Tuesday, August 30, 2011

On exploits and assessing security

Everything started on a summer Friday, August 19 2011, 23:23 BST (so it might be reported to Sat 20 Aug, depending on your timezone).
Kingcope, a Greek hacker known for a long stream of 0days and trash Greek music soundtracks, posts a very short email on the Full-Disclosure mailing list. The email has a promising title, Apache Killer, and an attachment: killapache.pl.

The script is extremely simple: 84 lines of perl, comments included, capable of killing any Apache web server.

The core of the attack is extremely simple: the script sends a (perfectly legit) request to the Apache server, including a Range header similar to this:

Range: bytes=0-,5-0,5-1,5-2,5-3,5-4,5-5,5-6,5-7,5-8,5-9,5-10,5-11,5-12,5-13,5-14,5-15,5-16,5-17,5-18,5-19, 5-20,5-21,5-22,5-23,5-24,5-25,5-26,5-27,5-28,5-29, -[..and many many more..].

What happens behind the scene is that Apache has to make separate copies of the response for each of those ranges and, guess what, it explodes.

The Range header (defined here in RFC 2616) has legitimate uses: for instance, PDF readers use it while you scroll down the page. Let us reiterate that the attack is a perfectly well formed request: "A byte range operation MAY specify a single range of bytes, or a set of ranges within a single entity."
On a side note, IIS only allows up to 5 ranges for each HTTP request, and it is not vulnerable - with the notable exception of IIS 5, but you don't have those around anymore, do you?

But back to our story. Apache reacted "quickly", issuing a security advisory [CVE-2011-3192] a mere 4 days after the release, on August 24 16:16 GMT [http://article.gmane.org/gmane.comp.apache.announce/58]. They acknowledged the bug, promising that "A full fix is expected in the next 48 hours.". Apache also provided a number of different mitigations, by using mod_rewrite, limiting the size of the request field or even a dedicated module.

+1 for having included so many mitigations inside the advisory, -1 for taking so long to give an official answer.

A number of different answers were surfacing by that time, including Spiderlab's suggestion of using ModSecurity (but truth be told the answer was already floating on some forums). The 24th was, in short, the moment where the general public started noticing the issue - also thanks to a The Register's article which, even though it drops a totally out of place comment on open source security, at least cites Zalewski.

And this is the interesting bit of the story: the bug was well known!
lcamtuf reported this exact issue in 2007 - just 4 years ago.

The reactions were, to say the least, wrong. William A. Rowe [of the Apache Software Foundation] answered: "With the host of real issues out there in terms of massively parallel DDoS infrastructures that abound, this is, as you say, quite a silly report.". Yeah, sure. The issue was also thought to be connected to the sliding window (Rowe: On the matter of your 1GB window (which is, again, the real issue), you have any examples of a kernel that permits that large a sliding window buffer by default).
Probably even lcamtuf did not understand the full impact of this issue - we are speaking about a full DOS against the default installation after all, even if it is very easily avoided - as he says: "William, again, this is not a critical issue; I did mention that, and if it were, I wouldn't report it that way". The issue was thus not really understood by both parties, or at least its full implications were not. If we go into the details of the bug report we must grant the fact that it focused more on the "network" attack vector - not exactly the one exploited by Kingcope's attack, which results in memory exhaustion. Siim Põder also tried to implement the attack, but failed. The thread then just died.

Bottom line: general failure of the community to identify and evaluate a risk in the absence of an exploit. So long for all the "risk assessment without full disclosure" talks: without an exploit, an entire community looking at the description of a bug in the most widespread web server of the world failed to evaluate risk.

At the same time, in the Apache dev@ mailing list there was not a single mention of the bug. Compare that to the mail thread [mod_deflate & range requests) from last August: http://mail-archives.apache.org/mod_mbox/httpd-dev/201108.mbox/browser, or here. This exploit, in the end, resulted in a ticket on the ietf with a request for changing the standards [http://trac.tools.ietf.org/wg/httpbis/trac/ticket/311]: that's what we call impact.

And all because an exploit was not developed in the first place...

3 comments: