Cross Site Request Forgery

In June of 2001, Peter Watkins defined the term Cross Site Request Forgery – pronounced Sea Surf. He keeps that discussion here: http://www.tux.org/~peterw/csrf.txt

I’d posted a copy of this text localy on my site and now I’ve now found I have a number of people linking to it.

So, I thought I’d turn it into an object lesson demonstration.

If you got to this link by clicking on a link to http://www.TheCodeCave.com/csrf.txt, you may be surprised to noticed that you are not looking a text file. That’s because I’ve intercepted your request and sent it to another location on my site. This is what a CSRF attack does however it bounces the attack back at you. I could have just as easily detected if you were an admin in any of the most popular open source projects out there, sent you to your site with an attack tailored to your software and then without taking a breath, put you into the text file again. Kinda scary isn’t it?

The trick is to address the danger by making sure that all of your web pages are secure. I’ve been planning for a long time to write a series of post describing what I’ve learned about PHP security. I just haven’t figured out a way to do it without creating a tutorial site. If you check back here: http://www.thecodecave.com/?cat=7 periodically, you can see what I’ve come up with.

From: Peter W <peterw@usa.net>
To: John Percival <john@jelsoft.com>
Cc: bugtraq@securityfocus.com, clambert@whitecrown.net, peterw@
tux.org
Subject: Cross-Site Request Forgeries (Re: The Dangers of Allowing 
Users to Post Images)
Message-ID: <20010615011542.C22677@usa.net>
References: <04f901c0f437$4911b610$9701a8c0@wellingtoncollege.
berks.sch.uk>
In-Reply-To: <04f901c0f437$4911b610$9701a8c0@wellingtoncollege.
berks.sch.uk>; from john@jelsoft.com on Wed, Jun 13, 2001 at 07:33:
04PM +0100

	Cross-Site Request Forgeries
		(CSRF, pronounced "sea surf")

I hope you don't mind if I expand on this a bit. You've come across 
the tip, in my opinion, of a rather large iceberg. It's another
Web/trust-relationship problem. Many Web applications are fairly 
good at identifying users and understanding requests, but terrible 
at verifying origins and intent.

The problem isn't the IMG tag on the message board, it's the 
backend app you seek to attack via the IMG tag. And I suspect lots 
of Web apps are vulnerable. Lots. I've been to training on highly-
regarded, widely-used, expensive Web app development frameworks, 
and none of the classes taught how to avoid the problems I will 
attempt to describe. In fact, they all seem to teach the "easy way" 
of handling what look like user requests, which is, of course, the 
vulnerable way. 

Anyway, let's look at how your post relates to what I call CSRF.

On Wed, Jun 13, 2001 at 07:33:04PM +0100, John Percival wrote:

> This exploit shows how almost any script that uses cookie 
> session/login data to validate CGI forms can be exploited if the 
> users can post images.  What is the problem? Well, by using an 
> [img] (or HTML <img> or <iframe> or <script src="">) tag, 
> the user is having anyone who views the thread access that image - 
> that is perform an HTTP GET on the URL specified for the image.
> Even if its not an image, it still can be accessed, but will 
> display a broken image. 

Depending on what's allowed, height/width and CSS/visibility tags 
can be used to hide the broken image icon.

> This means that the user can put a CGI script inside [img]
> tags.

** Learning from Randal's purple dinosaur?

The problem you describe is not uploading images, it's allowing 
users to post code that's inserted in an appropriate HTML tag 
attribute. This is something of a variation on Randal Schwartz's 
purple dinosaur hack,[2] but much more interesting and dangerous 
than even what you describe.

> This script will be called by whoever views that thread. 
> When used maliciously, it could force the user to: unknowingly 
> update their profile, respond to polls in a certain way, post new 
> messages or threads, email a user with whatever text they want, 
> the list goes on. This would be particularly worrying for a 'worm' 
> to spread through a forum, filling it with rubbish posts.

** The difference between XSS and CSRF

Right. There's something much larger going on here. Darnit, I 
wanted to make a nice formal paper out of this, but you're forcing 
my hand. :-) The problem is what I call CSRF (Cross-Site Request 
Forgeries, pronounced "sea surf"). Any time you can get a user to 
open an HTML document, you can use things like IMG tags to forge 
requests, with that user's credentials, to any Web site you want -- 
the one the HTML document is on, or any other.

This looks somewhat similar to Cross-Site Scripting (XSS), but is 
not the same. XSS aimed at inserting active code in an HTML 
document to either abuse client-side active scripting holes, or to 
send privileged information (e.g., authentication/session cookies) 
to a previously unknown evil collection site. 

CSRF does not in any way rely on client-side active scripting, and 
its aim is to take unwanted, unapproved actions on a site where 
the victim has some prior relationship and authority.

