0

Running a wargame and sleeping well at night

 8 months ago
source link: https://dustri.org/b/running-a-wargame-and-sleeping-well-at-night.html
Go to the source link to view the article. You can view the picture content, updated content and better typesetting reading experience. If the link is broken, please click the button below to view the snapshot at that time.

Running a wargame and sleeping well at night

In May 2016, Mantis and I started websec.fr, with the help of blotus, cutz and nurfed, as previously announced and detailed here.

Since for most of the levels, skilled players are able to gain arbitrary PHP code execution, sandboxing and isolation are in order, in ways as simple as possible: our free time is limited, and we'd rather spend it writing new challenges or doing interesting things instead of maintaining custom brittle contraptions and debugging weird snowflake issues. Most of the infrastructure is documented in various text files in /root: how to deploy a level, how to access the backups, along with working exploits for all the levels, so onboarding new admins is easy.

At the beginning, we were using grsecurity, to take advantage of RBAC, TPE and PAX_MPROTECT to prevent players from introducing new native code and execute it. But with grsecurity going dark, we had to spend some time trying to ghetto-replace™ the features we used.

Every challenge is running under an unique user, levelXX, via php-fpm, in an empty read-only chroot. Unfortunately, for some levels, having a session is required, and the simplest way to store them is to write them down in files, on the filesystem, that has thus to be writeable. Those levels are using a bind mount with MS_NOEXEC, to prevent the execution of new binaries in the chroot.

There are also crontabs regularly deleting session/uploaded files, and everything else that could be lying there. It's way simpler than trying to make use of quotas. PHP process are short-lived and not re-used a lot; along with some disable_functions/disable_classes common sense, this prevents a lot of dumb resource-hungry mistakes.

We used to have monit solving the challenges every hour, to make sure that everything worked, but in the end, it was brittle for challenges requiring several requests, and completely useless for the rest since nothing really broke because PHP takes retro-compatibility (too) seriously and levels tend to be pretty immutable.

Since PHP is a shitshow prone to local roots and other assorted happy little accidents, it's running via systemd, taking advantage of the sandboxing features, because unfortunately, running php-fpm with the master process under a non-root user is equally painful and fragile, so we pass USE_ZEND_ALLOC=0 in the environment to disable PHP's interesting custom memory management subsystem, and put a custom allocator like mimalloc in /etc/ld.so.conf. Amusingly, it's not really possible to use something like isoalloc or hardened_malloc, since they tend to crash PHP because of latent memory corruptions, at least last time we tried. But this doesn't really matter, since the goal here is only to prevent script-kiddies from copy-pasting PHP exploits to get some flags or try to mess with the master process. Unix sockets are used to communicate with php workers, preventing the usage of one level to talk to others without going through the reverse proxy.

Nothing fancy about it by the way: it's a simple nginx, with some regex-fu to factorise the configuration instead of copy-pasting a new block every time a new level is added. There is an administration interface at a secret URL looking like https://websec.fr/fd9dcc7f049aaae76bd277955eb585554f840…, with an equality random password, to publish level, manage writeups, … There are also some mitigations in place to detect, annoy and confuse automatic tooling like sqlmap and its friends.

Logging it into the box is done via ssh, running on the standard port, as root. The only hardening here is that the ciphers/KEX/HMAC/Key/… are tweaked to only allow modern ones: this keeps a ton of bots at bays.

The scoreboard is also running in a chroot, under an unique user, sandboxed. Since we allow and encourage players to publish writeups, we're processing them as markdown, then parse the resulting html using html5lib, and use an allow-list approach for tags, which should™ prevent XSS.

Upgrades are automatically installed via a crontab and the machine is ~automatically rebooted when a new kernel is installed. To prevent copy-pasted kernel exploits, LKRG is used, because ~nobody bothers with bypassing it.

The levels are backuped in a git repository, along with the scoreboard and website, on bitbucket, so should the box go down in a gigantic fire, ~nothing should be lost. Because of laziness, the wargame is running on a DigitalOcean droplet: it's cheap, reliable, simple, …

Everything has been running smoothly for the last 6 years, I can't remember the last time there was a significant issue infrastructure-wise. The next big thing is going to be the migration of the PHP5 levels to PHP7 when possible, or their archival, but this is a problem for future me.


About Joyk


Aggregate valuable and interesting links.
Joyk means Joy of geeK