node-sandbox

About

node-sandbox is a way of running untrusted code outside of your application's node process. You can interface with code running in the sandbox via RPC (or any library that works over the node Stream API).

Note that at the moment, the only protection is that the code is run in a separate process, and that the process is prevented from loading certain C bindings used by core modules. In the future, there will be container plugins available to block off functionality to the process itself (eg FS access, ability to open sockets, etc.) Containment will be done on the OS-level to reduce the attack surface of the sandbox.

Note that the method of containment used by the sandbox is far from bulletproof, and there is a very real possibility that malicious code could break out of the sandbox. In the future we will be working on improving it, but it will never be perfect.

License

This library is Licensed under the Academic Free License version 2.1

Documentation

Using node-sandbox is pretty straightforward. The source code is fairly well documented, and there a good number of test cases, so if you have any questions, feel free to dive in!

Basic Usage

Note: by default, the code being run won't have access to require() or anything. See "Specifying Permissions" for more info.

//create a new sandbox instance w/ default options
var sb = new Sandbox("./path/to/code.js");
 
//expose a method for the sandbox to call
sb.rpc.expose("someMethod", function(arg){
    console.log(arg);
});
 
//run the sandbox
sb.run();
 
//Wait for the sandbox to initialize.
//We can't call methods until the sandbox is ready,
//otherwise we'll get an error!
sb.on("ready", function(){
    //call a method exposed by the sandbox
    sb.rpc.call("someMethod", ["someArg"]).then(function(result){
        console.log(result);
    });
});

Basic Options

There are some basic options you should know about for fringe cases.

If your node command isn't in your PATH, you need to specify it manually. By default, it just uses node.

If for some reason Sandbox isn't detecting the path to shovel.js, you can specify the path manually (it's in node-sandbox/lib/shovel.js)

var sb = new Sandbox("path/to/code.js", {
    
    //the node command used to spawn the child process
    node_command: "node"
 
    //the path to the shovel (what bootstraps the child process)
    shovel: path.join(__dirname, "shovel.js"),
 
});

Other options are discussed in the relevant sections!

Specifying Permissions

Specifying permissions is somewhat complicated. In node, there's a function called process.binding() that's used by node's built-in modules to get the relevant C bindings for things like I/O, crypto, and others. process.binding() works like require() does; you pass a module as an argument, and it fetches the relevant C binding.

node-sandbox's permissions system will block off loading of the C bindings. To allow loading of them, we need to pass a permissions option to it, specifying which C bindings are allowed to be loaded.

The following are the permissions specified by default, and are needed for the sandbox to run.

var sb = new Sandbox("./path/to/code.js", {
    permissions: ["tty_wrap", "pipe_wrap"]
});

If you want to allow use of specific modules (eg fs or crypto), look at their file in node's lib/ directory, and see what bindings they load by searching for process.binding. Then pass any modules needed through permissions array.

Note that if you want access to require(), you'll need to pass the following to permissions:

var sb = new Sandbox("./path/to/code.js", {
    permissions: ["tty_wrap", "pipe_wrap"] //TODO: figure these out
});

Exposing RPC Methods

node-sandbox comes with a stock RPC library that uses JSON-RPC in a bi-directional way, so that both ends can expose methods for the other to call. We can access it through Sandbox.rpc.

Exposing methods in the main process is easy:

var sb = new Sandbox(/* options etc */);
sb.run();
 
//we can expose individual methods, eg:
sb.rpc.expose("addOne", function(arg){
    return arg+1;
});
 
//we can create namespaces by including a '.'
sb.rpc.expose("myNamespace.someMethod", function(){
    return true;
});
 
//we can unexpose methods like this!
sb.rpc.unexpose("myNamespace.someMethod");
 
//alternatively, we can opt to expose an entire object.
//This deletes anything that was already exposed previously.
sb.rpc.exposeObject({
    addOne: function(arg){
        return arg+1;
    },
    myNamespace: {
        someMethod: function(){ return true; }
    }
});
 
//unexposing still works the same way when
//using exposeObject
sb.rpc.unexpose("myNamespace.someMethod");

If you need to work with asynchronous libraries in the methods you expose, you can return a Promise object instead. You can use any Promise library you like, but I use Kris Zyp's node-promise.

