Testing Web Applications¶

In this chapter, we explore how to generate tests for Graphical User Interfaces (GUIs), notably on Web interfaces. We set up a (vulnerable) Web server and demonstrate how to systematically explore its behavior – first with handwritten grammars, then with grammars automatically inferred from the user interface. We also show how to conduct systematic attacks on these servers, notably with code and SQL injection.

A Web User Interface¶

Let us start with a simple example. We want to set up a Web server that allows readers of this book to buy fuzzingbook-branded fan articles ("swag"). In reality, we would make use of an existing Web shop (or an appropriate framework) for this purpose. For the purpose of this book, we write our own Web server, building on the HTTP server facilities provided by the Python library.

Running the Server¶

We run the server on the local host – that is, the same machine which also runs this notebook. We check for an accessible port and put the resulting URL in the queue created earlier.

The function start_httpd() starts the server in a separate process, which we start using the multiprocess module. It retrieves its URL from the message queue and returns it, such that we can start talking to the server.

Let us now start the server and save its URL:

Interacting with the Server¶

Let us now access the server just created.

Direct Browser Access¶

If you are running the Jupyter notebook server on the local host as well, you can now access the server directly at the given URL. Simply open the address in httpd_url by clicking on the link below.

Note: This only works if you are running the Jupyter notebook server on the local host.

Even more convenient, you may be able to interact directly with the server using the window below.

Note: This only works if you are running the Jupyter notebook server on the local host.

After interaction, you can retrieve the messages produced by the server:

We can also see any orders placed in the orders database (db):

And we can clear the order database:

Retrieving the Home Page¶

Even if our browser cannot directly interact with the server, the notebook can. We can, for instance, retrieve the contents of the home page and display them:

Placing Orders¶

To test this form, we can generate URLs with orders and have the server process them.

The method urljoin() puts together a base URL (i.e., the URL of our server) and a path – say, the path towards our order.

With urljoin(), we can create a full URL that is the same as the one generated by the browser as we submit the order form. Sending this URL to the browser effectively places the order, as we can see in the server log produced:

The web page returned confirms the order:

And the order is in the database, too:

Error Messages¶

We can also test whether the server correctly responds to invalid requests. Nonexistent pages, for instance, are correctly handled:

You may remember we also have a page for internal server errors. Can we get the server to produce this page? To find this out, we have to test the server thoroughly – which we do in the remainder of this chapter.

Fuzzing Input Forms¶

After setting up and starting the server, let us now go and systematically test it – first with expected, and then with less expected values.

Fuzzing with Expected Values¶

Since placing orders is all done by creating appropriate URLs, we define a grammar ORDER_GRAMMAR which encodes ordering URLs. It comes with a few sample values for names, email addresses, cities and (random) digits.

In the grammar, we make use of cgi_encode() to encode strings:

Using one of our grammar fuzzers, we can instantiate this grammar and generate URLs:

Sending these URLs to the server will have them processed correctly:

Fuzzing with Unexpected Values¶

We can now see that the server does a good job when faced with "standard" values. But what happens if we feed it non-standard values? To this end, we make use of a mutation fuzzer which inserts random changes into the URL. Our seed (i.e. the value to be mutated) comes from the grammar fuzzer:

Mutating this string yields mutations not only in the field values, but also in field names as well as the URL structure.

Let us fuzz a little until we get an internal server error. We use the Python requests module to interact with the Web server such that we can directly access the HTTP status code.

That didn't take long. Here's the offending URL:

How does the URL cause this internal error? We make use of delta debugging to minimize the failure-inducing path, setting up a WebRunner class to define the failure condition:

This is the minimized path:

It turns out that our server encounters an internal error if we do not supply the requested fields:

We see that we might have a lot to do to make our Web server more robust against unexpected inputs. The exercises give some instructions on what to do.

Extracting Grammars for Input Forms¶

In our previous examples, we have assumed that we have a grammar that produces valid (or less valid) order queries. However, such a grammar does not need to be specified manually; we can also extract it automatically from a Web page at hand. This way, we can apply our test generators on arbitrary Web forms without a manual specification step.

Searching HTML for Input Fields¶

The key idea of our approach is to identify all input fields in a form. To this end, let us take a look at how the individual elements in our order form are encoded in HTML:

We see that there is a number of form elements that accept inputs, in particular <input>, but also <select> and <option>. The idea now is to parse the HTML of the Web page in question, to extract these individual input elements, and then to create a grammar that produces a matching URL, effectively filling out the form.

To parse the HTML page, we could define a grammar to parse HTML and make use of our own parser infrastructure. However, it is much easier to not reinvent the wheel and instead build on the existing, dedicated HTMLParser class from the Python library.

