Fake Anthropic Preview Invites Target Google and Microsoft Credentials

The real weakness in the fake Anthropic preview campaign is how closely it follows normal workplace routines. An employee sees a polished invitation to an exclusive AI session, registers through familiar services and reaches a login page that resembles Google or Microsoft. Each step borrows credibility from a different brand, while the campaign is built to keep security scanners from following the path.

Cybersecurity firm Cofense documented the phishing operation, which uses GoTo Webinar branding to draw in enterprise users and ultimately harvest account credentials. The lure is a private, invitation-only discussion about AI systems, research and frontier models. That pitch fits a moment when companies are weighing how to use AI and employees are eager to get early access to new tools.

Frank Dickson of Dickson Research says the promise of exclusivity is doing much of the work. “Fear of missing out is the oldest lure in the book,” he said. The attackers selected Anthropic because it was prominent in the news; Dickson expects the same tactic could shift to another company as attention moves.

A familiar invitation masks a changing trail

GoTo Webinar’s routine meeting emails often pass through corporate email defenses, giving attackers a credible wrapper for the first click. Cofense found that the invitation appears to come from a legitimate sender and directs recipients to register. The link then routes them to a compromised URL, with a URL shortener and carefully constructed subdomains obscuring the destination.

The message borrows visual cues from GoTo Webinar and uses language designed to make the event sound selective. Its apparent sender address is customercare@gotowebinar[.]com, although GoTo Webinar’s website is www.goto.com. The sender display name offers a smaller clue: “Austin from nthropic,” with the first “A” missing. Cofense calls the misspelling a useful warning sign, even if it’s easy to overlook.

The link also appears to have been generated by Zapier, a legitimate automation service whose tools can connect with Claude, ChatGPT and other systems compatible with Model Context Protocol (MCP). That association can make the URL look plausible to someone expecting AI-related integrations. A short link and subdomains with dashes further make the real destination harder to read at a glance.

After the first click, a Cloudflare CAPTCHA appears. Completing it leads to a page styled as a Zoom invitation, where the visitor is asked to register with a work email address. Choosing “Continue with Google” or “Continue with Microsoft” brings up a second CAPTCHA before the credential prompt appears.

The sequence is designed to interfere with automated inspection. Cofense says the CAPTCHA pages appear to block VPN services, scanners, Tor browsers and other analysis tools. Splitting the journey across two CAPTCHA checks can also make it harder for security systems to follow the full chain. Dickson summed up the inversion: “A CAPTCHA was built to keep bots out. Here, it keeps the security scanners out and lets the victims in.”

The final pages imitate familiar sign-ins

The campaign’s Google and Microsoft credential pages are similar to the real services, but the data they collect differs. The Google imitation captures usernames but not passwords, which Cofense researchers suspect may reflect a setup error. The Microsoft imitation collects both. The researchers say both pages look remarkably similar to their legitimate counterparts.

The domains are designed to reward skimming. The fake Google address is goo[.]gle-access-eventregister[.]icu. The Microsoft page uses office[.]com-access-sharepoint[.]com, a construction that evokes Office and SharePoint while placing the actual domain elsewhere in the address. A familiar logo or page layout doesn’t establish that a login is authentic; the domain does.

Neither page gives users a confirmation message or sends them on to the real Google or Microsoft service after submission. By then, the credentials have already been sent to the attackers. A compromised account could expose sensitive documents or provide an initial foothold for further access inside the organization, Cofense warns.

The collection gap on the fake Google page may be temporary. Dickson expects the operators could correct it, making the page capable of capturing passwords as well. Organizations shouldn’t treat the current limitation as protection from the campaign’s broader approach.

Make the account harder to take over

Employees can watch for inconsistencies across the invitation, meeting page and sign-in screen. In this case, the path shifts from GoTo Webinar branding to a Zoom-style registration page, then to Google or Microsoft credentials. Dickson points to that changing identity as a clear signal: three brands appear in three clicks.

Cofense recommends checking that branding stays consistent, looking for small errors and verifying that the login domain belongs to the service being used. Those checks help, but the campaign shows why a defense that depends on every employee inspecting every URL has limits. AI companies’ gated previews and waitlists have made exclusive invitations familiar, and even careful staff may encounter a polished lure at a busy moment.

“Training helps, but I would not hang the defense on an employee squinting at a URL at 4:45 on a Friday,” Dickson said. His point is that an organization’s security should account for predictable human error instead of treating a clicked link as an individual failure.

Phishing-resistant multi-factor authentication, including passkeys, can make a stolen password much less useful. Dickson’s practical steps for enterprises are to confirm AI preview invitations through an account team or another known contact, treat changes in branding between email and login as a warning, and deploy phishing-resistant MFA. If an employee has entered credentials into a suspected phishing page, the organization should reset them and review the account’s activity.

The campaign uses familiar services to make a suspicious journey feel routine, then relies on lookalike pages to collect credentials. The durable response is to verify invitations through trusted channels and reduce what a stolen password can unlock.