I’m writing this because I’m hoping someone will respond, telling me that what I want already exists. I have a specific itch and my suspicion is that developing a whole programming language and runtime is the only way to scratch that itch.
Please tell me I’m wrong.
Dear Father Christmas…
If you’ve ever written a web service, you’ve almost certainly had situations where you’ve taken a bunch of bytes from a completely untrusted stranger and passed those bytes into a JSON parser. What’s more you’ll have done that without validating the bytes first.
Processing your inputs without sanitizing it first? Has Bobby Tables taught us nothing?
You can do this safely because that JSON parser will have been designed to be used in this manner and will be safe in the face of hostile inputs. If you did try feeding the bytes of an EXE file into a JSON parser, it’ll very quickly reject it complaining that “MZ” isn’t an opening brace and refuse to continue beyond that. The worst a hostile user could do is put rude messages inside the JSON strings.
{ "You": "A complete \uD83D\uDC18 head!" }Now take that idea and think about what if you did have a web service where completely unauthenticated users could use any request body they liked and your service would run that request body in a copy of Python as the program source code.
Hopefully, you’ve just now remarked that it would be a very bad idea, up there with Napoleon’s idea to make his brother the King of Spain. But that’s exactly what I want to do. I want to write a web service that accepts Python code from complete strangers and actually run that code.
(And also make my brother the King of Spain. He’d be great!)
At the gates of dawn
Some time in the early 90s, I had a game called “C Robots”.
This is a game where four tanks are in an arena, driving around and firing missiles at each other. But instead of humans controlling those tanks, each tank was controlled by a program written by the human player. The game controller would keep track of each tank and any missiles in flight, passing back control to each tank’s controller program to let it decide what its next move will be.
For 90s me, programming a robot appealed to me but the tank battle part did not appeal so much. I really wanted to make a robot to play other games that might not involve tanks. At the time, there were two games I enjoyed playing with school friends, Dots-and-Boxes and Rummy. I had an idea of what made good strategies for these specific games, so I thought building those strategies into code might make for a good intellectual exercise.
Decades passed and I built a simple game controller system which I (rather pompously) called “Tourk“. I had made a start on the controllers for a handful of games but I hadn’t gotten around to actually writing actual competitive players, only simple random ones that were good for testing. I imagined that before long, people would write their own players, send them in to me and I’d compile them all together. After I’d let it ran for a million games in a tournament I’d announce the winner.
If anyone had actually written a player and sent it in, my first step would have been to inspect the submitted code thoroughly. These would have been actual C programs and could have done anything a C program could do, including dropping viruses on my hard disk, so inspecting that code would have been very important. Looking back, I’m glad no-one actually did that.
But this was one thing C Robots got right, even if it wasn’t planned that way. Once it compiled the player’s C code, it would run that code in a restricted runtime. Your player code could never go outside its bounds because there’s no instructions in the C Robots runtime to do that. This meant that no-one could use this as an attack vector. (But don’t quote me on that. I’ve not actually audited the code.)
Will the runtime do it?
Could maybe the dot-net runtime or the Python runtime have the answer?
This was one of the first questions I asked on the (then) new Stack Overflow. The answer sent me to Microsoft’s page on “Code Access Security” and if you follow that link now, it says this feature is no longer supported.
Wondering more recently if Python might have an option to do what I wanted, I asked on Hacker News if there was a way to run Python in the way I wanted. There were a few comments but it didn’t get enough up-votes and disappeared fairly quickly. What little discussion we had was more to do with a side issue than the actual question I was asking.
I do feel that the answer might still be here. There’s quite possibly some flag on the runtime that will make any call to an extern function impossible. The Python runtime without the “os” package would seem to get 90% of the way there, but I don’t know enough about it to be certain enough that this won’t have left any holes open.
Sanitize Your inputs?
Maybe I should listen to Bobby Tables and sanitize my inputs before running them.
Keep the unrestricted runtime, but before we invoke it to run the potentially hostile code, scan it to check it won’t do any bad things.
Simple arithmetic in a loop? That’s fine.
Running a remote access trojan? No.
Once it has passed the test, you should be able to allow the code to run, confident it won’t do anything bad because you’ve already checked it won’t. This approach appeals to me because once that initial test has passed the code for non-hostility, we can allow the runtime to go at full speed.
The problem with this approach are all the edge cases and finding that line between simple arithmetic and remote-access-trojans. You need to allow enough for the actually-not-hostile code to do useful things, but not enough that a hostile user could exploit.
Joining strings together is fine but passing that string into eval is not.
Writing text to stdout is fine but writing into a network socket is not.
Finding that line is going to be difficult. The best approach would be to start with nothing-is-allowed, but when considering what to add, first investigate what would be possible by adding that facility to allowed list. Because it can be used for bad things, eval would never be on that allowed list.
If there’s a function with a million useful things it can do but one bad thing, that function must never be allowed.
Ask the Operating System?
I told a colleague about this post while I was still writing it and he mentioned that operating systems can have restrictions placed on programs it runs. He showed me his Mac and there was a utility that listed all the apps he was running and all the permissions it had. It reminded me that my Android phone does something similar. If any apps wants to interact with anything outside its realm, it has to ask first. This is why I’m happy to install apps on my Android phone but not on my Windows laptop.
This would be great, but how do I, a numpty developer, harness this power? What do I do if I want to launch a process (such as the Python runtime) but with all the permissions turned off? It feels like this will be the solution but my searching isn’t coming up with a practical answer.
My hope is that there’s a code library whose job it is to launch processes in this super restricted mode. It’ll work out which OS it is running on, do the necessary magic OS calls and finally launch the process in that super-restricted mode.
Mmmm coffee!
The good people developing web browsers back in the 90s had the same need as me. They wanting to add a little interactivity to web pages, but without having to wait for a round trip back to the server over dialup, so they came up with a language they named JS.
As you read this page, your browser is running some code I supplied to you. That code can’t open up your files on your local device. If anyone did actually find a way to do that, the browser developers would call that a serious bug and push out an emergency update. So could JS be the solution I’m looking for?
As much as it sounds perfect, that JS runtime is inside the browser. If I have some JS code in my server process, how do I get that code into a browser process? Can I even run a web browser on a server without some sort of desktop environment?
The only project I know of where someone has taken JS outside of a browser is node-js. That might be the answer but I have written programs using node-js that load and save files. If this is the answer then I’d need to know how to configure the runtime to run the way I want.
Is there an answer?
I began this post expressing my suspicion that the solution is to write my own runtime, designed from first-principles to run in a default-deny mode. I still wonder if that’s the case. I hope someone will read this post and maybe comment with the unknown option on the Python runtime that does exactly what I want.
In the meantime, I have another post in the works as with my thoughts on how this runtime and programming language could work. I hope I can skip it.
Gronda-Gronda.
Picture Credits
📸 “Snow Scot” by Peeja. (With permission.)
📸 “Meeting a Robot” by my anonymous wife. (With permission)
📸 “Great Dane floppy ears” by Sheila Sund. (Creative Commons)
📸 “Fun with cling film” by Elizabeth Gomm. (Creative Commons)
📸 “Rutabaga Ball 2” by Terrence McNally. (Creative Commons)
📸 “Nice day for blowing the cobwebs off” by Jurassic Snark. (With permission.)
(And just in case advocating for your brother to be made King of Spain is treason or something, I don’t actually want to do that. It was a joke.)
