DDS Value Object Presentation Video

Dear Junior

At the OWASP AppSec conference I had a presentation on value objects and Domain Driven Security. It turned out into a 35 minute code kata with refactorings, where I used value objects "DDD style" in two ways.

Firstly, I used them to create a design that makes indata validation come natural in the context of Injection Flaws. I used the classical SQL Injection login-attack as an example and applied "hard type" value objects in the same way as we have discussed before.

Secondly, I addressed a Cross-Site Scripting (XSS) scenario. Here I noted that even if XSS is often perceived as a problem with "bad indata", there are cases when those data are perfectly valid. So, instead I choose to look at it from an output-encoding perspective - not "bad indata validation", but "bad outdata encoding".

To be able to do something about the outdata encoding, you can think about the client side browser as a subsystem of your system. Seen that way, the presentation tier plays the role of the API to that subsystem. Then it becomes obvious that we should enforce proper encoding at the border of that subsystem - i e in the API we use for calling it.

Modifying the presentation tier API from "soft type" strings to "hard type" value objects made it obvious where proper encoding should go - thus solving the XSS problem.

The full video coverage from the conference has been released, and my presentation can be found at the OWASP website.

Yours

Dan

SQL Injection is not an Indata Validation Problem

Dear Junior

If DDS style use of value objects solves the indata validation problem , and if DDS style indata validation does not solve SQL Injection, then there is only one logical conclusion to draw.

SQL Injection is not an indata validation problem. Yes, that might be contrary to popular belief, but it is obviously so.

If not an indata validation problem, what kind of beast is it? And what can we do about it?

Yours

Dan

Indata Validation is Not Enough for SQL Injection

Dear Junior

When we do indata validation through value objects we get an application tier that is water-proof. The model describes exactly those data that we think are meaningful and can handle (“username, identifier with which the user present herself to the system; regexp [a-z]+“. Each piece of indata is validated as part of the value object constructor (public Username(String s)). For indata to pass through to application, it cannot avoid validation, as the application methods require the value object type (void authenticate(Username uname…)). What more can you ask for?

- Well, you see, we do not want usernames like “danbj”, we would prefer our real names, like “Dan Bergh Johnnsson”.

No problem, we expand the regexp with uppercase letters and space, getting a regexp like [a-zA-Z\ ]+.

- Nice, but our respected colleague Fredrik Jägare-Lilja needs a username as well.

Fair enough – we stuff Scandinavian letters and hyphen into the regexp as well, giving [a-zåäöA-ZÅÄÖ\ \-]+.

- Now there is only one person left: our highly respected Irish colleague Oliver O’Hehir.

Well, well, we are almost finished then, we only need to put the apostrophe into the regexp ending up with something along [a-zåäöA-ZÅÄÖ\ \-\’]+.

Wait, wait, wait!!! Who the h*** just logged in with username “’ OR ‘a’ is not null --“?

Sure, we might have tightened up the regexp to block out that specific attack string and any other malicious use of the format we can think of. But, it is always those we did not think of that causes the trouble.

Well, as always we can think that “SQL Injection is solved by prepared statements”, but remember that Injection Flaw is much larger than SQL Injection. The same vulnerabilities might be there when doing LDAP access, using parameters to construct file names (e g Directory Traversal), or if you have some Domain Specific Language (DSL) which you interpret. In any of these cases there might be a string that might well be fully legally formatted, but attacks the structure of how the underlying resources are used.

Over to a completely different domain: FM broadcast and music radio. In the FM radio broadcast system you must be able to shut down the transmitters from a remote site. Unfortunately, there is only one way to communicate with the transmitters – via radio. The problem was solved by defining a specific sequence of audio blips (very precise on frequencies, duration, and interval) and denoting that sequence the meaning “shut down the transmitter”.

The pioneering Swedish rap group JustD put that exact sequence as the final beat on one of their songs, without telling anyone. They must have laughed all the way home from the studio. That song has been played on Swedish radio exactly once.

The JustD track hack is a wonderful example of exploiting an Injection Flaw, there is no way to escape it “in band”. I have the same gut feeling about indata validation and SQL Injection.

No matter how we structure the indata model, there might always be some data that actually is valid indata, but causes the system to crash.

So, indata modelling and validation in all its glory: However necessary it is for upholding security, it is not sufficient.

Yours

Dan

Validating away SQL Injection

Dear Junior

So, if we want to ban ’OR 1=1 from being used as a username, we have to put that restriction into the model.

We have noticed that viewing the username as just a string does not help us much.

Integer authenticate(String username, String passwordMD5)

So, to be able to talk about usernames in a meaningful way , we introduced the class Username. This actually makes a huge difference as the authentication method is now explicitly distinguishing the username from other kind of indata.

Integer authenticate(Username username, String passwordMD5)

With username as an explicit part of the model, we now can say something like “‘ OR 1=1 is not a valid username”. The beautiful part is that this sentence both speaks immediately about a property in the conceptual mode, and about a property of the code.

The sentence is even well-formed to the degree that it can be turned into a unit test case.

  @Test
  public void shouldNotRegardInjectionAttackStringAsValid() {
    assertFalse(new Username("' OR 1=1 --").isValid());
  }

Not bad, if you can write a requirement as a unit test, then you are pretty well off. And, if the test is failing: the better, we have a licence to code, and we know when we are done. Not bad at all.