Where XSS sought to steal your online trading cookies so an attacker 
could manipulate your portfolio, CSRF seeks to use your cookies to 
force you to execute a trade without your knowledge or consent (or, 
in many cases, the attacker's knowledge, for that matter). [Just an 
extreme example there; I do not have any idea if any trading sites 
are vulnerable. I have not tested *any* applications or sites that I 
don't have some personal involvement in the design and maintenance 
of. Don't ask me to.]

<img src="https://trading.example.com/xfer?from=MSFT&to=RHAT
&confirm=Y">
<img src="https://books.example.com/clickbuy?book=ISBNhere
&quantity=100">

** Ubiquity of attack channels

Since HTML documents are popping up everywhere (even in 
corporate email systems!!!), and it's impossible to discern what IMG
or HREF values might be direct CSRF attacks, or redirect users to 
unwittingly do dangerous things via CSRF redirects, the fix has to 
be in the applications that do the interesting things.

> For example, if a user posted something along these lines:
> [img]http://your.forums/forums/newreply.cgi?action=newthread&
> subject=aaa&body=some+naughty+words&submit=go[/img]
> Then the post would go through, under the name of whoever 
> viewed the image.
> This is of particular danger when an administrator views an image, 
> which then calls a page in an online control panel - thus granting 
> the user access to the control panel.

** Impossible to filter content

Right, and as I say, the site you act against can be somewhere else 
entirely. Here's what a CSRF attack might look like:
 <img src="http://example.net/logo.gif" height=0 width=0 alt="">
That's it. When your client requests logo.gif - exposing no cookies 
- the example.net server redirects you to a URL like the one you 
show, above. So the end result us the same as if the attacker had 
embedded the more obvious URL inside the IMG tag. 

If an attacker wants, he can also use a simple, innocent looking 
hyperlink and hope the victim clicks on it (http://example.net/
kyotoanalysis.htm). You don't allow hyperlinks? Well, someone might 
copy/paste the link, and be stung that way. They'd notice? Maybe 
not -- the URL could be a mostly useful page, with a tiny frameset 
sliver that loads your attack URL.

> How can it be fixed? Well, there are a couple of ways to stop it, 
> but the easiest (in PHP at least) seems to be to have most of the 
> variables used by scripts be used through $HTTP_POST_VARS. So 
> instead of checking for $action in a script, $HTTP_POST_VARS
> ['action'] would be checked. This forces the user to use a POST 
> request, not a GET. 

which means the attacker reverts to using Javascript, or entices 
the victim to click on an image that's acting as a submit control 
in a <form>.  Requiring POST raises the bar, but doesn't really 
fix the problem.

> Alternatively, the sessionid could be required to come with the 
> GET/POST request variables, rather than by cookie.

...thereby exposing an important piece of authentication 
information to history files and proxy servers; I really don't like 
URL mangling for authentication purposes, especially in non-SSL 
systems. A combination of cookie + URL mangling might not be bad, 
though in the message board case, a CSRF attacker could use an 
intermediate redirect (as described earlier) to get the URL 
mangling (from the Referer), and redirect back to the messageboard 
with the proper mangling as well as all cookies that might be 
expected/needed. So in your example case, URL mangling would buy 
nothing. :-(

> Finally, in the specific case of [img] tags, the use of ? or & in 
> the img URL can be disabled by some regexes.

Not at all adequate. Browsers follow redirects on IMG tags, so I 
redirect you to http://example.net/logo.gif which in turn redirects 
you to the final URL, as described earlier.

> If the software that you run is not secure, we recommend that 
> you disable HTML and/or [img] tags, until the fixes have been 
> implemented.

It's much worse than that.

Please see the following URLs for an introduction to the dangers 
of CSRF, and some discussion of countermeasure strategies. 

 http://www.astray.com/pipermail/acmemail/2001-June/000803.html
 http://www.astray.com/pipermail/acmemail/2001-June/000808.html
 http://www.astray.com/pipermail/acmemail/2001-June/000804.html

** Server-Side Countermeasures

The fix MUST be implemented on the backend that's being attacked. 
In your example, newreply.cgi needs to be intelligent enough to 
detect and stop CSRF attacks. 

We've talked about how an attacker can post a message to the 
messageboard with innocent looking URLs. But an attacker can also 
simply send the victim a piece of HTML email including the full 
attack IMG URL. No amount of IMG tag filtering in your 
messageboard posting system can stop that.

** Three-phase tests before acting

When it comes to generic CSRF attacks, any application that 
uses a two-phase approach to action approval is vulnerable (the 
two phases being [1] do you possess authentication information 
and [2] are all the required arguments present). What's needed is 
a third test: is the user really using a proper application form to 
generate the request?

** The 90% solution: Referer tests

For many sites, you can achieve a high level of protection by 
checking the HTTP Referer header. This would prevent things like 
attacks via email. But it would also mean locking out any user whose 
requests did not contain Referer information.[1] As long as the 
values in the allowed Referer list are all coded with XSS and CSRF 
in mind, this could be adequate.  Referer checks should be as 
specific as possible, e.g. you might require the Referer to begin 
with: 
"https://example.com/admin/admin.cgi" or "https://example.com/
admin/" instead of simply "https://example.com/".

