Making WebWorkers a safe environment

Question

In a quest to have an interface capable of running arbitrary javascript code inside the browser, without having a security hole the size of a typical yo-mama joke, Esailija proposed using Web Workers. They run in a semi-sandboxed environment (no DOM access and already inside the browser) and can be killed so the user can't put them in an infinite loop.

Here's the example he brought up: http://tuohiniemi.fi/~runeli/petka/workertest.html (open your console)

jsfiddle (Google chrome only)

Now, this seems like a good solution; however, is it a complete (or approaching complete) one? Is there anything obvious missing?

The entire thing (as it's hooked up to a bot) can be found on github: worker, evaluator

main:

workercode = "worker.js";

function makeWorkerExecuteSomeCode( code, callback ) {
    var timeout;

    code = code + "";
    var worker = new Worker( workercode );

    worker.addEventListener( "message", function(event) {
        clearTimeout(timeout);
        callback( event.data );
    });

    worker.postMessage({
        code: code
    });

    timeout = window.setTimeout( function() {
        callback( "Maximum execution time exceeded" );
        worker.terminate();
    }, 1000 );
}

makeWorkerExecuteSomeCode( '5 + 5', function(answer){
    console.log( answer );
});

makeWorkerExecuteSomeCode( 'while(true);', function(answer){
    console.log( answer );
});

var kertoma = 'function kertoma(n){return n === 1 ? 1 : n * kertoma(n-1)}; kertoma(15);';

makeWorkerExecuteSomeCode( kertoma, function(answer){
    console.log( answer );
});

worker:

var global = this;

/* Could possibly create some helper functions here so they are always available when executing code in chat?*/

/* Most extra functions could be possibly unsafe */

    var wl = {
        "self": 1,
        "onmessage": 1,
        "postMessage": 1,
        "global": 1,
        "wl": 1,
        "eval": 1,
        "Array": 1,
        "Boolean": 1,
        "Date": 1,
        "Function": 1,
        "Number" : 1,
        "Object": 1,
        "RegExp": 1,
        "String": 1,
        "Error": 1,
        "EvalError": 1,
        "RangeError": 1,
        "ReferenceError": 1,
        "SyntaxError": 1,
        "TypeError": 1,
        "URIError": 1,
        "decodeURI": 1,
        "decodeURIComponent": 1,
        "encodeURI": 1,
        "encodeURIComponent": 1,
        "isFinite": 1,
        "isNaN": 1,
        "parseFloat": 1,
        "parseInt": 1,
        "Infinity": 1,
        "JSON": 1,
        "Math": 1,
        "NaN": 1,
        "undefined": 1
    };

    Object.getOwnPropertyNames( global ).forEach( function( prop ) {
        if( !wl.hasOwnProperty( prop ) ) {
            Object.defineProperty( global, prop, {
                get : function() {
                    throw new Error( "Security Exception: cannot access "+prop);
                    return 1;
                }, 
                configurable : false
            });    
        }
    });

    Object.getOwnPropertyNames( global.__proto__ ).forEach( function( prop ) {
        if( !wl.hasOwnProperty( prop ) ) {
            Object.defineProperty( global.__proto__, prop, {
                get : function() {
                    throw new Error( "Security Exception: cannot access "+prop);
                    return 1;
                }, 
                configurable : false
            });    
        }
    });




onmessage = function( event ) {
    "use strict";
    var code = event.data.code;
    var result;
    try {
        result = eval( '"use strict";\n'+code );
    }
    catch(e){
        result = e.toString();
    }
    postMessage( "(" + typeof result + ")" + " " + result );
};

Zales

Wouldn't they still be able to send AJAX requests? – Centrum

@Centrum The worker has XHR set to null – Contentment

@Centrum You can manually remove a bunch of insecure properties yourselves, as shown in the example. – Zales

It appears that if you delete a native/host object, it will be restored in its original state. "delete XMLHttpRequest; XMLHttpRequest;" Will return the original XMLHttpRequest object. There must be a way around this :/ – Contentment

Nvm, a work around was suggested by @copy. Editing in – Contentment

Frankly, this is brilliant. I wish I could give you more rep for this. – Flatten

You have the blacklist of functions, that are unsafe. It doesn't look good. What if new standard or browser defines new unsafe functions? – Lederman