During parsing, we search for <form> tags and save the associated action (i.e., the URL to be invoked when the form is submitted) in the action attribute. While processing the form, we create a map fields that holds all input fields we have seen; it maps field names to the respective HTML input types ("text", "number", "checkbox", etc.). Exclusive selection options map to a list of possible values; the select stack holds the currently active selection.

While parsing, the parser calls handle_starttag() for every opening tag (such as <form>) found; conversely, it invokes handle_endtag() for closing tags (such as </form>). attributes gives us a map of associated attributes and values.

Here is how we process the individual tags:

• When we find a <form> tag, we save the associated action in the action attribute;
• When we find an <input> tag or similar, we save the type in the fields attribute;
• When we find a <select> tag or similar, we push its name on the select stack;
• When we find an <option> tag, we append the option to the list associated with the last pushed <select> tag.

Our implementation handles only one form per Web page; it also works on HTML only, ignoring all interaction coming from JavaScript. Also, it does not support all HTML input types.

Let us put this parser to action. We create a class HTMLGrammarMiner that takes a HTML document to parse. It then returns the associated action and the associated fields:

Applied on our order form, this is what we get:

From this structure, we can now generate a grammar that automatically produces valid form submission URLs.

Mining Grammars for Web Pages¶

To create a grammar from the fields extracted from HTML, we build on the CGI_GRAMMAR defined in the chapter on grammars. The key idea is to define rules for every HTML input type: An HTML number type will get values from the <number> rule; likewise, values for the HTML email type will be defined from the <email> rule. Our default grammar provides very simple rules for these types.

Our grammar miner now takes the fields extracted from HTML, converting them into rules. Essentially, every input field encountered gets included in the resulting query URL; and it gets a rule expanding it into the appropriate type.

Let us show HTMLGrammarMiner in action, again applied on our order form. Here is the full resulting grammar:

Let us take a look into the structure of the grammar. It produces URL paths of this form:

Here, the <action> comes from the action attribute of the HTML form:

The <query> is composed of the individual field items:

Each of these fields has the form <field-name>=<field-type>, where <field-type> is already defined in the grammar:

These are the query URLs produced from the grammar. We see that these are similar to the ones produced from our hand-crafted grammar, except that the string values for names, email addresses, and cities are now completely random:

We can again feed these directly into our Web browser:

We see (one more time) that we can mine a grammar automatically from given data.

A Fuzzer for Web Forms¶

To make things most convenient, let us define a WebFormFuzzer class that does everything in one place. Given a URL, it extracts its HTML content, mines the grammar and then produces inputs for it.

All it now takes to fuzz a Web form is to provide its URL:

We can combine the fuzzer with a WebRunner as defined above to run the resulting fuzz inputs directly on our Web server:

While convenient to use, this fuzzer is still very rudimentary:

• It is limited to one form per page.
• It only supports GET actions (i.e., inputs encoded into the URL). A full Web form fuzzer would have to at least support POST actions.
• The fuzzer is build on HTML only. There is no Javascript handling for dynamic Web pages.

Let us clear any pending messages before we get to the next section:

Crawling User Interfaces¶

So far, we have assumed there would be only one form to explore. A real Web server, of course, has several pages – and possibly several forms, too. We define a simple crawler that explores all the links that originate from one page.

Our crawler is pretty straightforward. Its main component is again a HTMLParser that analyzes the HTML code for links of the form

<a href="<link>">

and saves all the links found in a list called links.

The actual crawler comes as a generator function crawl() which produces one URL after another. By default, it returns only URLs that reside on the same host; the parameter max_pages controls how many pages (default: 1) should be scanned. We also respect the robots.txt file on the remote site to check which pages we are allowed to scan.

We can run the crawler on our own server, where it will quickly return the order page and the terms and conditions page.

We can also crawl over other sites, such as the home page of this project.

Once we have crawled over all the links of a site, we can generate tests for all the forms we found:

For even better effects, one could integrate crawling and fuzzing – and also analyze the order confirmation pages for further links. We leave this to the reader as an exercise.

Let us get rid of any server messages accumulated above:

Crafting Web Attacks¶

Before we close the chapter, let us take a look at a special class of "uncommon" inputs that not only yield generic failures, but actually allow attackers to manipulate the server at their will. We will illustrate three common attacks using our server, which (surprise) actually turns out to be vulnerable against all of them.

HTML Injection Attacks¶

The first kind of attack we look at is HTML injection. The idea of HTML injection is to supply the Web server with data that can also be interpreted as HTML. If this HTML data is then displayed to users in their Web browsers, it can serve malicious purposes, although (seemingly) originating from a reputable site. If this data is also stored, it becomes a persistent attack; the attacker does not even have to lure victims towards specific pages.

