Confusion Attacks: Exploiting Hidden Semantic Ambiguity in Apache HTTP Server

https://blog.orange.tw/2024/08/confusion-attacks-en.html

by tptacek • 2 years ago

120 16 2 years ago

blog.orange.tw

tptacek2• 2 years ago

These are pretty beautiful attacks, which build off what I think are two basic issues Apache deals with that other webservers don't:

* A maybe ill-advised architecture where lots of independently-developed modules cooperate to handle a request, all working on an over-complicated under-specified request "object".

* The requirement to honor `httpd.conf` file structures dating back to the 1990s, which has led to legacy support code paths that do surprising things.

The core of this article is a two-step attack that exploits (1) the fact that modules alternately interpret `request->filename` as a filename and as a URL, which allows attackers to truncate URLs at path separator characters, which by itself can slip you past some basic ACLs, and (2) the fact that Apache will try to serve request both with and without the `DocumentRoot` applied, which, combined with the first attack pattern, means that config directives that have attacker-controlled targets give attackers full filesystem access. Hilarity ensues.

There's a really slick attack here that takes you from CRLF-injection to SSRF and RCE, which is neat, because CRLF-injection ("header splitting") would ordinarily not be that interesting of an attack.

This is great stuff that just makes me wish more people ran Apache in 2024.

Vecr• 2 years ago

Why would you want more people to run Apache? If you're running it you should cover it up with some sort of carefully configured reverse proxy.

tptacek2• 2 years ago

So that I could play with bugs like these!

wepple• 2 years ago

“Carefully” is very load bearing.

A similar but different class of issues common to web stacks is when you have proxies and multiple layers of HTTP “things”, especially when they parse data differently (eg: nginx fronting Apache)

chrsig• 2 years ago

Because I wasn't hurting enough without having to worry about parser alignment.

kindawinda• 2 years ago

Are we reading 'The Mind of a Hacker'?

tommiegannert• 2 years ago

So my public web servers will start seeing a lot more spam requests that contain %0D%0A or end in %3F?

Really impressive to find that many exploitable issues without once relying on a buffer overflow.

xg15• 2 years ago

The lesson I've learned from this: Ambiguity may not always resolve. If some technical details in a spec seem confusing or ambiguous between two contradictory options, then half of the implementations will probably do option #1, the other will half will do option #2 and some may do both at the same time.

flemhans• 2 years ago

The AddHandler stuff I've seen exploited or attempted in the last 10 years

1oooqooq• 2 years ago

do they all depends on rewrite engine (.*) $1, or i missed one that doesn't?

idoubtit2• 2 years ago

Some attacks did not rely on mod_rewrite. For instance, the splitting of response headers.

The Apache vulnerabilities list provides a short summary of each: https://httpd.apache.org/security/vulnerabilities_24.html The fixes are in versions 2.4.59 and 2.4.60. Then a fix of the fix of CVE-2024-39884 in 2.4.61. Then a better fix of CVE-2024-39884 in 2.4.62.

yazzku• 2 years ago

Also slides: https://i.blackhat.com/BH-US-24/Presentations/US24-Orange-Co...

KronisLV2• 2 years ago

In big letters, near the beginning:

Why Apache HTTP Server Smells Bad?

In the end:

The entire development process must be carefully built upon a pile of legacy code and technical debt. So, if any Apache HTTP Server developers are reading this: Thank you for your hard work and contributions!

Quite the tone shift, a bit amusing, but definitely an attention catching choice and immensely cool that people are doing research like this.

I'm fairly sure that most large systems have a bunch of potential vulnerabilities, for example you can most likely do more with Apache than you can with Nginx due to its dynamic nature and module system, but at the same time that comes with dangers like this.

idoubtit2• 2 years ago

you can most likely do more with Apache than you can with Nginx

Apache having more features than Nginx is debatable. Apache has probably more unofficial modules, but that's irrelevant to the core developpers. With the official modules, I suppose Apache and Nginx more or less on par, but it's hard to compare. Apache has many useless modules in its distribution, e.g. "cern_meta" or "dialup" (for 56K modems!). And some modern features are lacking in Apache: e.g. HTTP3 was an experimental feature of Nginx 4 years ago, it's stable for more than one year, and AFAIK it's not even planned in Apache.

due to its dynamic nature and module system,

Nginx has a module system, like Apache. E.g. "ngx_echo" is an unofficial module.

but at the same time that comes with dangers like this.

As you noted, the post explains that the problem was not the features surface, it was the legacy code and the technical debt. Nginx is better on these aspects, and Caddy probably even more so.

troyvit• 2 years ago

Apache has probably more unofficial modules, but that's irrelevant to the core developpers.

In a way it's irrelevant, but in a way not. The myth was that they called it Apache because it's "A patchy" web server.[1] IOW it was designed to use modules to do most of its work.

[1] https://dan.hersam.com/2004/07/29/a-patchy-web-server/

KronisLV2• 2 years ago

that's irrelevant to the core developers

That's relevant to me! If I want Apache to act as an OpenID connect relying party, for example, I can just use mod_auth_openidc https://github.com/OpenIDC/mod_auth_openidc installing which is pretty much trivial and requires little work on my part:

  sudo apt install libapache2-mod-auth-openidc

If I want something similar for Nginx, I can take a look here https://openid.net/developers/certified-openid-connect-imple... and realize that I'll need to use OpenResty with LuaRocks, which does add a bit more complexity: https://github.com/zmartzone/lua-resty-openidc

   luarocks install lua-resty-openidc

If I want a basic WAF with something that runs ModSecurity rules, on Apache I can just do:

  libapache2-mod-security2

and can use the OWASP CRS with no issues: https://owasp.org/www-project-modsecurity-core-rule-set/

With Nginx that is no longer properly supported and needed their Plus version anyways: https://docs.nginx.com/nginx-waf/admin-guide/nginx-plus-mods...

  sudo apt-get install nginx-plus-module-modsecurity

If I want to do that myself with the regular Nginx, I suddenly have to compile a bunch of stuff myself, which becomes far less pleasant and a bit more long winded: https://www.linode.com/docs/guides/securing-nginx-with-modse...

  sudo apt-get install bison build-essential ca-certificates curl dh-autoreconf doxygen \
  flex gawk git iputils-ping libcurl4-gnutls-dev libexpat1-dev libgeoip-dev liblmdb-dev \
  libpcre3-dev libpcre++-dev libssl-dev libtool libxml2 libxml2-dev libyajl-dev locales \
  lua5.3-dev pkg-config wget zlib1g-dev zlibc libxslt libgd-dev
  sudo apt install git
  cd /opt && sudo git clone https://github.com/SpiderLabs/ModSecurity
  cd ModSecurity
  sudo git submodule init
  sudo git submodule update
  sudo ./build.sh
  sudo ./configure
  sudo make
  sudo make install
  cd /opt && sudo git clone --depth 1 https://github.com/SpiderLabs/ModSecurity-nginx.git
  cd /opt && sudo wget http://nginx.org/download/nginx-1.14.0.tar.gz
  sudo tar -xvzmf nginx-1.14.0.tar.gz
  cd nginx-1.14.0
  sudo ./configure --add-dynamic-module=../ModSecurity-nginx <Configure Arguments>
  sudo make modules
  sudo mkdir /etc/nginx/modules
  sudo cp objs/ngx_http_modsecurity_module.so /etc/nginx/modules

(old versions in the example, but same principle)

At that point I might as well look in the direction of OWASP Coraza: https://coraza.io/

When it comes to ACME and Let's Encrypt, I can use mod_md out of the box with Apache: https://httpd.apache.org/docs/2.4/mod/mod_md.html

It's experimental but works nicely in my experience, whereas with Nginx, I have to go with Certbot separately: https://certbot.eff.org/

If I want everything in a single package, I can just use Caddy instead (which is a lovely web server).

And that's before I get into weird Nginx behavior when I try to use it with containers, where the whole web server crashes, just because it cannot find the DNS record for a container that hasn't been created yet, along the lines of:

  nginx: [emerg] host not found in upstream "my-app"

where trying to use the suggested fix leads to:

  nginx: [emerg] "proxy_redirect default" cannot be used with "proxy_pass" directive with variables

and using other values often breaks app redirects.

Due to these experiences, I'll absolutely say that you can do more with Apache, in ways that will be less of a pain, except for the whole "getting your entire web server hacked" or "brought down by any sizeable amount of traffic" bit. Nginx is really good for hosting static content or acting as a basic reverse proxy, but will be harder to get various bits of the greater ecosystem into it, as opposed to how Apache works.

Then again, I fully acknowledge that someone will come along and say "ModSecurity isn't very useful, you should just put your site behind Cloudflare if you need a WAF." or maybe even "Don't handle OpenID Connect in your ingress, even if that means that in the case of 5 APIs you'll need OIDC libraries and configuration for each of them, which will look quite different in various languages." but that doesn't change the fact that some things are just more doable in one option instead of another.

OpenResty or Caddy increasingly feel like better choices for a lot of stuff, but not bare Nginx. Use Nginx when you want to handle simple use cases (reverse proxy, serving static files) with good performance, security and a sane configuration format.

Confusion Attacks: Exploiting Hidden Semantic Ambiguity in Apache HTTP Server

Related Stories