An html sanitizer for C#

sanitizer After 3 months gestation and some bug fixes, HtmlSanitizer is reporting no hacking successes.

Does it mean that it rocks? I don’t think so, but it is probably strong enough to sail in stormy waters.

See my previous post to know how it works, but mainly test it online with the Patapage playground.

Being honest I received some complaints concerning the black list approach to CSS styles, but no one has hacked the current version (yet 🙂 ). In any case the code is open to changes, and I’m happy to receive your feedbacks.

Now we are proud to announce that there is a porting to C# by Beyers Cronje (thank you). You can find C# sources here (and Java ones here).  Warning: source code is already patched as suggested by Isaiah.

Other portings are welcome!

Solving the “halting problem”…

When I asked Gino (alias Roberto Baldi mostmobbed) a software solution for the “halting problem”, he told me “should not be so difficult”!

“In computability theory, the halting problem is a decision problem which can be stated as follows: given a description of a program and a finite input, decide whether the program finishes running or will run forever, given that input.” (from Wikipedia)

Alan Turing proved in 1936 that a general algorithm to solve the halting problem for all possible program-input pairs cannot exist; but as Gino says, Turing was a pessimist…

Why I’m scratching about the halting problem? Because I need power at BugsVoice user’s fingertips. In particular, I need to allow the execution of user’s JavaScript code on my server, and knowing if these scripts end may be comforting.20080102_9458

If you already heard about XSS  you probably know that “third party” code execution, authorized or not, could be a nightmare even in the client… can you imagine how can be ugly on your server a “pirate” code?

Going more in detail, BugsVoice is a service that receives a “bug” from a customer’s server, process the request locally, serves a friendly feedback to the user and stores the bug in its database.

JavaScript (JS from here on) server-side execution is involved in the “processing” phase.

We supply a pre-filled and certified set of “rules” for processing bugs, but we even allow customers to create their own rules.

JS gives you the power of inspecting the error to understand what happened and gives your customer an error better than a “500 server error” in order to comfort it and recover a situation where your application credibility is going down. An interesting reading about error recovering and error feedback is  the book “Defensive design for the web” from 37 Signal.

The complete BugsVoice process includes mainly three parts:

1) on the customer’s server side,  an error page that catches the exception, collects as much information as possible (logged user, time, server status, database status, memory etc.) and redirects the user to our BugsVoice server (see how to configure an error trapping page on BugsVoice blog for more details).

2) our server, reading user preferences recovers the error template. Each template is fully dynamical and customizable; it introduces some “variables” that can be filled from the error happened. Then our server creates two JS objects: the “bug” object filled with the error collected and the “template” object filled from layout skeleton.

3) the JS rules are executed to fill “template” from “bug” or for rejecting the request.

4) the layout is rendered to the user by  using “template” and “bug” objects. The bug is stored on our server.

5) the user feedback is collected and stored.

6) a “thank you” page is displayed to the user.

Then there is the error management but this is interesting “only” for BugsVoice’ users, not for this post. Here some error pages from BugsVoice:

image image image image

So every user can create its own rules in order to inspect, for instance, the received bug’ stacktrace trying to discover if a database problem happens, or if there is a problem with the latest version of  some browser.

Coming back to rules execution:
during step 3) we get rules from the user configuration and we execute them on our server. We use the Java SE 6 scripting features supplying an ECMAScript engine to run rules.  A scripting engine instance is isolated from the JVM environment and you must declare the resource (libraries) you want to made available in the execution context.

Before executing them, the context is fed by “bug” and “template” objects. Then we run the rules…(drum roll!).

A basic (and friendly) rule example :

if (bug.code==404)  errorPage.errorMessage="Page missing: you get this error because of...";

Of course this code is safe, but what happens if an evil user composes a pleasant rule like

function snake(s){
  return "s"+snake(s);

… or even worst?

Sadly Turing beats Gino 1-0, and there is no general solution to the question “does this rule ends?”.

The only possible solution is to narrow the scope of the problem by introducing some fences.

A solution is to set up an external observer using  multi-threading and watch dogs in order to kill processes after a while, but the best solution is to avoid infinite loop situations.

Rules in our context are used mainly for discovering string patterns in the error stacktrace and for building better feedback; we do not need to iterate or create complex functions, so reducing the set of possible JS statement is possible without loosing “power”.

Luckily in JS there is a limited set of statements for iteration and recursion; so if we are able to “kill” bad intentions by forbidding dangerous statement like “while”, “for” or function definition we can run rules with confidence.

This way  we reduce the complex halting problem to the (quite) easy problem of  HTML sanitization  (where you must  remove some unaccepted tags. See XSS war: a Java HTML sanitizer ).

Actually identifying “while” or “for” statements in a complex code is not as easy as finding the “while” string. The find/replace approach  it’s too rough, and here we need a more accurate solution in order to understand the difference between

while (true);


var dummy= “while(true)”;

that is obvious for us but not for a string searcher…

You must use something to analyze the code token by token.

ANTLR 3 supply all you need for tokenizing, parsing and walking your code. You need a JS grammar and then ANTLR will build all the stuff. We used the ES3 grammar from Xebic Reasearch  (BSD license) based on the original work of Patrick Hulsmeyer, that fits perfectly our needs.

With the AS3 grammar we built  parser, lexer and walker to analyze rule’s code to intercept every forbidden statement and avoid accepting dangerous scripts (at least I hope this). Only the rules that pass the test will be saved on the system and will be available to the script engine.

Ok, I can confess you, the post’ title is a little misleading, there is no way to solve the halting problem at least without cheating!

XSS war: a Java HTML sanitizer

Yesterday I was testing one of our forthcoming online services, in order to check XSS (Cross Site Scripting) robustness.

