Hi Peter, Thanks for your comments! >The gist of your suggestion is to use different base URLs >for the untrusted content, so that "same origin" policies >act as a sort of firewall. You propose different hostnames; >back in 2001, the acmemail webmail project did something >similar, but rather than hostnames, we chose to offer the >option of using different port numbers. Many of us ran >acmemail on https URLs, and that meant either using wildcard >certs for https (which would expose other hosts to any >flaws in acmemail) or different ports. You can see the source here: > >http://acmemail.cvs.sourceforge.net/viewvc/acmemail/acmemail/AcmemailConf.pm?view=log > >Revision 1.27 on 18 Aug 2001 introduced the change: > > # For better protection against JavaScript attacks in messages > # and attachments, it is recommended that you configure your > # Web server to listen to two ports. One of these ports should > # be designated as the "control" port, where acmemail will display > # pages it has high confidence have safe content. The other will > # be designated the "message" port, and will be used to display > # emails and their attachments > >IIRC, acmemail used querystring/URL arguments to pass authentication >tokens in the requests to the "message" host:port requests; our hope >was that all (? important?) cookies would only go to the "control" > URLs. Interesting. I'll mention this in the revised paper. >Using different ports can be a little tricky; corporate firewall admins >are very fond of disallowing https to atypical ports, for instance. > Your >hostname suggestion has other benefits if you're able to mitigate other >risks (e.g., SSO cookies scoped for all RegisteredDomain hostnames) -- Good point, but this should not be a problem if the application service provider uses a dedicated RegisteredDomain for the particular application. >being able to sandbox each document+viewer combo is great. I think you >should do some usability testing with your suggestion that the file >retrieval session record be deleted when the document is accessed, > though. >This is very likely to cause problems with user agents like Internet > Explorer >that have aggressive anti-caching stances for https content, and I > imagine >could easily cause trouble for things like chunked partial requests. Very good point!!! Plus, this makes me think of another problem with deleting the record: what if the user wants to go back to the file using the back arrow, or the browser's history, or a bookmark. > I'd >tend to treat the retrieval keys more like typical web session objects >-- in fact, I'd probably stick a hashtable of filename -> hostkey > values >in each user's web session objects, so the keys would remain valid as >long as the user was still logged in. My motivation for deleting the file retrieval session record was that the extended hostname is recorded in the browser history. So if the user neglects to log out, and is using a laptop, and the laptop is stolen (even if turned off), the thief can access the file from the history until the login session times out. But the chunked request problem you brought up trumps this. So I think now that the file retrieval session record should not be deleted until the login session record is deleted, and the user will have to be careful about logging out before leaving the laptop unattended. Also, the file retrieval session record should now be specific to a particular file, so it should have a field for the filepath, which should be checked before downloading the file. (This is equivalent to your hashtable, but I like to think of sessions as implemented by normalized relational database records, in this case by the login session record plus the collection of file-retrieval session records that refer to the login session record.) It remains to solve the back-arrow/history/bookmark problem. Here is what I propose for that: if the file retrieval session ID does not map to a file retrieval session record, the application redirects the browser to the standard user file URL. If the user is logged in, the redirected request will come in with the user-file authentication cookie, and the application will create a file retrieval session record and redirect to a new extended user-file URL. Yes, that's two redirects for each download from a bookmark, but hopefully that will not cause a noticeable additional delay, especially if keepalive is used. Will add all this to the revised paper. Thanks again, Francisco