sb.rpc.expose("asyncMethod", function(){
    var p = new Promise();
 
    setTimeout(function(){
        p.callback("Success!");
        //or
        p.errback("Failure!");
    }, 2000);
 
    return p;
});

Exposing methods within the sandbox works the same way, except we use the global rpc variable instead of Sandbox.rpc. Inside the sandbox, Promise is accessible globally for convenience.

rpc.expose("myMethod", function(){
    var p = new Promise();
    // etc. etc.
    return p;
});

Calling RPC Methods

To call methods, you can use Sandbox.rpc.call, or Sandbox.rpc.notify. Both methods will call the remote method, but call will give you a return value, while notify won't (see the JSON-RPC docs if this is confusing).

It's important to note that you can't call methods until after Sandbox has emitted a ready event! See the "Basic Usage" code snippet for an example.

call will return a Promise object to give you the result asynchronously. See the docs for Kris Zyp's node-promise for all available methods.

//this will call the method like so: myMethod(1, 2, 3);
sb.rpc.call("myMethod", [1, 2, 3]).then(
    function(result){
        console.log("Success! "+result);
    },
    function(error){
        console.log("We got an error: "+error.message);
    }
);
 
//if we pass call() an object as arguments, it'll call the method like this: myMethod(myObj);
var myObj = {foo: "bar"};
sb.rpc.call("myMethod", myObj);
 
 
//here's an example using notify(). It's arguments are identical to call()
sb.rpc.notify("myMethod", [1, 2, 3]); //no returned value

Again, the API is identical from within the sandbox. Just use the rpc global variable instead of Sandbox.rpc

rpc.call("myMethod", [1, 2, 3]);

RPC Call Timeouts

Sometimes, we might want to specify a timeout for method calls, just in case. We can do this one of two ways.

The first is to use the Promise api:

var p = rpc.call("myMethod", [1, 2, 3]);
p.timeout(10000); //timeout after 10 seconds
p.then(/* ... */);

The second is to add an additional call_timeout option, eg:

var sb = new Sandbox("path/to/code.js", {
    call_timeout: 10000 //10 seconds
});

By default, call_timeout is -1, which disables timeouts to method calls. The value specified in call_timeout will also be applied to the rpc class inside the sandbox.

Detecting when the sandbox exits

We can detect when the sandbox exits using the exit event.

sb.on("exit", function(){
    console.log("the sandbox ended!");
});

Lockup detection & killing the sandbox

node-sandbox has built in lockup detection, so if a stray while() loop locks up the sandbox, we can react to it.

To kill the sandbox, we can use Sandbox.kill().

var sb = new Sandbox("path/to/code.js", {
    permissions: [/*...*/],
    
    //Here are some relevant options for lockup detection.
    //All time is specified in milliseconds.
    //Set any of these values to -1 to disable them.
 
    //how long we should wait for a reply
    //before emitting a 'lockup' event. (default: 10 seconds)
    lockup_timeout: 10000,
    
    //how long we should wait after killing the process
    //to kill -9 it. (default: 10 seconds)
    kill_with_fire_timeout: 10000,
    
    //how frequently we should check the sandbox (default: 10 seconds)
    ping_interval: 10000,
    
    //how long we should wait before assuming
    //the sandbox failed to start (locked up immediately)
    //(default: 10 seconds)
    startup_timeout: 10000,
 
});
 
sb.run();
 
sb.on("lockup", function(){
    //restart the sandbox
    sb.kill();
    sb.run();
});

Detecting output on STDERR

If something ever goes wrong within the sandbox, by default it doesn't get printed to the main process' STDOUT. Instead, you need to listen on the stderr event and do it yourself, eg:

sb.on("stderr", function(text){
    console.err(text);
});

You can also pass this on to any logging library you use.

Pinging the Sandbox

If you want to ping the sandbox to figure out latency, you can use Sandbox.ping(), which returns a Promise object.

sb.ping().then(function(time){
    console.log("Sandbox latency: "+time);
}, function(err){
    console.log("Failed to ping sandbox!");
});

Plugins

node-sandbox has a full featured plugin system, and a lot of it's features are provided by built-in plugins. Built-in plugins can be found in lib/plugins/, and include the following:

• _base: not meant to be loaded, but provides base functionality to other plugins
• rpc: provides JSON-RPC functionality (exposed through Sandbox.rpc) over the Stream between the parent and child processes exposed by Sandbox.
• lockup_detection: provides lockup detection functionality, including the on("lockup") event. Relies on rpc.
• wrapper: wraps process.binding so that any unauthorized modules aren't allowed to be loaded.

