OBJ05-J. Limit extensibility of classes and methods with invariants to trusted subclasses only

Many methods offer invariants, which are guarantees made about what the method can do, and what state the object must be in when the method completes. For instance, the % operator, which computes the remainder of a number provides the invariant that

0 <= abs(a % b) <= abs(b), for all integers a, b where b != 0

Many classes also offer invariants, which are guarantees made about the state of their objects' fields upon the completion of all their methods. For instance, classes whose member fields may not be modified once they have assumed a value are called immutable classes. An important consequence of immutability is that the invariants of instances of these classes are preserved throughout their lifetimes.

A fundamental principle of object-oriented design is that when a subclass extends a superclass, the subclass's methods should preserve the invariants provided by the superclass. Unfortunately, attackers are not constrained by design principles. They can construct malicious classes which extend benign classes and provide methods that violate the invariants of the benign classes.

For instance, if an immutable class is not declared final, it might be possible to write a malicious subclass capable of changing the state of the immutable object. It would definitely be possible for the malicious subclass to impersonate the immutable object without actually being immutable. Such malicious subclasses would be capable of violating program invariants on which clients depend, thus introducing security vulnerabilities.

Therefore, classes with invariants that other code depends on ought to be declared final, thereby preventing malicious subclasses from subverting their invariants. Furthermore, immutable classes must be declared final.

Some classes (hereafter known as superclasses) must permit extension by trusted subclasses while simultaneously preventing extension by untrusted code. Therefore, declaring such superclasses to be final is not an option, because it would prevent extension by trusted code. Such problems require careful design for inheritance.

Consider two classes belonging to different protection domains; one is malicious, and estends the other, which is trusted. Consider an object of the malicious subclass with a fully qualified invocation of a method defined by the trusted superclass, and not overridden by the malicious class. In this case, the trusted superclass's permissions are examined to execute the method, with the consequence that the malicious object gets the method invoked inside the protection domain of the trusted superclass. [[Gong 2003]].

One commonly suggested solution is to place code at each point where the superclass can be instantiated to ensure that the instance being created has the same type as the superclass. When the type is found to be that of a subclass instead of the superclass's type, the checking code performs a security manager check to ensure that malicious classes cannot misuse the superclass. This approach is insecure because it allows a malicious class to add a finalizer and obtain a partially-initialized instance of the superclass. This attack is detailed in guideline OBJ04-J. Do not allow access to partially initialized objects.

For non-final classes, the method that performs the security manager check must be invoked as an argument to a private constructor to ensure that the security check is performed before any superclass's constructor can exit.

When the parent class has members that are declared private or are otherwise inaccessible to the attacker, the attacker must use reflection to exploit those members of the parent class. Declaring the parent class or its methods final prohibits this level of access.

A method which receives an untrusted, non-final input argument must beware that other methods or threads might modify the input object. Some methods try to prevent modification by making a local copy of the input object. This does not provide sufficient protection, as a shallow copy of an object may still allow it to point to mutable sub-objects, which may still be modified by other methods or threads. Some methods go farther and perform a deep copy of the input object. This mitigates the problem of modifiable sub-objects, but the method might still be passed a mutable object that estends the input object class.

Noncompliant Code Example

In this noncompliant code example, a malicious class can extend the public non-final parent class, NonFinal. Consequently, the attacker can invoke any of the parent class's accessible instance methods, can access the parent class's protected fields, and can even override any of the parent class's accessible non-final methods.

public class NonFinal {
  public NonFinal() {
    // ...   
  }
}

Noncompliant Code Example (Security Manager)

This noncompliant code example installs a security manager check in the constructor of the non-final parent class. The security manager denies access when it detects that a subclass without the requisite permissions is attempting to instantiate the superclass [[SCG 2007]]. It also compares class types, in compliance with [OBJ06-J. Compare classes and not class names].

public class NonFinal {
  public NonFinal() {
    Class c = getClass();  // java.lang.Object.getClass(), which is final
    // Confirm class type
    if (c != NonFinal.class) {
      // Check the permission needed to subclass NonFinal
      securityManagerCheck(); // throws a security exception if not allowed
    }
  // ...   
  }
}

Unfortunately, throwing an exception from the constructor of a non-final class is insecure because it allows a finalizer attack (see guideline OBJ04-J. Do not allow access to partially initialized objects).

Compliant Solution (Java SE 6+, public and private constructors)

This compliant solution invokes a security manager check as a side-effect of computing the boolean value passed to a private constructor (as seen in guideline OBJ04-J. Do not allow access to partially initialized objects). The rules for order of evaluation require that the security manager check must execute before invocation of the private constructor. Consequently, the security manager check also executes before invocation of any superclass's constructor. Note that the security manager check is made without regard to whether the object under construction has the type of the parent class or the type of a subclass (whether trusted or not).

This solution thwarts the finalizer attack. It is specific to Java SE 6 and onwards, where a finalizer is prevented from being executed when an exception is thrown before the java.lang.Object constructor exits [[SCG 2009]].

public class NonFinal {
  public NonFinal() {
    this( check( this.getClass())); // throws a security exception if not allowed
  }
  
  private NonFinal(boolean securityManagerCheck) {
    // regular construction goes here
  }

  private static boolean check(Class c) {
    // Confirm class type
    if (c != NonFinal.class) {
      // Check the permission needed to subclass NonFinal
      securityManagerCheck(); // throws a security exception if not allowed
    }

    return true;
  }
}

Risk Assessment

Permitting a non-final class or method to be inherited without checking the class instance allows a malicious subclass to misuse the privileges of the class.

Related Vulnerabilities

Search for vulnerabilities resulting from the violation of this guideline on the CERT website

Related Vulnerabilities

Search for vulnerabilities resulting from the violation of this guideline on the CERT website.

Bibliography

[[Bloch 2008]] Item 1: "Consider static factory methods instead of constructors"
[[Gong 2003]] Chapter 6: "Enforcing Security Policy"
[[Lai 2008]]
[[McGraw 2000]] Chapter Seven Rule 3: "Make Everything Final, Unless There's a Good Reason Not To"[[SCG 2007]] Guideline 1-2 "Limit the extensibility of classes and methods"
[[SCG 2009]]

OBJ04-J. Do not allow access to partially initialized objects      04. Object Orientation (OBJ)      OBJ06-J. Compare classes and not class names