Here is an example of a (simple) HTML injection. For the name field, we not only use plain text, but also embed HTML tags – in this case, a link towards a malware-hosting site.

If we use this grammar to create inputs, the resulting URL will have all of the HTML encoded in:

What happens if we send this string to our Web server? It turns out that the HTML is left in the confirmation page and shown as link. This also happens in the log:

Since the link seemingly comes from a trusted origin, users are much more likely to follow it. The link is even persistent, as it is stored in the database:

This means that if anyone ever queries the database (for instance, operators processing the order), they will also see the link, multiplying its impact. By carefully crafting the injected HTML, one can thus expose malicious content to numerous users – until the injected HTML is finally deleted.

Cross-Site Scripting Attacks¶

If one can inject HTML code into a Web page, one can also inject JavaScript code as part of the injected HTML. This code would then be executed as soon as the injected HTML is rendered.

This is particularly dangerous because executed JavaScript always executes in the origin of the page which contains it. Therefore, an attacker can normally not force a user to run JavaScript in any origin he does not control himself. When an attacker, however, can inject his code into a vulnerable Web application, he can have the client run the code with the (trusted) Web application as origin.

In such a cross-site scripting (XSS) attack, the injected script can do a lot more than just plain HTML. For instance, the code can access sensitive page content or session cookies. If the code in question runs in the operator's browser (for instance, because an operator is reviewing the list of orders), it could retrieve any other information shown on the screen and thus steal order details for a variety of customers.

Here is a very simple example of a script injection. Whenever the name is displayed, it causes the browser to "steal" the current session cookie – the piece of data the browser uses to identify the user with the server. In our case, we could steal the cookie of the Jupyter session.

The message looks as always – but if you have a look at your browser title, it should now show the first 10 characters of your "secret" notebook cookie. Instead of showing its prefix in the title, the script could also silently send the cookie to a remote server, allowing attackers to highjack your current notebook session and interact with the server on your behalf. It could also go and access and send any other data that is shown in your browser or otherwise available. It could run a keylogger and steal passwords and other sensitive data as it is typed in. Again, it will do so every time the compromised order with Jane Doe's name is shown in the browser and the associated script is executed.

Let us go and reset the title to a less sensitive value:

SQL Injection Attacks¶

Cross-site scripts have the same privileges as web pages – most notably, they cannot access or change data outside your browser. So-called SQL injection targets databases, allowing to inject commands that can read or modify data in the database, or change the purpose of the original query.

To understand how SQL injection works, let us take a look at the code that produces the SQL command to insert a new order into the database:

sql_command = ("INSERT INTO orders " +
    "VALUES ('{item}', '{name}', '{email}', '{city}', '{zip}')".format(**values))

What happens if any of the values (say, name) has a value that can also be interpreted as a SQL command? Then, instead of the intended INSERT command, we would execute the command imposed by name.

Let us illustrate this by an example. We set the individual values as they would be found during execution:

and format the string as seen above:

All fine, right? But now, we define a very "special" name that can also be interpreted as a SQL command:

What happens here is that we now get a command to insert values into the database (with a few "dummy" values x), followed by a SQL DELETE command that would delete all entries of the orders table. The string -- starts a SQL comment such that the remainder of the original query would be easily ignored. By crafting strings that can also be interpreted as SQL commands, attackers can alter or delete database data, bypass authentication mechanisms and many more.

Is our server also vulnerable to such attacks? Of course, it is. We create a special grammar such that we can set the <name> parameter to a string with SQL injection, just as shown above.

These are the current orders:

Let us go and send our URL with SQL injection to the server. From the log, we see that the "malicious" SQL command is formed just as sketched above, and executed, too.

All orders are now gone:

This effect is also illustrated in this very popular XKCD comic:

{width=100%}

Even if we had not been able to execute arbitrary commands, being able to compromise an orders database offers several possibilities for mischief. For instance, we could use the address and matching credit card number of an existing person to go through validation and submit an order, only to have the order then delivered to an address of our choice. We could also use SQL injection to inject HTML and JavaScript code as above, bypassing possible sanitization geared at these domains.

To avoid such effects, the remedy is to sanitize all third-party inputs – no character in the input must be interpretable as plain HTML, JavaScript, or SQL. This is achieved by properly quoting and escaping inputs. The exercises give some instructions on what to do.

Leaking Internal Information¶

To craft the above SQL queries, we have used insider information – for instance, we knew the name of the table as well as its structure. Surely, an attacker would not know this and thus not be able to run the attack, right? Unfortunately, it turns out we are leaking all of this information out to the world in the first place. The error message produced by our server reveals everything we need:

The best way to avoid information leakage through failures is of course not to fail in the first place. But if you fail, make it hard for the attacker to establish a link between the attack and the failure. In particular,

• Do not produce "internal error" messages (and certainly not ones with internal information).
• Do not become unresponsive; just go back to the home page and ask the user to supply correct data.