** The more difficult cases

Some other applications are more difficult to secure. Consider 
webmail apps. So webmail.example.com decides only "message 
delete" requests from webmail.example.com pages will be accepted: 
well, if the attacker sends a CSRF message to your webmail account, 
then when you read it via webmail, the Referer in the CSRF image 
request (your client thinks it's an image request) says it's indeed 
from the proper webmail server (even in the case of an intermediate 
redirect; check the bugtraq archives for past discussion of 
anonymizing hyperlinks, redirects vs. client-pull, etc.), so the 
request gets through. Basically, any application that allows posting 
of URLs needs more sophisticated protection than Referer checks. 
This would also include messageboards and discussion sites like 
Slashdot. 

> Known Vulnerable: Infopop's UBB 6.04e (probably the whole 6.xx 
> series), ezboard 6.2, WWW Threads PHP 5.4, vBulletin 2.0.0 
> Release Candidate 2 and before (later versions are safe). Probably 
> many more bulletin boards and CGI scripts out there, but those are 
> the main ones that we have been tested positive.

** One-time authorization codes

The URLs I list above outline a server-side one-use-token approach 
to closing the hole. For instance, the page that users are expected 
to use for drafting messages (in your newreply.cgi example) would 
create a one-time use token, good for a limited time. The newreply.
cgi processing script would require this value be present, correct, 
and in time. So while the attacker knows that action, subject, body, 
and authcode values are required, the attacker does not know, and 
cannot ascertain, the proper value needed for the authcode 
argument.[0] These tactics tend to introduce certain 
inconveniences (e.g., preventing use of the "back" button) so you  
may wish to analyze the various actions your application can take 
and provide varying levels of protection. For example, in a webmail 
system sending and deleting messages need more protection than 
displaying messages.

** Unpredicatable argument names?

Other tactics may be possible. For instance, consider
"action=newthread&subject=aaa&body=some+naughty+words&
submit=go". On the server side, you could have an "argument map 
table" for each session, e.g. pick random surrogates for the normal 
argument names. For one user, the system might look for 
"876575665" as an argument name instead of the predictable 
"action", "9876dafd987" for "body", etc. There may be some 
tricks vis-a-vis anonymizing referers if the labels are constant 
throughout a session, but it might be possible to do something
like this to make it more difficult to construct a valid URL for a 
CSRF malicious action.

** Attacking sites behind corporate firewalls

Want more fun? CSRF tactics can be used to attack servers 
behind corporate firewalls. It's not just your public Web apps that 
are at risk. 
<img src="http://intranet/admin/purgedatabase?rowslike=%2A&
confirm=yes">
If the attacker knows enough to make a URL and can get you to 
open a message, that's all it takes. Here we see that HTTP 
Referer headers can be a double-edged sword. Earlier we 
described how Referer tests can add security to many apps 
relatively easily. But Referer headers can also leak information 
about "private" sites if those sites use non-anonymized hyperlinks 
and external document references.

I'm afraid CSRF is going to be a mess to deal with in many cases. 
Like trying to tame the seas.

** Workarounds

Most of us probably depend on applications that won't be fixed
anytime soon. So what can you do to prevent a CSRF attack 
from making your browser request something without your 
approval?
 - Do not use an email client that renders HTML
 - Do not use a newsgroup client tied to your Web browser
 - Do not allow your browser to save usernames/passwords
 - Do not ask Web sites you care about to "remember" your login
 - Be sure to "log off" before and after using any authenticated
   Web site that's important to you [or your employer ;-)], 
   even if that means exiting your Web browser completely
 - Consider using something like Windows 2000's "Run As" 
   shortcut feature or my "runxas" shell script (available at the 
   tux.org URL listed below) to run a Web browser for casual 
   use.

My apologies for the somewhat rambling nature of this post; I may 
yet clean this up and put it in a proper paper, and do some real 
editing... but I hope even in this rough form it makes some sense, 
and helps folks design better, safer applications.

-Peter
http://www.tux.org/~peterw/

[0] Not unless the page that included the authcode is readable, 
e.g. if the composition page had XSS bugs that would facilitate 
construction of a URL for a CSRF attack.

[1] As discussed earlier (http://www.securityfocus.com/arch
ive/1/41653), client-pull pages usually result in no Referer 
information being sent by the client. So if your application 
allows a request with no Referer, an attacker need only direct 
the victim to an HTML document that uses a client-pull META 
tag to send the victim to the CSRF attack URL. This might be 
tricker to pull off, but remains feasible. So if you want to use 
Referer checks, you really ought to go all the way and deny 
every request that lacks a Referer header.

[2] http://www.stonehenge.com/merlyn/ [3]

[3] fellow cornfed users: the horror! footnotes referenced in 
reverse order!