Why Blocklists Can't Keep Up With AI-Built Phishing Sites
Attackers are spinning up throwaway phishing pages faster than defenders can list them. Researchers at Push Security argue the fix is watching what a page does, not where it lives.
Key points
- Push Security says AI tools now let criminals generate fresh phishing sites and lures faster than security vendors can add them to blocklists.
- Blocklists work by naming known-bad web addresses, so a site that only lives for a few hours often never gets listed in time.
- Push argues detection should shift into the browser, spotting the tricks a phishing page uses rather than its address.
- Email filters and web gateways still lean heavily on domain reputation, which is exactly where the gap opens.
- Treat any login page reached by clicking a link with fresh suspicion, even if it looks familiar.
Phishing has always been a numbers game. The attacker sends millions of lures, defenders publish a blocklist of bad web addresses, and the race is on. AI just tilted the field.
A writeup from Push Security, first reported by BleepingComputer, lays out why the old defence is consistently arriving late. Generative AI lets criminals produce convincing fake login pages and disposable domain names in minutes. Each campaign can burn through a fresh site that only lives a few hours before being replaced.
That's the whole point. If a site is dead before it hits a blocklist, the blocklist never protects anyone.
What is a blocklist, and why is it failing?
A blocklist is a running catalogue of web addresses known to be malicious. Email filters and web proxies check outgoing traffic against it and block matches. It works well against reused infrastructure but falls apart against disposable infrastructure.
Push's argument is that AI has pushed attackers firmly into the disposable camp. Toolkits now auto-generate the domain, the lure copy, and the HTML. One campaign can burn through hundreds of unique sites. By the time a vendor spots one and adds it to a feed, the attackers are already on the next batch.
The researchers also flag a second problem: AI-written phishing emails read cleanly in any language, stripping out the clumsy grammar that used to give lures away. Our 22 July story on CrowdStrike's intrusion data showed the same pattern playing out across network defences more broadly.
What does browser-level, technique-based detection mean?
It means catching a phishing page by how it behaves, not by its address. A browser extension or built-in check can watch for the signs of a credential-stealing page: a login form pointed at an unexpected destination, a fake sign-in page rendered on an unrelated domain, a page trying to capture a one-time code and relay it in real time.
Those techniques stay stable even when the domain changes hourly. A brand-new site using the same trick still triggers the same detection. That stability is the whole argument for shifting detection into the browser, and it's why the fake-site operation we covered on 25 July was so effective: it assembled its payload in browser memory, well clear of network-based reputation checks.
Push sells a product in this space, so treat the framing accordingly. The underlying point stands on its own: reputation lists degrade the moment attackers can mass-produce unique infrastructure, and AI has made that trivial.
Old defence vs new defence
| Approach | What it checks | Where AI hurts it |
|---|---|---|
| Domain blocklist | Is this web address on a known-bad list? | Sites are burned before listing |
| Email content filter | Does the message look suspicious? | AI writes clean, native-language lures |
| Browser technique detection | Is this page behaving like a credential thief? | Behaviour is stable across new domains |
What should ordinary users do?
Assume any login page you reached by clicking a link is suspect. Type your bank's or work portal's address yourself, or use a saved bookmark. If a page asks for your password and then a one-time code straight after, and the address bar looks off, stop and go directly to the real site.
Password managers help here quietly. They only auto-fill on the exact domain they saved the password for. If yours refuses to fill a login form, that's a signal, not a bug.
Blocklists aren't dead. They'll keep catching lazy attackers and old campaigns. They just can't be the main line of defence any more.