XSS and CSRF (Cross Site Request Forgery) are the two scary “black beasts” of online services in good company with scalability,  but while even nerd-programmers think about scalability, almost nobody takes care of XSS and CSRF, at least until the first attack 🙂

The XSS becomes a serious  problem when you allow your users to add contents on your site/service; typical situations are blogs, forums, online chats etc.: when you are about to “print” user contributions you must do it conscientiously, you may found some “easter egg”.

Of course if you limit user contributions to plain text you will solve the problem in minutes, by just encoding every html tag (the idea is to change every “>” in “&gt;” and “<” in “&lt;” and so on). But things quickly get harder when you try to accept some basic html tags.

In our case, Patapage is a service that allows to add comments, threads and more options to existing  web sites, it is tailored to appealing interfaces so it is a MUST to allow users to insert a wide set of html tags. We have found on a good balance from admitted tags and refused ones; as usual Jeff  Atwood is prodigal of precious hints, but in our case while his set is fine for “patacomments” it is too restrictive for the scope of “patacontents”.

As you can imagine, I’ve found a lots of holes, so going upstream as usual and not giving a damn on, probably, the best Jeff’ suggestion (“do not re-write you own implementation”, an advice which he too is not following) I started writing code basing my implementation on these articles.

Another aspect of the problem is the conservation of layout: users can break the page by inserting unclosed or misplaced tags. If you hope that these flaws do not impact on security you are on the wrong path. A well planned css attack can, for instance, layer your application for click stealing or simply “switch” two buttons of your application with funny(?) results.

Of course I’ve looked around to find something matching my needs, but I found only two kind of approaches:

The first approach isvery basic focused on removing “< s c r i p t>” tags from your html; obviously this is a silly approach, you can inject js code without a “script” tag: you have a bunch of ways to do this by using dom events (onclick, onload etc.).
The second approach is to parse HTML to understand exactly what is happening in the code in order to remove un-allowed tags; this approach is logically the right one, you cannot expect to do better. Your code analyses the HTML, extracts the tags, and then you walk down the tree dropping out unwanted ones. Unluckily this approach requires a HTML parser library really “strong” with respect to malicious code, and usually these libraries are built for a different scope. Another aspect is that parser, lexer and walker are quite complicated pieces of code, so it is not a joke to test them completely. I’ve tested a couple of parsers with unsatisfying results.

This is why we wrote our own sanitizer by hand. Our approach is to remove unwanted tags and properties without testing HTML correctness in deep.

First step is to tokenize the code: a token could be one of : tag start (), comment (), tag content (blah blah), a tag closing ().

For instance <p style=”color:red” align=”center”>test</p> generates three tokens:

  1. <p style=”color:red” align=”center”>
  2. test
  3. </p>

The tokenize method looks for <…> pairs or comments, and that its fine for our scope of restricting accepted tags: if a <…> pair is badly closed the tag will be html-encoded.

Having the tokens list, we will test every single token whether it is acceptable; again, we do not perform tag matching at first, so for us <b>test</i> its fine, we are working on security, not on syntax – we’ll also fix that afterward, as it is easy in any case to add a tag-pair counter to close at the end unclosed tags, which you actually find done in our code.

We loop for every single token and we test it with regular expressions. The flow is:

  1. if token is a comment discard it.
  2. if token is a start tag (<p style=”color:red” > ) extract the tag (p) and attributes (style=”color:red” align=”center”)
    1. if tag is forbidden it will be removed
    2. if tag is allowed we will extract every attribute performing a check
      1. check “href” and “src” for admitted tags (a, img, embed only) and check url validity (only http or https)
      2. check “style” attribute looking for “url(…)” parameter, and eventually discarding it
      3. remove every “on…” attribute – e.g. onClick, onLoad, …
      4. encode attribute’ value for unknown ones
      5. push the tag on the stack of open tags
    3. else the tag is unknown and will be removed
  3. if the token is an end tag (</p> ) extract the tag (p) and check if the corresponding tag is already open. Eventually close those that are still open.
  4. else it is not a tag and we will encode it

In order to avoid js injection on user inserted URLs we will accept tag only if the “href” attribute is there and points to a valid URL; we test url correctness with Apache UrlValidator, and this will cut out every “javascript:” tries. Same approach for the and tags.

Finished my hard work coding the shelter, I give the happy news to our design department that the sanitizer was done, and after a first minute of excitement Matteo ( told me that they have three different usages of user input: for displaying an HTML page on the front office, for displaying a textual abstract in lists in the backoffice, and of course for storing contents on the database.

So a sanitizer needs three different outputs, html-encoded with tag, text-only without tag, and the “original” version for the database. This is why the latest version of the sanitizer returns “.html”, “.text”  and “.val”. Why you should store “.val” instead of the original input or “.html”?  Because the original input may be “dangerous”, and may mislead  the user in believing that all tags are allowed. The encoded value is not suitable in case of subsequent modification because of  double encodings (e.g. “>” –> “>” –> “>” and so on). On the other side “.val” removes only forbidden tags maintaining all other user oddities (strange tags, comments, etc.).

We have set-up a public playground for testing our sanitization code: This page allows you to input a text and by pressing “test” your input will be printed (sanitized) on the page.

Source of our sanitizer are released under MIT license (i.e. free as free beer, just keep the attribution); see the complete code here.

These tags will be accepted, others will be encoded, and printed. If you like challenges try  to inject a js in your text and, for instance, get an alert. Tell us about your victories, if any.

BTW, I’ve tested my code using XSS Me plugin for Firefox, and it passed all (about 15o) tests 🙂