26 May
2010
26 May
'10
9:39 a.m.
The work on the new right manager user interface looks very promising.
Good work!
But I also think that the underlying security model must urgently be
revised.
Although I must say that I love the idea that the users can write
their own applications on top of XWiki, the current security just
doesn't support this.
When executing a script, who should be held accountable for the
actions are taken by the script?
In traditional operating systems, a user is expected to know and trust
the programs that he or she chooses to run. Considering how much
problems there are with "viruses" I'd say that this security model
works badly in traditional environments.
But in a wiki, scripts are executed at any page load. A user who
casually browse the wiki cannot know when and what scripts are
executed. Thus, it is a complete disaster to use a security model
where the users are held accountable for whatever actions the script
is doing.
For instance, to take over a wiki that is open for anonymous
commenting, just post a script that sits and waits for a user with
admin or programming rights to view the page.
How can the security model be fixed?
There are two roles that should be considered involved in the
execution of a script: user and programmer. Both the programmer and
the user should be held accountable for whatever actions are taken by
the script, but neither trust the other. The programmer says "go
ahead and run this script, but you wont be able to do anything that
you don't already have rights to". The user says "sure, I'll go ahead
and run your script, but I don't give you any rights that you don't
already have."
Hence, a script should be executed with the intersection of the
rights that the two roles possess.
But it should be possible for a programmer that have carefully
sanitized the user input to lend his or hers full privileges to the
user.
It should also be possible for users to indicate that they have
reviewed a script and indicate that they trust a certain scripts or
that they trust some programmer. In other words, the users should be
able to lend their full rights to a particular script or to scripts
written by a trusted programmer, when executed by the user.
What about contents that is generated by a script?
If the user who runs a script becomes the programmer of any scripts
that are generated and saved, we have just added one level of
indirection for attackers: An anonymous commenter could post a script
that waits for an administrator. This script should in turn post a
script (whos programmer would then be the administrator) that waits
for an administrator.
Since both user and programmer should be held accountable for the
actions taken by a script, it might be reasonable that they are also
both credited for any contents generated by the script. Thus both the
user and the programmer could be considered "the programmer" for any
new scripts generated by a script. But this might be unnecessarily
complex.
Instead, a distninction could be made on documents that is "saved
with a programmer set" and those that are not. When rendering a document
that does not have a programmer, all rights should be denied for
scripts in the document. Saving a document with yourself as "the
programmer" must not be allowed without first prompting the user for a
confirmation, where the user must input a password. This to prevent
attacks where the user is tricked into saving content with to the
programmer set.
How can this be implemented in XWiki?
A document has a "creator", which is the person who saved the first
revision of the document and each revision of the document has an
"author", which is the person who saved the revision. We have to
extend the document format to support a "programmer".
"The user" is of course the logged in user that requests the page.
The user is authenticated the ordinary way and both programmer (if
any) and user is noted in the context. Thereafter any rights check is
made by checking if both user and programmer has the right. If there
is no programmer, all right checks should return "deny", except maybe
for the "view"-right that should be checked against the user to allow
the include macro without saving with a programmer. (But this opens
up for an attack where page content which aren't viewable by everyone
can be extracted, so maybe not. It could be a configurable option.)
Obviously, it is important that comments to a page are not rendered
with a programmer set. This will make it impossible to execute
privileged scripts in comments. I don't think that is a useful
feature, anyway.
This will also require some additional user interaction. The default
should be to not set any programmer when saving a document, and this
will of course not require anything special from the user.
But when saving a page, it should be possible to select "save with me
set to the programmer". If the user chooses this option, she will be
forwarded to a confirmation page "You are about to save this content
with you set to the programmer. Please confirm by entering your save
password."
It should also be possible to select "save with me as programmer and
lend my rights to users executing scripts in this page", whereupon the
confirmation page should be marked with a big warning sign and contain
a harsh lecture on sanitizing user input. It might be a good idea to
introduce a special "save setuid" right that the user must have in
order to at all be allowed to do this. A user that have programming
rights should additionally have the option to set the programmer to an
arbitrary user. A setuid script should, of course, have a programmer
with as little privileges as possible. A set of dummy users with
varying privileges could be prepared for this purpose.
In order to allow saving documents with "the programmer" set in bulk
(for instance in the import application), there should also be a
confirmation page for allowing "saving with programmer set" throughout
a request.
Is this sufficient?
Maybe. We have at least made it a lot harder for attackers, because
now they must trick a privileged user to confirm a "save with myself
as programmer". If the same password is used for saving pages as for
logging in, this can be accomplished by spoofing a login-page. Thus,
there should be a separate "save password" which must differ from the
ordinary login password.
It might even be possible to allow users to save javascript
extensions. But there should at least be a configurable option to
demand programming rights for this. Javascripts opens up many
attack paths.
Tasks:
1. Add "programmer" attribute to document.
2. Add "execute scripts with the programmers privilege (setuid)"-flag
to document.
3. Change the RightService to:
1. if a programmer is not set, deny everything except "view", which
(as a configurable wiki option) can be checked only for the user
(to allow using the include macro without saving with a
programmer),
2. check rights only for the programmer, if the setuid-flag is set,
3. check rights only for user, if the programmer is "trusted" or
the user "trusts" the document,
4. check rights for both user and programmer, otherwise.
Although the current right service implementation can be kludged
into supporting this, it really needs to be rewritten from scratch,
both for clarity and speed.
The special treatment when the document author happens to have
programming rights must be removed.
How to determine if a programmer or the document is "trusted" is an
open question, but I guess that users with programming or admin
rights, at least, can be considered trusted.
4. Add "save password" to the user profile, which must be set to a
non-empty string that differs from the login password for the user
to be allowed to "save with myself as programmer" or "save
setuid".
5. Add user interface controls for saving a page with programmer set.
6. Add confirmation page for confirming "save with myself as
programmer".
7. Add user interface controls and confirmation page for "save setuid".
8. Add wiki-option to demand programming rights for saving javascript
extensions.
9. Fix all applications that undoubtly will be broken by this change.
Do you other people agree with me when I say that the security model
must be replaced immediately? What do you think about my suggestion?
Please, poke at it to try to find any attacks that I have not
thought of.
I have started on a new right service implementation with a
cacheing front-end. I'll create a new module under core which I'll
call 'xwiki-security' and move the right service to the new
architecture while I'm at it.
Best regards,
Andreas Jonsson