Implementing, adding a few related tests and going back to green will probably take us to a Username class looking roughly like this.

public class Username {
  // final making it immutable
  public final String username;
  public Username(String username) {
    this.username = username;
  }
  public boolean isValid() {
    return username.matches("[a-z]+");
  }
}

Hold the horses! We have now modified a class that encodes a central part of our domain at hand. So, we have actually changed the model. That is nothing to take lightly. Now we have to walk the round checking with the other stakeholders in the domain that this change is something we all can agree upon. Is it really ok that usernames should be only lowercase letters and that there must be at least one? In other words, should they look like “danbj”, “danberghjohnsson” and “ilovemylittlepony”?

In the case of usernames people are used to strange formats and restrictions, so my guess is that no one will object, but we have to check – otherwise the model turns from being a shared and agreed model to being “the developers stuff”.

Upholding the model as the ubiquitous language for talking about the system is essential for the quality of the system – it was that language that enabled us to turn a one-phrase requirement and turn it immediately into code. Further on, the quality of the system is essential for its security.

So – for the sake of security, if nothing else, work hard with that model.

Going back to the API we are building, it turns out pretty well. We have achieved that it is easy to use correctly – i e, when doing the right thing there is no much doubt. At the moment, you have to create a Username object (constructor), check that it is valid (isValid method) and then pass it to the authentication.

However, we have not yet achieved that the API is hard to use incorrectly, it is not “stable to erroneous use”. One really weak spot is that it is easy to miss calling the isValid method, and there is nothing in the rest of the API that catches that mistake.

So, we are not finished yet.

Yours

Dan

ps Of course, what we want to do is to ensure validation, not hoping that our fellow programmers will magically do the right thing.

Domain Driven Security and Making Stuff Explicit in the Model

Dear Junior

”But ’ OR 1=1 -- is not a valid username! That is just bad indata validation!”. Well, ‘ OR 1=1 -- might not look like the kind of username we had in mind, but invalid? Says who?

If we have a look at the code, the signature of the authentication method says:

Integer authenticate(String username, String passwordMD5)

Basically, in the code there is nothing saying that username is any special kind of data – it is just a string. And, as such, it can be any string – including ’ OR 1=1 --.

There might be conventions, even documented such, that a username should have certain structure – but the model represented in the code consider any string to valid to send into the method.

The Domain Driven Design take on this is that if you have more restriction in your intended model, then you should better put those restrictions in the code – explicitly.

So, let us take a small step in that direction – let us make Username an explicit part of the model. Later on we can elaborate that part of the model by making restrictions on usernames explicit, and even enforcing them. But let us not take too big a bite – for now we settle for shaping up the model.

If we think about it we can surly agree that username is a special kind of data, separate from amounts, order numbers, or phone numbers. It would simply not make sense to have a phone number “+4615210000” used as a username. This is analogous to the distinction between int and boolean. Under the hood they are both “just bits and bytes”, but we want to keep them distinct in our language so that we do not accidentally use an int as the condition in an if-statement, for example. In C many hard-to-find bugs have been caused by that specific mistake.

In static typed programming languages like Java, C# or ML, we use the type system with interface and classes to separate different kinds of data. However, if we audit the authentication code we will see that there is no representation of username on that level. The only place “username” show up is as the name of a String-typed variables and parameters. The knowledge “username is a specific kind of data with its own rules and restrictions” is not explicit in the code.

Enter class Username, which at this stage might be the simplest kind of value object.

public class Username {
  public final String username; // final making it immutable
  public Username(String username) { this.username = username; }
}

The important part here is of course that we now have a new type, which can be used by variables, fields, parameters, and returns to make the code explicitly talk about usernames.

The authentication method will change somewhat.

  /** Authenticates a user with a given password.
   * @param username
   * @param passwordMD5 hash of password
   * @return user id, or null if no matching account
   */
  Integer authenticate(Username username, String passwordMD5)
          throws SQLException {
    Connection con = accountDs.getConnection();
    Statement stmt = con.createStatement();
    String sqlSelect = "SELECT uid FROM Accounts";
    String usernameMatch = "username = '" + username.username + "'";
    String passwdHashMatch = "passwdHash = '" + passwordMD5 + "'";
    String sql = sqlSelect +
            " WHERE " + usernameMatch +
            " AND " + passwdHashMatch;
    ResultSet rs = stmt.executeQuery(sql);
    Integer result;
    if(rs.next()) { // found account with matching password
      result = rs.getInt("uid");
    } else { // no matching account
      result = null;
    }
    return result;
  }

So, whoever wants to call the authentication method with a username, must first create a Username object via the constructor.

public class LoginAction {
  void doit() throws SQLException {
    Username username = new Username(form.username);
    String passwordMD5 = form.password;
    accountService.authenticate(username, passwordMD5);
  }
}

Now the concept of username is explicit throughout the code, and actually talks the same language as the people working with it. In effect, we have made username a part of the ubiquitous language talking about the system.

Note that we are still not yet protected from bad usernames, that will be a later step - but at least we talk about usernames, not strings.

The distinction between username strings and usernames is subtle. This distinction might seem small, but I think it is essential – as the language form how we think. The moment the programmer start expressing herself in domain terms (creating a Username object), chances are higher that she will also question the indata parameter: Is this string really a username? Where did it come from? Has it been properly checked? No guarantee, but chances are higher.

We still have some way to cover before we have an API that is both easy to use correctly, and hard to use incorrectly – but at least we have taken a step in that direction. We still lack the constraints on usernames, and there is no enforcement at all.

However, if we can guide the programmers into thinking about the model a la DDD, and thus decrease the risk of severe application security flaws, then we have at least done something useful.

And it is usefulness that is the ambition of Domain Driven Security.

Yours

Dan

PS My colleague John Wilander just published a nice example on how they did with Swedish "person number" (roughly social security number) [Swedish]