- Cofense says a fake Anthropic preview invitation, active since September 2026, steals Google and Microsoft logins and has reached inboxes behind Proofpoint and Microsoft Defender.
- The campaign borrows a GoTo Webinar sender address, Zapier's link shortener and a Cloudflare CAPTCHA, then adds a custom CAPTCHA that blocks scanners.
- Check whether your email tools can follow links behind CAPTCHAs, and train staff to spot brand mismatches between the email and the login page.
Email security tools mostly judge two things: who sent a message and where its links lead. A phishing campaign documented by Cofense shows what happens when an attacker borrows both answers from companies the tools already trust.
What Cofense found
Cofense says that since September 2026, an unidentified attacker has sent emails inviting people to a private preview of a new Anthropic AI tool. The goal is to steal Google and Microsoft login details. Dark Reading, reporting on the Cofense research, says the emails got through Proofpoint's Secure Email Gateway and Microsoft Defender for Office 365.
This is one campaign, not a measure of how well those products work overall. But it shows how a weak message can still get through.
Three borrowed pieces of trust and one scanner-blocker
First, the sender. The email appears to come from customercare@gotowebinar[.]com, a real GoTo Webinar address. Dark Reading describes it as a shared address that any GoTo Webinar user might send from. Cofense notes that attackers abused the same address and the service's free trial in 2020.
Second, the link. It was shortened with a Zapier tool, which hides the real destination from a scanner. Cofense adds a detail: Zapier built this shortener to work with AI assistants such as Claude and ChatGPT. A tool built for AI assistants carries a respectable image, which may help it avoid suspicion.
Third, a Cloudflare CAPTCHA, the test that checks for a human visitor. It appears before any phishing content. Cofense notes that people see CAPTCHAs on real sites and may read one as a sign of safety.
Fourth, and different in kind, a custom CAPTCHA with a sliding control. It borrows no one's reputation. Cofense says it also appears good at detecting and blocking popular VPNs, scanners, the Tor browser and other defender tools. With two checkpoints, an automated tool may not see the final page.
How the trap works
A recipient who clicks lands on a fake Zoom meeting page at us968web-zoom[.]us. It asks them to register with a work email and then to choose "Continue with Google" or "Continue with Microsoft".
The Microsoft page captures both username and password. Its address, office[.]com-access-sharepoint[.]com, imitates real Microsoft domains. The Google page, at goo[.]gle-access-eventregister[.]icu, captures only usernames. Cofense thinks it is probably misconfigured.
Cofense lists what a stolen login can give an attacker: access to a trusted account, the emails kept in it, and a base from which to go after the wider organization.
The warning signs were visible to people
The sender name reads "Austin from nthropic", with the first letter of Anthropic missing. A GoTo Webinar email also leads to a Zoom page, a brand mismatch Cofense highlights. Dark Reading calls the campaign "pretty lazy".
Cofense could not confirm that AI built the campaign. Researcher Jacob Malimban told Dark Reading it was likely, given the many variations of the Zoom page. He added that from October 28 the attacker began spoofing Anthropic's own website instead of Zoom's.
The lesson: filters grade the envelope, people grade the story
This shows a gap in how many organizations think about phishing defense. Tools that rate reputation can be passed by an attacker who rents reputation. A shared sender address, a shortener and a CAPTCHA each look ordinary alone.
The lure is chosen with equal care. Cofense notes that many organizations are deciding how AI fits their business, so early access to an AI tool tempts people to click before they inspect. The email asks for no money and carries no malware. It asks an employee to sign in.
What to ask your team
Ask your security team whether your email tools can follow a link through a shortener and past a CAPTCHA. If the vendor says yes, ask how it handles the sliding kind that blocks scanners.
Ask how fast an employee can report a login page they typed into by mistake. A username entered on a fake page is still a signal worth acting on.
Cofense advises teaching staff to check that the brand stays the same from email to login page. It also advises checking the full domain, since logos are easy to copy. Cofense sells email defense products, so weigh its advice with that in mind. Its two checks cost little to teach.
A fake invitation that fails every human check can still pass the machine checks. Defenses need both.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Cofense.