Eventually I'd like to include a plugin that can set up a secure container for the child process using OS features (eg SELinux). If you know a lot about this sort of thing and would like to contribute, please let me know!

Plugins can be specified using the plugins option:

var sb = new Sandbox("path/to/file.js", {
    plugins: [
        //Note: these are the default plugins that are loaded. If you want to load an extra plugin, you should include these built-in ones too!
        "rpc", "lockup_detection", "wrapper"
    ]
});

Note that plugin hooks are executed in the order that they're provided in the array, so make sure "wrapper" goes last, otherwise a plugin might not have access to the resources it needs to initiate itself!

Some plugins in the future may take additional arguments, but all the built-in ones at the time of writing read from the main arguments passed to the sandbox (for the sake of ease-of-use). Here's an example on how to pass custom arguments to a plugin:

var sb = new Sandbox("path/to/file.js", {
    plugins: [
        {
            name: "my_plugin",
            options: {foo: "bar"}
        },
        "some_other_plugin"
    ]
});

To load an external plugin, simply pass a path to the directory containing all the plugin's files (manifest.js, etc.) instead of a plugin name. It can either be an absolute path, or relative to the lib/PluginManager.js file:

var sb = new Sandbox("path/to/file.js", {
    plugins: [
        {
            name: "/path/to/my_plugin",
            options: {foo: "bar"}
        },
        //OR
        "/path/to/my_plugin"
    ]
});

Writing Custom Plugins

The plugin system lets you hook into the sandbox and add any functionality you want. Things you can do include:

• Extend the sandbox by adding custom options and methods, and pretty much override/wrap any function/variable you want
• Run code during certain events in the parent process (eg when the child process exits, when the process writes to stderr, etc.)
• Run code inside the child process (eg when the process spawns, after the code is loaded, after the code is executed, etc.)
• Pass extra arguments to the child process via startData

Using this, you can write your own RPC plugin, container plugin, or anything else you need. I strongly suggest looking in the lib/plugins directory for examples, especially at the _base plugin, which documents when each hook is called.

Plugins consist of three files:

• plugin_name/manifest.js: a manifest file that provides information about the plugin.
• plugin_name/ParentHooks.js: a class that provides hooks for the parent process and allows you to hook into each Sandbox instance.
• plugin_name/ShovelHooks.js: a class that provides hooks for shovel.js, which is what's run to create our child process.

ParentHooks and ShovelHooks should both extend the respective classes in the _base plugin. manifest.js can simply be copied to your plugin's directory.

The PluginManager class loads the plugins, and does some basic dependency/conflict checks. Below is the manifest.js file from the _base plugin:

module.exports = {
    name: "_base", //the name of the parent directory our plugin is in
    provides: [], //features that it provides. This is flexible, so it can be something like "container" or "rpc".
    conflicts: [], //features or specific plugins this plugin conflicts with.
    depends: [] //features or specific plugins that this plugin requires to run.
}

To put this into practice, lets say we want to write a replacement for the RPC plugin. If we put "rpc" in the provides array, the plugin manager will throw an error if any other plugins that provide the "rpc" functionality are loaded. This way two plugins won't fight over access to Sandbox.rpc.

If we don't provide RPC functionality, but for one reason or the other we conflict with the RPC module (maybe our plugin wants to use the Stream between parent and child processes exclusively), we can put it in the conflicts array instead.

If our plugin depended on the RPC class in order to pass data between the parent and child processes, we could put "rpc" in the depends array. Note that any plugin that provides the "rpc" functionality would satisfy this requirement. For this reason, it's strongly suggested that any module that provides a specified functionality should have an identical base API (extra functions are allowed to be implemented).

To implement subclasses of ParentHooks and ShovelHooks, take a look inside of the respective class definitions in the _base class. Everything is well documented in there, and will explain how to access things through member variables, and which methods get called when.