar showing accounts.google.com or an Okta tenant. Island says a live operator then collects passwords and MFA codes in real time.
- Who is targeted. Agency staff, media buyers and administrators of manager accounts, whose logins reach multiple downstream client accounts. Island says victims are probably directed to the pages through fake invitation emails that impersonate the brands.
- The motive. Stolen accounts can be used to spend available balances on fraudulent campaigns or sold to other criminals, especially when they have a clean spend history. Island says attackers typically add their own administrators and downgrade the legitimate owner.
- A wider operation. The AI ad pages share a Next.js and Socket.IO stack and common endpoints with fake Google Ads refund and payment-confirmation pages and recruitment lures using Tesla, Louis Vuitton, Nike and Adecco. Misconfigured public GitHub repositories exposed older source code, tracing activity to March.
- What to do. Island recommends phishing-resistant authentication, review of advertising control changes, and scrutiny of AI integrations before accounts are connected.
Timeline
| Date | Event |
|---|---|
| March | Activity on the phishing platform can be traced back to March, according to Island, after older source code was found in misconfigured public GitHub repositories. |
| 16 September | The domain museads.ai appears, a little over a week after Meta launched its Muse AI agent, presenting itself as an AI ads manager for paid media workflows. |
| 6 October | Island publishes its report on the human-operated phishing platform and the domains tied to it. |
How the fake portals work
Island says the sites pose as ad products for ChatGPT, Gemini, Claude, Perplexity, Meta Muse and Manus, each with its own pitch: campaign briefs, manager-account support, spend audits or a private Meta integration. Every one funnels the visitor to the same action, a "Connect" button, which opens a browser window rendered inside the real browser while the genuine page stays on the phishing domain.
The kit restyles itself for Windows, macOS, iOS and Android, including dark mode, and fingerprints the device, sending the result to the attacker over Socket.IO. A human operator then controls the session: prompting for a password up to three times, asking for an SMS or authenticator code, triggering Okta push requests or Google approvals, showing a QR code, rejecting submitted codes or holding the victim on a waiting screen.
The researchers note one simple check. A real sign-in pop-up can be dragged outside the browser window and resized; an embedded frame cannot. Island also found that the phishing platform supports Google, Meta, TikTok and Okta sign-in flows, and that its interface rebuilds the provider's login page locally rather than proxying traffic through to it.
Who is affected and what recovery involves
The campaign is aimed at agency staff, media buyers and manager-account administrators, whose accounts can extend to multiple downstream clients. Island says the card is usually the easy part to fix, while account recovery can take weeks or months and, for a manager account, the damage reaches the agency's clients while the account keeps serving ads.
The researchers place the campaign in a wider pattern of ad account theft. A Mimecast report from July 2026 on malware families including VietCredCare, DuckTail, NodeStealer and PXA Stealer said they had turned ad account theft into widespread commodity crime: budgets drained and accounts with good reputations resold on underground markets.
Shared infrastructure and separate activity
All pages tied to the operation use Next.js and Socket.IO and reach the same API endpoints, with many on Vercel frontends and Railway or Render backends. Alongside the AI ad portals, Island describes two other lures: Google Ads-themed refund claims and payment confirmations, and recruitment pages naming Tesla, Louis Vuitton, Nike and Adecco. The attackers' Telegram control channel had received hundreds of victim submissions, which Island says does not necessarily reflect the number of successful compromises.
Island separately reported abuse of paid search results that sent users to custom GPTs or shared AI chat content and then to a fake Cloudflare verification page, using ClickFix-style lures to deliver NetSupport RAT, a remote access trojan. Over three months to August 2026 that cluster included about 850 paid-ad landings, 26 lookalike ChatGPT destinations and 71 Google Ads campaign IDs. The researchers said no flaw in ChatGPT or Google was required. No source links this to the ad-portal campaign.
What is not known yet
- Who operates the phishing platform and where the operators are based.
- How many advertising accounts were actually compromised rather than merely submitted.
- How much ad spend was taken, and whether stolen accounts were resold.
- Whether the fake domains, the Telegram channel and the GitHub repositories are still live.
- Whether Google, Meta, TikTok or Okta have detected or blocked the stolen credentials on their side.
What to do
If you manage advertising or agency accounts
- Move Google, Meta, TikTok and Okta logins for ad accounts to phishing-resistant authentication such as passkeys or hardware security keys, rather than SMS codes or push approvals.
- Review the user list and recent admin changes on each ad and manager account, removing administrators you do not recognise.
- Check recent campaigns and spend for activity you did not authorise.
- Verify an AI advertising tool's domain independently before connecting any ad account to it.
If you run IT or security for an agency
- Block or alert on the domains listed in Island's report.
- Brief media buyers that a genuine sign-in pop-up can be dragged outside the browser window and resized, and that an embedded fake cannot.
- Require a review step before staff connect ad accounts to third-party AI products or grant them manager-account access.
For everyone else
- Prefer phishing-resistant authentication wherever it is offered, because a live operator can relay codes and approve push prompts.
- Treat unsolicited invitations, refund notices and payment confirmations that lead to a sign-in page as suspect and reach the service through its own site instead.
Update log
- First published.
Sources
New incident reports by email
One email a week with the reports published since the last one.