Dave Korn announced on the Full Disclosure and Bugtraq security lists that Microsoft is bypassing local lookups for some hosts, meaning that you can't locally block some sites through your HOSTS file. All of these sites are MicroSoft controlled sites:

DomainScreenList:
windowsupdate.microsoft.com
windowsupdate.com
microsoftupdate.com
download.microsoft.com
update.microsoft.com

HostsScreenList:
microsoft.com
www.microsoft.com
support.microsoft.com
wustats.microsoft.com
microsoftupdate.microsoft.com
office.microsoft.com
msdn.microsoft.com
go.microsoft.com
msn.com
www.msn.com
msdn.com
www.msdn.com

A quick check suggests that this behavior debuted with XP SP2, and is present on 2003 SP1 as well. (I haven't looked at 2003 RTM, but it would be interesting if someone please would.) Although one could argue that this measure is intended to thwart attempts to block updating Microsoft products, it's indefensible because:

1) It's a point-in-time, cat-and-mouse defense against an ephemeral malware technique, a change that causes permanent headaches in situations like yours, and the potential for negative publicity as a result.

2) As far as I know, their malicious software removal tool didn't exist back when this behavior was created, so what good was keeping access to Microsoft open going to do an infected system? What good does it do to install a patch for a vulnerability that's already been exploited onto the computer of the archetypal "home user"?

3) Although it falls in line with removing raw sockets and limiting half-open TCP connections, making these Microsoft hosts and domain unfilterable is even more egregious because of the implications you mentioned, and because this behavior was never publicly documented.

4) Their selectiveness seems unfair. I'm sure all the antivirus/antispyware companies whose domains regularly end up in hosts-files would love to be added to the list, too. (So would everyone else whose software reports "anonymous usage statistics" and all the other companies making money from web advertising.*) Going back to #3, it would have been more disruptive but less controversial if they had removed regard for the hosts-file entirely, or made the resolver only consult the hosts-file after all else failed, thereby preventing it from being used for blocking. It's not a great analogy, but this move is sort of like if they had only blocked raw IP packets headed for a Microsoft IP address, instead of raw sockets entirely.