DemoSee the dashboard, report widget and takedown live
DE Log in Free fraud scan
BasicsAnalysis

AI-built phishing sites: what has changed, and what hasn't

Generative AI tools now clone a login page from a text prompt. What that means for detection, user training and how fast you need to respond.

Malte Zander Written with AI, reviewed by the author 6 min read
Contents
  1. What can AI tools do for phishers today?
  2. What has really changed?
  3. What hasn't changed?
  4. What should defenders change?
  5. Will AI help defenders too?
  6. FAQ
In short

Generative AI has made convincing phishing sites cheap and fast to build. In 2025, Okta saw attackers create working copies of sign-in pages from a text prompt. What hasn't changed: every fake still needs a domain, a host and a way to reach people. Defenders should stop relying on people spotting bad copies and win on speed instead: detect early, report everywhere at once, and make stolen passwords useless.

Key facts
  • On 1 July 2025, Okta reported attackers using Vercel's generative AI tool v0 to build sites that impersonate the sign-in pages of several brands, including an Okta customer (Okta).
  • The attackers hosted logos and other assets on the same trusted platform, and public GitHub repositories offer clones of the tool (Okta).
  • APWG observed about 3.8 million phishing attacks in 2025 (APWG).

What can AI tools do for phishers today?

They remove most of the craft that a convincing fake used to need:

  • Clone a page from a prompt. Okta documented working replicas of sign-in pages built with simple text instructions (Okta).

  • Write flawless text in any language. The spelling mistakes that used to give phishing away are gone.

  • Produce endless variants. New versions of the same page help attackers get past blocklists that match known copies.

  • Hide in trusted places. Okta saw phishing pages and their assets hosted on a reputable platform, which may help them avoid detection.

What has really changed?

Four things, and all of them hit the human side of defense first.

  1. Quality is no longer a warning sign. A copy with your exact layout and perfect wording looks like the real thing. Okta concludes that organizations can no longer rely on teaching users to recognize imperfect imitations (Okta).

  2. Time to launch shrinks. Building the page is no longer the slow step.

  3. More people can do it. Cloned tools and do-it-yourself guides on GitHub spread these capabilities to less skilled attackers (Okta).

  4. Platforms become part of the attack. A fake hosted on a well-known service inherits some of its reputation.

What hasn't changed?

The infrastructure. However the page was built, the attack still needs the same chain:

  • An address: a domain or a subdomain on a platform. New domains and their TLS certificates leave public traces, because Chrome requires publicly trusted certificates to be logged (Chrome).

  • A host that can be notified and has rules against phishing.

  • Distribution: an email, a text, an ad or a QR code that brings people to the page.

  • A place where stolen data goes.

Each link in that chain is a place to detect and to report. AI made the page cheap, not the chain invisible. Our article on the full phishing chain walks through every link.

What should defenders change?

  1. Retrain people for the new reality. Teach them to check the domain, to start from bookmarks or the app, and to report anything odd. Spotting typos is no longer a skill worth training.

  2. Make stolen passwords worthless. CISA calls FIDO/WebAuthn the only widely available phishing-resistant authentication (CISA). Okta recommends phishing-resistant factors and limiting sensitive apps to managed devices (Okta). More in why two-factor login doesn't stop every phishing page.

  3. Detect at the infrastructure layer. Watch new domains, certificates and hosting, and compare finds with your real page automatically. See how to spot cloned login pages.

  4. Report the same day, in parallel. When a replacement takes minutes, every day a copy stays online matters more. Our 24-hour response plan shows the sequence.

  5. Report abuse to the platform as well. When a fake runs on a site builder or hosting platform, its abuse team can act on the account. In the v0 case, Vercel restricted access to the identified sites (Okta).

Will AI help defenders too?

Yes, and the advantage is real. The same kind of model that copies a page can compare it with the original, name what was copied and rate the risk in seconds. Our article on visual similarity detection explains how that comparison works.

The limit is judgment. A tool can flag a reseller, a partner or a parody as a fake. That is why AI Fraud Protect lets detection run automatically, while a person reviews and sends every report.

See AI against AI.

In a demo, we show how AI Fraud Protect compares a find with your real page, names the copied elements and prepares reports for host, registrar and Google in one step.

Request a demo

FAQ

Can AI website builders be used to create phishing sites?

Yes. Okta documented attackers using the AI tool v0 to build copies of sign-in pages in 2025. The platform restricted access to the identified sites after Okta's findings.

Are AI-generated phishing sites harder to detect?

For people, yes, because quality no longer gives them away. For infrastructure monitoring, not necessarily: they still need domains, certificates and hosting, which leave traces.

Is security awareness training still worth it?

Yes, with new content. Teach people to verify the address, to use known entry points and to report suspicious links, instead of hunting for typos.

What protects logins best against AI-built phishing?

Phishing-resistant authentication such as FIDO/WebAuthn and passkeys. They are bound to the real domain, so they do not work on a copy, however perfect it looks.

Sources

Back to the blogAuf Deutsch lesen