One more time, the exercises give some instructions on how to fix the server.

If you can manipulate the server not only to alter information, but also to retrieve information, you can learn about table names and structure by accessing special tables (also called data dictionary) in which database servers store their metadata. In the MySQL server, for instance, the special table information_schema holds metadata such as the names of databases and tables, data types of columns, or access privileges.

Fully Automatic Web Attacks¶

So far, we have demonstrated the above attacks using our manually written order grammar. However, the attacks also work for generated grammars. We extend HTMLGrammarMiner by adding a number of common SQL injection attacks:

We see that several fields now are tested for vulnerabilities:

We create a SQLInjectionFuzzer that does it all automatically.

Our attack was successful! After less than a second of testing, our database is empty:

Again, note the level of possible automation: We can

• Crawl the Web pages of a host for possible forms
• Automatically identify form fields and possible values
• Inject SQL (or HTML, or JavaScript) into any of these fields

and all of this fully automatically, not needing anything but the URL of the site.

The bad news is that with a tool set as the above, anyone can attack websites. The even worse news is that such penetration tests take place every day, on every website. The good news, though, is that after reading this chapter, you now get an idea of how Web servers are attacked every day – and what you as a Web server maintainer could and should do to prevent this.

Lessons Learned¶

• User Interfaces (on the Web and elsewhere) should be tested with expected and unexpected values.
• One can mine grammars from user interfaces, allowing for their widespread testing.
• Consequent sanitizing of inputs prevents common attacks such as code and SQL injection.
• Do not attempt to write a Web server yourself, as you are likely to repeat all the mistakes of others.

We're done, so we can clean up:

Next Steps¶

From here, the next step is GUI Fuzzing, going from HTML- and Web-based user interfaces to generic user interfaces (including JavaScript and mobile user interfaces).

If you are interested in security testing, do not miss our chapter on information flow, showing how to systematically detect information leaks; this also addresses the issue of SQL Injection attacks.

Background¶

The Wikipedia pages on Web application security are a mandatory read for anyone building, maintaining, or testing Web applications. In 2012, cross-site scripting and SQL injection, as discussed in this chapter, made up more than 50% of Web application vulnerabilities.

The Wikipedia page on penetration testing provides a comprehensive overview on the history of penetration testing, as well as collections of vulnerabilities.

The OWASP Zed Attack Proxy Project (ZAP) is an open source website security scanner including several of the features discussed above, and many many more.

Exercises¶

Exercise 1: Fix the Server¶

Create a BetterHTTPRequestHandler class that fixes the several issues of SimpleHTTPRequestHandler:

Part 1: Silent Failures¶

Set up the server such that it does not reveal internal information – in particular, tracebacks and HTTP status codes.

Part 2: Sanitized HTML¶

Set up the server such that it is not vulnerable against HTML and JavaScript injection attacks, notably by using methods such as html.escape() to escape special characters when showing them.

Part 3: Sanitized SQL¶

Set up the server such that it is not vulnerable against SQL injection attacks, notably by using SQL parameter substitution.

Part 4: A Robust Server¶

Set up the server such that it does not crash with invalid or missing fields.

Part 5: Test it!¶

Test your improved server whether your measures have been successful.

Exercise 2: Protect the Server¶

Assume that it is not possible for you to alter the server code. Create a filter that is run on all URLs before they are passed to the server.

Part 1: A Blacklisting Filter¶

Set up a filter function blacklist(url) that returns False for URLs that should not reach the server. Check the URL for whether it contains HTML, JavaScript, or SQL fragments.

Part 2: A Whitelisting Filter¶

Set up a filter function whitelist(url) that returns True for URLs that are allowed to reach the server. Check the URL for whether it conforms to expectations; use a parser and a dedicated grammar for this purpose.

Exercise 3: Input Patterns¶

To fill out forms, fuzzers could be much smarter in how they generate input values. Starting with HTML 5, input fields can have a pattern attribute defining a regular expression that an input value has to satisfy. A 5-digit ZIP code, for instance, could be defined by the pattern

<input type="text" pattern="[0-9][0-9][0-9][0-9][0-9]">

Extract such patterns from the HTML page and convert them into equivalent grammar production rules, ensuring that only inputs satisfying the patterns are produced.

Exercise 4: Coverage-Driven Web Fuzzing¶

Combine the above fuzzers with coverage-driven and search-based approaches to maximize feature and code coverage.