I think you need whitelist - remove all functions, except those that you know they are safe. – Lederman

@Lederman I just noticed this when I enumerated over global object in worker. I am making a whitelist approach. Thanks – Contentment

@Lederman we are using whitelist approach now, how does it look? – Contentment

As a paranoid I would add some kind of runtime test, that your method actually blocks something on working system. Otherwise, looks okay, but I'm not an expert. – Lederman

How about using something like: code.google.com/p/google-caja ? – Southeasterly

Maybe running a VM such as continuum could be an idea? – Oversoul

@IdoGreen That'd require incorporating a huge outside framework. A native solution is preferred. – Zales

@FlorianMargaine That's learning to the preferable, but comes with a huge performance kill – Zales

What exact workaround was added to prevent delete keyword from undoing it? @Contentment – Reputed

Answer 1

The current code (listed below) has been now in use in the Stackoverflow javascript chat room for a while and so far the toughest problem was Array(5000000000).join("adasdadadasd") instantly crashing some browser tabs for me when I was running the code executor bot. Monkeypatching Array.prototype.join seems to have fixed this and the maximum execution time of 50ms has worked for any other attempt to hog memory or crash the browser.

var global = this;

/* Could possibly create some helper functions here so they are always available when executing code in chat?*/

/* Most extra functions could be possibly unsafe */

var wl = {
    "self": 1,
    "onmessage": 1,
    "postMessage": 1,
    "global": 1,
    "wl": 1,
    "eval": 1,
    "Array": 1,
    "Boolean": 1,
    "Date": 1,
    "Function": 1,
    "Number" : 1,
    "Object": 1,
    "RegExp": 1,
    "String": 1,
    "Error": 1,
    "EvalError": 1,
    "RangeError": 1,
    "ReferenceError": 1,
    "SyntaxError": 1,
    "TypeError": 1,
    "URIError": 1,
    "decodeURI": 1,
    "decodeURIComponent": 1,
    "encodeURI": 1,
    "encodeURIComponent": 1,
    "isFinite": 1,
    "isNaN": 1,
    "parseFloat": 1,
    "parseInt": 1,
    "Infinity": 1,
    "JSON": 1,
    "Math": 1,
    "NaN": 1,
    "undefined": 1
};

Object.getOwnPropertyNames( global ).forEach( function( prop ) {
    if( !wl.hasOwnProperty( prop ) ) {
        Object.defineProperty( global, prop, {
            get : function() {
                throw "Security Exception: cannot access "+prop;
                return 1;
            }, 
            configurable : false
        });    
    }
});

Object.getOwnPropertyNames( global.__proto__ ).forEach( function( prop ) {
    if( !wl.hasOwnProperty( prop ) ) {
        Object.defineProperty( global.__proto__, prop, {
            get : function() {
                throw "Security Exception: cannot access "+prop;
                return 1;
            }, 
            configurable : false
        });    
    }
});

Object.defineProperty( Array.prototype, "join", {

    writable: false,
    configurable: false,
    enumerable: false,

    value: function(old){
        return function(arg){
            if( this.length > 500 || (arg && arg.length > 500 ) ) {
                throw "Exception: too many items";
            }

            return old.apply( this, arguments );
        };
    }(Array.prototype.join)

});


(function(){
    var cvalues = [];

    var console = {
        log: function(){
            cvalues = cvalues.concat( [].slice.call( arguments ) );
        }
    };

    function objToResult( obj ) {
        var result = obj;
        switch( typeof result ) {
            case "string":
                return '"' + result + '"';
                break;
            case "number":
            case "boolean":
            case "undefined":
            case "null":
            case "function":
                return result + "";
                break;
            case "object":
                if( !result ) {
                    return "null";
                }
                else if( result.constructor === Object || result.constructor === Array ) {
                    var type = ({}).toString.call( result );
                    var stringified;
                    try {
                        stringified = JSON.stringify(result);
                    }
                    catch(e) {
                        return ""+e;
                    }
                    return type + " " + stringified;
                }
                else {
                    return ({}).toString.call( result );
                }
                break;

