Four versions of work since 2.0. In short: far fewer false alarms, a much better read on addresses, and a button that behaves itself at the edges of the window.

Fewer false alarms. Version 2 rated google.com, search results and most modern web apps F. They now rate A. Checks ask whether suspicious things happen together, not whether they merely appear somewhere on the page — and news articles about scams, banks' fraud guides and "never share your recovery phrase" warnings are recognised for what they are, instead of being read as the attack they describe.

Better at spotting phishing. The whole host name is read now, not just the domain, so faceb00klogin.blogspot.com is caught where nothing used to look past blogspot.com. Look-alike letters are folded together, so robiox reads as roblox. Measured against 234 live phishing addresses and 115 ordinary ones: 68% rated Suspicious or worse from the address alone, and not one legitimate address fell below Probably Safe.

A steadier button. It opens on whichever side has room, and no longer jumps sideways or slides off the edge when dragged. The report scrolls smoothly and reopens instantly. Ratings also keep up with the page now — on apps like ChatGPT, a warning raised on a dialog you have since dismissed clears itself instead of lingering.

Still no account, no tracking, and no network requests of any kind.