implemented html sanitisation

[ikiwiki] / doc / security.mdwn

Let's do an ikiwiki security analysis..

If you are using ikiwiki to render pages that only you can edit, do not
generate any wrappers, and do not use the cgi, then there are no more
security issues with this program than with cat(1). If, however, you let
others edit pages in your wiki, then some possible security issues do need
to be kept in mind.

# Probable holes

## XSS holes in CGI output

ikiwiki has not yet been audited to ensure that all cgi script output is
sanitised to prevent XSS attacks.

## image files etc attacks

If it enounters a file type it does not understand, ikiwiki just copies it
into place. So if you let users add any kind of file they like, they can
upload images, movies, windows executables, css files, etc (though not html
files). If these files exploit security holes in the browser of someone
who's viewing the wiki, that can be a security problem.

Of course nobody else seems to worry about this in other wikis, so should we?

## web server attacks

If your web server does any parsing of special sorts of files (for example,
server parsed html files), then if you let anyone else add files to the wiki,
they can try to use this to exploit your web server.

## multiple accessors of wiki directory

If multiple people can write to the source directory ikiwiki is using, or
to the destination directory it writes files to, then one can cause trouble
for the other when they run ikiwiki through symlink attacks.

So it's best if only one person can ever write to those directories.

## setup files

Setup files are not safe to keep in subversion with the rest of the wiki.
Just don't do it. [[ikiwiki.setup]] is *not* used as the setup file for
this wiki, BTW.

## svn commit logs

Anyone with svn commit access can forge "web commit from foo" and make it
appear on [[RecentChanges]] like foo committed. One way to avoid this would
be to limit web commits to those done by a certian user.

It's actually possible to force a whole series of svn commits to appear to
have come just before yours, by forging svn log output. This could be
guarded against by using svn log --xml.

ikiwiki escapes any html in svn commit logs to prevent other mischief.

## page locking can be bypassed via direct svn commits

A [[lock]]ed page can only be edited on the web by an admin, but
anyone who is allowed to commit direct to svn can bypass this. This is by
design, although a subversion pre-commit hook could be used to prevent
editing of locked pages when using subversion, if you really need to.

----

# Hopefully non-holes

(AKA, the assumptions that will be the root of most security holes...)

## exploting ikiwiki with bad content

Someone could add bad content to the wiki and hope to exploit ikiwiki.
Note that ikiwiki runs with perl taint checks on, so this is unlikely.

## publishing cgi scripts

ikiwiki does not allow cgi scripts to be published as part of the wiki. Or
rather, the script is published, but it's not marked executable (except in
the case of "destination directory file replacement" below), so hopefully
your web server will not run it.

## suid wrappers

ikiwiki --wrapper is intended to generate a wrapper program that
runs ikiwiki to update a given wiki. The wrapper can in turn be made suid,
for example to be used in a [[post-commit]] hook by people who cannot write
to the html pages, etc.

If the wrapper script is made suid, then any bugs in this wrapper would be
security holes. The wrapper is written as securely as I know how, is based
on code that has a history of security use long before ikiwiki, and there's
been no problem yet.

## shell exploits

ikiwiki does not expose untrusted data to the shell. In fact it doesn't use
system() at all, and the only use of backticks is on data supplied by the
wiki admin and untainted filenames. And it runs with taint checks on of
course..

## cgi data security

When ikiwiki runs as a cgi to edit a page, it is passed the name of the
page to edit. It has to make sure to sanitise this page, to prevent eg,
editing of ../../../foo, or editing of files that are not part of the wiki,
such as subversion dotfiles. This is done by sanitising the filename
removing unallowed characters, then making sure it doesn't start with "/"
or contain ".." or "/.svn/". Annoyingly ad-hoc, this kind of code is where
security holes breed. It needs a test suite at the very least.

## CGI::Session security

I've audited this module and it is massively insecure by default. ikiwiki
uses it in one of the few secure ways; by forcing it to write to a
directory it controls (and not /tmp) and by setting a umask that makes the
file not be world readable.

## cgi password security

Login to the wiki involves sending a password in cleartext over the net.
Cracking the password only allows editing the wiki as that user though.
If you care, you can use https, I suppose.

# Fixed holes

_(Unless otherwise noted, these were discovered and immediatey fixed by the
ikiwiki developers.)_

## destination directory file replacement

Any file in the destination directory that is a valid page filename can be
replaced, even if it was not originally rendered from a page. For example,
ikiwiki.cgi could be edited in the wiki, and it would write out a
replacement. File permission is preseved. Yipes!

This was fixed by making ikiwiki check if the file it's writing to exists;
if it does then it has to be a file that it's aware of creating before, or
it will refuse to create it.

Still, this sort of attack is something to keep in mind.

## symlink attacks

Could a committer trick ikiwiki into following a symlink and operating on
some other tree that it shouldn't? svn supports symlinks, so one can get
into the repo. ikiwiki uses File::Find to traverse the repo, and does not
tell it to follow symlinks, but it might be possible to race replacing a
directory with a symlink and trick it into following the link.

Also, if someone checks in a symlink to /etc/passwd, ikiwiki would read and
publish that, which could be used to expose files a committer otherwise
wouldn't see.

To avoid this, ikiwiki will skip over symlinks when scanning for pages, and
uses locking to prevent more than one instance running at a time. The lock
prevents one ikiwiki from running a svn up at the wrong time to race
another ikiwiki. So only attackers who can write to the working copy on
their own can race it.

## symlink + cgi attacks

Similarly, a svn commit of a symlink could be made, ikiwiki ignores it
because of the above, but the symlink is still there, and then you edit the
page from the web, which follows the symlink when reading the page, and
again when saving the changed page.

This was fixed by making ikiwiki refuse to read or write to files that are
symlinks, or that are in subdirectories that are symlinks, combined with
the above locking.

## underlaydir override attacks

ikiwiki also scans an underlaydir for pages, this is used to provide stock
pages to all wikis w/o needing to copy them into the wiki. Since ikiwiki
internally stores only the base filename from the underlaydir or srcdir,
and searches for a file in either directory when reading a page source,
there is the potential for ikiwiki's scanner to reject a file from the
srcdir for some reason (such as it being a symlink), find a valid copy of
the file in the underlaydir, and then when loading the file, mistekenly
load the bad file from the srcdir.

This attack is avoided by making ikiwiki scan the srcdir first, and refuse
to add any files from the underlaydir if a file also exists in the srcdir
with the same name. **But**, note that this assumes that any given page can
be produced from a file with only one name (`page.mdwn` => `page.html`).

If it's possible for files with different names to produce a given page, it
would still be possible to use this attack to confuse ikiwiki into
rendering the wrong thing. This is not currently possible, but must be kept
in mind in the future when for example adding support for generating html
pages from source with some other extension.

## XSS attacks in page content

ikiwiki supports [[HtmlSanitistion]], though it can be turned off.