If you have the ability to set environment variables then it's basically already game over, with or without the existence of the antigravity module.

You could set PATH to change which files are executed in certain scenarios. You could set SSLKEYLOGFILE which logs session keys to an arbitrary file, essentially nullifying TLS/SSL protections. On Linux you can just set PROMPT_COMMAND to whatever you want and it'll be executed any time a bash prompt is printed.

It's an interesting attack vector, but a vulnerability requires impact, and I'm not sure this has very much.

Related to the idea "innocuous at first glance", that time when folks got cute with the man command:

https://git.savannah.nongnu.org/cgit/man-db.git/commit/src/m...

Turns out it was breaking a users automated tests - https://unix.stackexchange.com/questions/405783/why-does-man...

I'm unconvinced. "The ability to turn this into arbitrary code execution depends on what other executables are available on the system" is doing a LOT of heavy lifting here.

Remember: you have control over only the environment variables, and you do not have the ability to alter the arguments. In order for this to represent arbitrary code execution, you need for the system to have an executable on it that, when executed with the argument "https://xkcd.com/353/", grants you arbitrary code execution.

So, you have full control over the environment variables, and that's it. How do you turn effectively

    [binary] "https://xkcd.com/353/"

into arbitrary code execution, where [binary] is an executable already on the machine?

There is a proof of concept given, "Figure-4: arbitrary code execution achieved using multiple environment variables against Python 2 and Python 3":

    $ docker run -e 'PYTHONWARNINGS=all:0:antigravity.x:0:0' -e 'BROWSER=perlthanks' -e 'PERL5OPT=-Mbase;print(`id`);exit;' python:2.7.18 python /dev/null
    uid=0(root) gid=0(root) groups=0(root)
    Invalid -W option ignored: unknown warning category: 'antigravity.x'

    $ docker run -e 'PYTHONWARNINGS=all:0:antigravity.x:0:0' -e 'BROWSER=perlthanks' -e 'PERL5OPT=-Mbase;print(`id`);exit;' python:3.8.2 python /dev/null
    uid=0(root) gid=0(root) groups=0(root)
    Invalid -W option ignored: unknown warning category: 'antigravity.x'

If you can influence arbitrary environment variables, you usually have a lot more options - you could set PYTHONHOME or PYTHONPATH (influence where Python finds imports), LD_LIBRARY_PATH (influence where shared libraries come from), PATH (influence where the Python command itself comes from), etc. etc.

In this particular case, they happened to be in a situation where they couldn't easily figure out how to create new files but they could set environment variables. From a design perspective, as a reviewer, I would not believe that this makes it safe to have attacker-controlled environment variables, although I admit I wouldn't know exactly how (and this post is a pretty clever approach to making the attack work).

And specifically, although it is nowadays uncommon for Python scripts to be run the way old-school WSGI web services were, with a script being fired up for every request and mapping user-controlled query parameters to environment variables, this was at one time actually the motivating example for https://www.python.org/dev/peps/pep-0333/ ! It's a common enough attack pattern that https://capec.mitre.org/data/definitions/77.html exists and specifically cautions applications against trusting environment variables. There are very few places in Python where upon importing a module, an environment variable is trusted and triggered; all of them should be seen as security holes.

Surely this is only a security risk if antigravity is imported?

What cause would there be to import antigravity in a CGI service – or anything other than a terminal in which you could already execute arbitrary code?

I am struggling to understand, even with the proof of concept, how the situation could arise where an attacker could realistically exploit this, based largely on the uselessness of the antigravity module.

See the "Making Progress With PYTHONWARNINGS" section in the article posted - setting the environment variable PYTHONWARNINGS to a value mentioning the antigravity module causes it to be imported, even if there's no "import antigravity" statement in the actual source.

Not sure how I missed that – it was even in the proof of concept in the message I replied to. Thanks for the help.

There are multiple failings that don’t seem like a big deal taken in isolation:

- The PYTHONWARNINGS and whole Warning Filters system.

- The webbrowser module executing a given executable from an environment variable.

- The antigravity module running webbrowser.open() on import.

All of these sort of seem a bit like sacrificing good taste for convenience. But in combination, in a situation where you don’t control the environment variables, they do lead to arbitrary code execution.

In other words, it’s not all antigravity’s fault, but those side effects on import make this possible.

You're right about sacrificing good taste for convenience, but since this is an easter egg, it's sacrificing good taste for a laugh, and I think that pushes it over the precipe over into being a bad idea.

I left one window open, but what is to guarantee a burgler would use that window??

Absolutely, I agree now that I see it, and these seem like numerous innocent-seeming footguns just waiting for abuse.

Add the not-too-far-fetched-seeming assumption that it’s fine to let the Internet define your environment variables, and it’s a plausible exploit.

This is why, as another commenter points out, defense in depth is so important.

TL;DR: 'import antigravity' is fine, but Python can be tricked with environment variables to use Perl as the web browser, which has arbitrary code execution through environment variables.

And Perl (specifically "perlthanks") was just the first thing they found that worked. Bash had similar problems in the past, and I imagine there's other "installed by default" stuff that can be coaxed into running code in environment variables.

Defense in depth would disagree that it is fine. I'm pretty sure this vulnerability can go much further if you combine it with, for example, the `https_proxy` environment variable.

Interesting, that's pretty cool. Thanks!

The next section of the link literally goes into how to achieve all of that using a call to perl.