A while ago I read a post entitled CSRF protections are not doomed by XSS by Stefan Esser which proposed an interesting method of using domain boundaries to stop an XSS hole in the main domain being used to extract form tokens and circumvent CSRF protections, and it would even go so far as being able to stop an XSS vuln in one form being able to circumvent the CSRF protections of another.
And even if it is more difficult to implement than simple token protections, it is still feasible if you use a wildcard DNS entry, and have a check on each form which checks if the $_SERVER['HTTP_HOST'] is the appropriate one, and if not then redirect to the appropriate one.
And I was even going to implement an example, until I realised one simple flaw; its all still hosted on the same server, and the only thing separating it is the HTTP Host header which can easily be forged via XMLHttpRequest or FlashRequest, so this protection can easily be beaten.
Which got me thinking; what would you need to add for it to work as intended? Well, you could put all the forms on separate servers, but that doesn't seem at all practical, the only really viable solution is to add more authentication mechanisms which are readable only by the specific subdomain and no others.
And thats really the biggest issue - you have to create an extra session key for every single form you have on your site, and set a cookie for them, and they all have to be set at login, and can never be regenerated because XMLHttpRequest and (possibly) FlashRequest can read response headers and extract the cookies being set on subdomains.
So while it is still of course possible to implement this, it seems completely impractical to create yet another session id which the server has to keep track of for every single form.
But if anyone has actually implemented something similar to this in any environment I'd really love to hear about it.