Sandbox

Scripts are generally considered to be untrusted and executed exclusively in a child process. It’s important to note that there’s a basic sandbox that’s active on every operating system, and there’s a second line of defense on supported operating systems.

The first line of defense is the restrictive stdlib. It’s assumed that and attacker gains full control over the lua code and is able to call any function with arbitrary arguments. The stdlib only provides functions that are considered safe, so for example it’s not possible to start a process or open a file.

The second line of defense is supposed to make sure the system isn’t compromised even if the first layer is fully broken and an attacker gains full control over the child process.

Right now this is only supported on linux and openbsd.

Linux

On linux we use seccomp to filter all syscalls that we don’t need. We also use chroot to disable filesystem access. It’s recommended to install the sn0int binary with cap_sys_chroot to make sure unprivileged users can use chroot. The chroot location is hard coded and all capabilities are removed after the chroot is done or if no chroot is going to happen.

OpenBSD

On openbsd we’re using pledge to restrict syscalls and unveil to restrict filesystem access.

IPC Protocol

The parent process and the child process communicate using an IPC protocol that is line-based json.

For a simple hello world the parent process is only going to send a single line to the child process. This line contains:

  • The function argument
  • The dns config
  • Keys that the module has been given access to
  • The module metadata and code
  • Options, if any
  • A socks5 proxy, if any
  • The log level
{"arg":null,"dns_config":{"ns":["1.1.1.1:53","1.0.0.1:53"],"tcp":false,"timeout":{"nanos":0,"secs":3}},"keyring":[],"module":{"author":"anonymous","description":"basic selftest","keyring_access":[],"name":"selftest","script":{"code":"-- Description: basic selftest\n-- Version: 0.1.0\n-- License: GPL-3.0\n\nfunction run()\n    -- nothing to do here\nend\n"},"source":null,"version":"0.1.0"},"options":{},"proxy":null,"verbose":2}

Saving this line in a file called start.json and sending it to a sandbox process should result in the following output:

$ sn0int sandbox foobar < start.json
{"Exit":"Ok"}
$

This line tells us that the script terminated successfully.

There are some functions that cause a notification to the parent process. We are going to add a call to the info() function to our module:

{"arg":null,"dns_config":{"ns":["1.1.1.1:53","1.0.0.1:53"],"tcp":false,"timeout":{"nanos":0,"secs":3}},"keyring":[],"module":{"author":"anonymous","description":"basic selftest","keyring_access":[],"name":"selftest","script":{"code":"-- Description: basic selftest\n-- Version: 0.1.0\n-- License: GPL-3.0\n\nfunction run()\n    info('ohai')\nend\n"},"source":null,"version":"0.1.0"},"options":{},"proxy":null,"verbose":2}

This is going to print an additional event:

$ sn0int sandbox foobar < start2.json
{"Log":{"Info":"\"ohai\""}}
{"Exit":"Ok"}
$

There are some functions that block the child process until the parent process sent a reply. These functions are mostly database related functions, since the child doesn’t have direct database access. To demonstrate this, we’re going to write two lines to our file this time, one is the init line and the second one is the reply for the database event:

{"arg":null,"dns_config":{"ns":["1.1.1.1:53","1.0.0.1:53"],"tcp":false,"timeout":{"nanos":0,"secs":3}},"keyring":[],"module":{"author":"anonymous","description":"basic selftest","keyring_access":[],"name":"selftest","script":{"code":"-- Description: basic selftest\n-- Version: 0.1.0\n-- License: GPL-3.0\n\nfunction run()\n    x = db_add('domain', {value=\"example.com\"})\n    info(x)\nend\n"},"source":null,"version":"0.1.0"},"options":{},"proxy":null,"verbose":2}
{"Ok":1337}

Results in the following output:

$ target/release/sn0int sandbox foobar < start3.json
{"Database":{"Insert":{"Domain":{"value":"example.com"}}}}
{"Log":{"Info":"1337.0"}}
{"Exit":"Ok"}
$

The first line is a database event and indicates that the child wants to insert data. After printing this line the child tries to read a line from stdin, this is why we needed to write two lines to our json file this time. In the second line the child learns if the insert was successful and which id was assigned to that entity.

Limitations

There are some limitations that you should be aware:

  • Network access is available and network namespaces aren’t isolated. This means scripts have access to your local network, the internet and also your localhost loopback interface.
  • If chroot is unavailable an attacker could connect to unix domain sockets.

Diagnosing a sandbox failure

You might experience a sandbox failure, especially on architectures that are less popular. This usually looks like this:

[sn0int][example][kpcyrd/ctlogs] > run
[-] Failed "example.com": EOF while parsing a value at line 1 column 0
[+] Finished kpcyrd/ctlogs (1 errors)

A module that never finishes could also mean an IO thread inside the worker got killed by the sandbox.

You can try to diagnose this yourself with strace:

strace -f sn0int run -vv ctlogs 2>&1 | tee strace.log

Open strace.log, look out for syscalls that didn’t return by searching for = ? and ignore calls to exit and similar. You are looking for something like this:

seccomp(SECCOMP_SET_MODE_FILTER, 0, {len=48, filter=0xdd59094e490}) = 0
write(1, "[+] activated!\n", 15[+] activated!
)        = 15
getresuid( <unfinished ...>)            = ?
+++ killed by SIGSYS (core dumped) +++

This would indicate a call to getresuid which was not allowed by the seccomp filter.

If you don’t want to diagnose this yourself open a new bug report with as much information as possible, specifically which distro, which release and which architecture you’re using.