Mozilla has released security updates to Firefox 3.5 and 3.6 that include defenses for an old, little-known, but serious security hole: cross-site data theft using CSS. These defenses have a small but significant chance of breaking websites that rely on “quirks mode” rendering and use a server in another DNS domain (e.g. a CDN) for their style sheets.
In this article I’ll describe the attack, what we’re doing about it, how you can ensure that your site will continue to work, and how you can protect your users who have not upgraded their browsers yet.
The diagram above shows the course of a CSS data theft attack, from the perspective of a network hub that can see all the traffic. Before the attack begins, a victim user (the laptop in the middle) logs into their favorite website, Clockworks (mockup icon, on the right). Clockworks sends down a session cookie.
Some time later, while the victim is still logged into Clockworks, they click on an ad for dancing hamsters, and get sent to the attacker’s website (Badenov, on the left). The attacker’s website sends down an innocent-looking webpage that contains a
<link> tag whose URL points to the victim’s private-messages page at Clockworks.
The victim’s browser duly requests the private-messages page from Clockworks; since the victim is still logged in, it sends the session cookie, so the reply will include information meant only for the victim. The query string, chosen by the attacker, causes Clockworks’ server to inject strings into the HTML on either side of an interesting piece of secret information.
Because the attacker’s website is being rendered in quirks mode, the victim’s browser ignores the
Content-Type header and feeds HTML to the CSS parser. Of course, the very first HTML tag in the file causes a CSS syntax error, but CSS has predictable, lenient rules for recovering from syntax errors. The attacker’s injected strings make the CSS parser ignore most of the target page, and capture the secret as the value of the CSS
Finally, since the
background property applies to the
body tag, the browser needs to download the image it specifies in order to render the attacker’s website. The image URL has been wrapped around the secret information that the attacker wants. So the browser sends that secret to the attacker’s server as a query string.
This attack has been known for some time. The earliest public description I have found was by GreyMagic Corporation in 2002. It has been rediscovered at least twice since then: by Matan Gillon in 2005, and by pseudonymous blogger ‘ofk’ in 2008 (article in Japanese). There are many variations, some of which no longer work, and some of which only work in IE. The variation I’ve described works everywhere that hasn’t deployed a defense; security researchers at CMU were able to use this attack to steal the contents of private messages from a bulletin board and two different webmail providers, with victims using unpatched versions of all the popular browsers.
Since the attack relies on the CSS parser’s error recovery behavior, sites may be immune because of accidental properties of their page structure. For instance, most browsers do not allow newlines in
url() literals. If there had been a newline in the middle of the secret information in the diagram, just because that’s the way Clockworks generates its HTML, the attack would only work against victims using IE.
This attack works because a webpage in quirks mode can load anything as a style sheet, even if it’s really a HTML page coming from someone else’s server. If the attacker’s page were in standards mode, the browser would pay attention to the HTTP
Content-Type header on the target page, declaring it not to be CSS, and refuse to load it as a style sheet.
The attacker, of course, controls whether their page is in quirks mode. But the attacker’s page is on a different server than the target page, which means the attack can be blocked by an extension of the same-origin policy. Even if a page is in quirks mode, it’s not allowed to load a style sheet with a
Content-Type header declaring it to be something other than CSS, unless that sheet comes from the same origin. Firefox 4 will implement this rule, and IE has also adopted it (see below).
Unfortunately, there are a few websites out there that are rendered in quirks mode, and load their style sheets from a different origin, and put a
Content-Type header on those style sheets that says they’re not CSS. These sites aren’t common—the aforementioned CMU security researchers found 62 in the Alexa top 100,000, and most of those have been fixed already—but Firefox 4 will break them.
To give folks more time to fix their sites, while blocking the attack as soon as possible, we implemented a more lenient rule in Firefox ≥3.5.11 and ≥3.6.7. If a page is in quirks mode, loads a style sheet cross-origin, and that sheet has the wrong
Content-Type, we’ll start parsing it as CSS anyway … but we’ll stop and throw the sheet away if we encounter a syntax error before the first complete rule has been parsed. HTML tags cause CSS syntax errors, so unless the attacker can inject text at the very beginning of a page, they won’t be able to make the attack work. Safari, Google Chrome, and Opera have also adopted this rule.
It’s possible that this rule could break sites too. For instance, if a style sheet begins with an
@-rule that Firefox 3.5 does not understand, that will count as a syntax error, and the sheet will be discarded.
Fixing your website
You only have to worry about your site being broken by the defense if you load your style sheets from a different server than the HTML and you use quirks mode. If your site works with a Firefox 4 beta, you’re fine. Current versions of Firefox 3.5 and 3.6 will warn you in the error console when they see a site that will break in Firefox 4, so you can also test that way. (Unfortunately, due to limitations of our translation process, part of this warning will always be in English.)
If your site breaks, all you have to do to fix it is make sure that your style sheets are being served with
Content-Type: text/css in the HTTP headers. Please also consider switching to standards-mode rendering. If you cannot fix your website, we want to hear from you.
Protecting your users
If a browser tries to load a style sheet, and the HTTP response it gets has no
Content-Type header, it will just assume that it has been sent some CSS, even if it’s a cross-origin load. Therefore, your users are not fully safe from the attack, even if they all have browsers with the defense, unless your servers put Content-Type headers on all content requiring authentication. Check your web services as well as human-readable content.
You should also make sure that those headers are correct. Most importantly, ensure that if your server can’t figure out what Content-Type to put on a response, it falls back to
text/plain. Certain other possibilities (for instance,
application/x-unknown-content-type) may be treated the same as if you hadn’t sent a
Content-Type at all.
It is also vital to provide an accurate
charset= option in your
Content-Type headers for all textual data. If you don’t, an attacker can bypass your XSS filters by encoding injected strings in UTF-7. Declaring the charset in a
meta tag or
<?xml...?> instruction is not enough to defend against a CSS data theft attack encoded in UTF-7; the CSS parser doesn’t pay any attention to them.
To protect users that are still using browsers that have no defense against CSS data theft, you should block this attack in your filters for user-submitted content. All you have to do is add
@ to the set of characters that get replaced with equivalent HTML entities (
@, respectively). If you can’t be sure that you are always producing
Content-Type headers with the correct
charset= option, you should also entity-encode
Chris Evans rediscovered this exploit in late 2009 and has been instrumental in getting it fixed. He has two other blog posts that go into more detail. Collin Jackson and his team at CMU have also been very helpful in understanding the full scope of the attack and ensuring all major browsers fixed it. Their paper will appear at the ACM Computer and Communications Security conference in October.