RE: Comments on: Access Control for Cross-site Requests
Close, Tyler J. <tyler.close <at> hp.com>
2008-01-02 18:26:03 GMT
Ian Hickson wrote:
> On Mon, 31 Dec 2007, Close, Tyler J. wrote:
> > 1. Browser detects a cross-domain request
> > 2. Browser sends GET request to /1234567890/Referer-Root
> > 3. If server responds with a 200:
> > - let through the cross-domain request, but include a
> > header. The value of the Referer-Root header is the relative URL /,
> > resolved against the URL of the page that produced the request. HTTP
> > caching mechanisms should be used on the GET request of step 2.
> > 4. If the server does not respond with a 200, reject the
> > request.
> This is a very dangerous design. It requires authors to be able to
> guarentee that every resource across their entire server is capable of
> handling cross-domain requests safely. Security features with the
> potential damage of cross-site attacks need to default to a
> safe state on
> a per-resource basis, IMHO.
Sure, but the question is: "Who's responsibility is it?". In my opinion, it is the server's responsibility
to ensure a safe default for each resource. You seem to have the perspective that it's the client's responsibility.
With the "OPTIONS *" request, or "GET /special/URL" request, I am just trying to establish that the client
and server know what the other is saying. We can then leave each to protect its own interests. This division
of labour requires the least amount of coordination between client and server and puts the enforcement
task with the party who most wants a policy enforced. In my opinion, these are the two high order bits when
judging a solution to this problem.
By giving the thumbs up to the client's query about understanding of the Referer-Root header, the server is
saying: "Yes, I've got a filter in place to control access to individual resources. Go ahead and send the
requests through, but label them with their originator information. I've got it covered from there." I
imagine the sysadmin for the server would setup this filtering and then provide some guidelines to web
page authors on how to activate cross-domain requests. These guidelines may very well look much like the
ones this WG has designed. The difference is that they are an agreement between web page authors and their
own server, not between web page authors and all client software for the Web. The former means Web content
publishers can develop their own policy and enforcement mechanisms, and have confidence in how they are
being used, regardless of which client software any user may have.
We can twiddle with the details of how client and server establish that both understand the meaning of the
new Referer-Root header. I've offered 2 possibilities: one that fits web-arch better, and one that fits
the WG's self-imposed rule of no modifications to server software. I only offered these possibilities to
show that the problem can be solved in this way. If the WG really needs me to design the exact details of this
handshake, I'll do so, but I imagine WG members are equally capable to this task. I suspect this handshake
can also be designed so as not to require modification of the server's software. I don't think this is a high
priority, but I undertand that this WG does. For example,
> Furthermore, relying on a 200 OK response would immediately
> expose all the
> sites that return 200 OK for all URIs to cross-site-scripting attacks.
Fine, so have the server respond with something unmistakeable. In the extreme, the server could be
required to respond with a text entity containing a large random number whose value is specified in the rec.
> (There is also the problem that any fixed URI has -- /robots.txt
> /w3c/p3p.xml, etc are all considered very bad design from a
> URI point of
> view, as they require an entire domain to always be under the
> control of
> the same person, whereas here we might well have cases where
> partitions of
> a domain are under the control of particular users, with the different
> partitions having different policies.)
In which case it is only necessary that whoever has control of the special URL has coordinated with the other
users of the host. These are all arrangements that can be made server side, without exposing the details to clients.
Again, I think the "OPTIONS *" request is the more acceptable design, since it doesn't use a well-known URL,
but I also don't think this issue is that big a deal.
> Furthermore, there is a desire for a design that can be applied purely
> static data where the user has no server-side control whatsoever. With
> your proposal, even publishing a single text file or XML file
> with some
> data would require scripting, which seems like a large onus to put on
> authors who are quite likely inexperienced in security matters.
Again, this is server-side setup. Particular servers may well choose to deploy technology much like what
this WG has created. We just don't have to say that everyone has to do it that way. We don't need that broad an
agreement. These technology choices can be confined to the server-side. We only need a way for client and
server to signal the presense of such a mechanism, in particular, declaring that each understands the
meaning of the Referer-Root header.