        }

    }

    onmessage = function( event ) {
        "use strict";
        var code = event.data.code;
        var result;
        try {
            result = eval( '"use strict";\n'+code );
        }
        catch(e) {
            postMessage( e.toString() );
            return;
        }
        result = objToResult( result );
        if( cvalues && cvalues.length ) {
            result = result + cvalues.map( function( value, index ) {
                return "Console log "+(index+1)+":" + objToResult(value);
            }).join(" ");
        }
        postMessage( (""+result).substr(0,400) );
    };

})();

Contentment

Why are Math, and parse* functions disabled? – Neau

@Neau why do you think they are disabled – Contentment

@Contentment So far, I only found cross-domain Math.random() prediction. But since the attacker cannot communicate to the outside world, it seems to be harmless. – Neau

Btw, I am also disallowing XMLHttpRequest, Worker, WebSocket and importScripts. How do you guys cope with that? – Neau

@Neau I think you are misunderstanding - this is a whitelist approach. Anything not listed is disallowed while only the things that are listed are allowed. Whitelist is better because it doesn't require updating when new APIs are implemented. – Contentment

What would happen if any of the insecure properties are already non-configurable? – Mccully

@Mccully that is not happening in the use case – Contentment

How do you know that for future APIs? :-) But I've checked it now, Object.defineProperty would throw and prevent the onmessage handler from being set up. – Mccully

I tried this in Chrome, and it reported "cannot redefine property Intl". Adding it to the whitelist fixed it. – Neau

One more stupid question: What is dangerous about allowing console? Is that also part of making it "future-proof", or can it currently be exploited? – Neau

@Neau I don't think there is a console in a web worker – Contentment

@Contentment there is, at least on Chrome and IE. Been using it for quite a while already. I am guessing that it is running on its own thread. – Neau

@Contentment bad news: There is a Chrome bug (or feature?) where a whole bunch of globals are not listed by Object.getOwnPropertyNames, including all *Timeout and *Interval functions. This JSFiddle shows a rather long list of such globals (at least in Chrome). This script won't run in IE because it uses in-line Workers (The current IE version disallows running workers from Blob URLs, for some stupid reason). – Neau

@Neau that's not a bug and those are found by Object.getOwnPropertyNames( global.__proto__ ) – Contentment

@Contentment That is not true. global.__proto__ only finds two additional globals in my Chrome: constructor and postMessage. The rest is still missing, including *Timeout and *Interval functions. See this updated JSFiddle for proof. – Neau

@Neau *Timeout and *Interval function are listed jsfiddle.net/WuhrP/2 (I am using version 33.0.1750.152) – Contentment

@Contentment Apologies, my code probably had a bug then. Anyway, would you mind taking a look at this related question? – Neau

@Contentment The code in your answer looks like the one in the question. You might want to check out this answer. – Ensile

Is there any way for code in the WebWorker to undo the disabling, eg with delete keyword? – Reputed

Hi @Esailija, are you still using this code in stackoverflow chat? – Kenlee

Answer 2

The code currently (2014-11-07) shown in the question despite ostensibly disallowing access to XMLHttpRequest (since it is not whitelisted), still allows code to access it.

If I put the code in the question (or the accepted answer) in a web page and worker combo and execute the following code on Chrome 38:

makeWorkerExecuteSomeCode('event.target.XMLHttpRequest', function (answer) { console.log( answer ); });

The result is:

function XMLHttpRequest() { [native code] } 

However it does not work in FF. Bug in Chrome?

Another thing I found but which does not seem to lead very far down the rabit hole is reinstating console.log. This works on FF 31 but not Chrome 38:

makeWorkerExecuteSomeCode(
    'var c = self.__proto__.__proto__.__lookupGetter__("console").call(self); c.log("FOO");', 
    function (answer) { console.log(answer) });

This would log "FOO" to the console without passing through the fake console.log that the web worker provides. The code above uses self, which can be blacklisted (by removing it from the whitelist) but this and global also work. I've found that attempts to blacklist global fail on FF and Chrome: the worker dies with an error.

Note: Chrome refuses to blacklist Intl so it has to be added to the whitelist for the code to run at all.

Ensile

Very good point concerning the hole in Chrome 38, but that is rather easy to fill: Just put a closure around the try/catch in onmessage and redefine event with var event; within the closure. – Habitancy

So is there any real solution? – Reputed