Nice proposal. Some comments below:
Here is a proposal for the XMLHttpRequest 2.0 specification for how to
enable cross-site XMLHttpRequests. I just thought I'd throw it out there
for everyone to see what security problems they can spot. :-) It's based
in part of Yahoo!'s JSONRequest proposal and on the comments that proposal
generated on the WHATWG list.
If the open() method is called with a URI that points to a third-party
domain, then:
setRequestHeader() must be prevented from setting the Domain and
Referer headers.
But, it can set "Cookie" and "Authorization" headers? That
should be okay.
send() must include a Domain: header that consists of the scheme,
hostname and port portions of the original domain of the global scope
of the script originating the request in every request in causes to be
sent to a third-party host. The scheme must be given first, in
canonical case, followed by a U+003A COLON. If the scheme supports
hosts and ports then the colon must be followed by two U+002F SOLIDUS
characters followed by the canonical punycode representation of the
host name. If the port is not the standard port for the given scheme,
then the host name must be followed by a U+003A COLON and the port
number.
I assume that IP address literals are also supported (w/ "[]"
escaping for IPv6 address literals).
If the method is GET, then the send() method must fetch the resource
indicated by the URI, then must perform an access check (see below),
and if that is positive, it must make the document available (see
below). If the check is negative, it must act as if there was a
network-level failure.
If the method is not GET, then the send() method must make a GET
request to the third-party host at the given URI, without any payload.
If the request is successful, the user agent must perform an access
check on the returned document (see below). If the check is positive,
then the UA must send the actual request (with the correct method and
complete payload), perform an access check on _that_ document (see
below), and if it is again positive, must make the return document
available (see below). If either document fails the access check,
then the user agent must act as if there had been a network-level
failure.
Note: The initial request should be cached to the extent allowed by
its response headers, so that subsequent non-GET requests can use the
cached response instead of having to hit the network twice for every
request.
Access check: If there are response headers with the name
"Access-Control", then they must have their values parsed as the data
part of an <?access-control?> PI.
Are there any cases where the allowed contents of the PI would
not be allowed as a HTTP header value?
If there are no such headers, but the document has an XML MIME type,
then the document must be processed according to the Access Control
specification.
In either case, if, according to the processing model in the Access
Control specification, the headers (if any) or PIs (if any) indicate
that the originating site is allowed to access the document, then the
access check is positive, otherwise it is negative (failed check).
Can you clarify what happens when an "Access-Control" header
does not agree with an <?access-control?> PI? Or, suppose
there are multiple "Access-Control" headers.
Making the document available: if the access check for the document
is positive, then the various responseXXXX headers must be initialised
appropriately. The security context of the document must be the same as
that of the originating site. The getAllResponseHeaders() method must
return the empty string, and the getResponseHeader() method must not
return any header values except for the following headers:
Cache-Control, Content-Language, Content-Type, Expires, Last-Modified,
Pragma. The document.cookie attribute must return the empty string,
and must do nothing when set.
It seems like it should be okay for getAllResponseHeaders() to
return the subset of response headers that getResponseHeader()
may see.
-Darin