The Database Hacker's Handbook

John Wiley & Sons

Chapter One

In the introduction, we discussed the reasons why we consider database security to be important. In this chapter, we provide a brief overview of several broad categories of security issues, with a few specific details and some discussion of general defenses. We also briefly discuss how to go about finding security flaws in database systems. Before we do so, we should discuss some emerging trends in database security.

In recent years, with the explosion in web-based commerce and information systems, databases have been drawing ever closer to the network perimeter. This is a necessary consequence of doing business on the Web - you need your customers to have access to your information via your web servers, so your web servers need to have access to your databases. Databases that were previously accessible only via several insulating layers of complex business logic are now directly accessible from the much more fluid - and much less secure- web application environment. The result of this is that the databases are closer to the attackers. With the constant march toward a paperless business environment, database systems are increasingly being used to hold more and more sensitive information, so they present an increasingly valuable target. In recent years, database vendors have been competing with each other to provide the most feature-richenvironment they can, with most major systems supporting XML, web services, distributed replication, operating system integration, and a host of other useful features. To cap all of this, the legislative burden in terms of corporate security is increasing, with HIPAA, SOX, GLBA, and California Senate Bill No. 1386 imposing an ever-increasing pressure on companies to ensure that their networks are compliant.

So why care about database security? Because your databases are closer to the attacker, present a more valuable target, have more features to configure, and are more closely regulated than they have ever been before.

Which Database Is the Most Secure?

All of the databases we cover in this volume have had serious security flaws at some point. Oracle has published 69 security alerts on its "critical patch updates and security alerts" page - though some of these alerts relate to a large number of vulnerabilities, with patch 68 alone accounting for somewhere between 50 and 100 individual bugs. Depending on which repository you search, Microsoft SQL Server and its associated components have been subject to something like 36 serious security issues - though again, some of these patches relate to multiple bugs. According to the ICAT metabase, DB2 has had around 20 published security issues - although the authors of this book have recently worked with IBM to fix a further 13 issues. MySQL has had around 25 issues; Sybase ASE is something of a dark horse with a mere 2 published vulnerabilities. PostgreSQL has had about a dozen. Informix has had about half a dozen, depending on whose count you use.

The problem is that comparing these figures is almost entirely pointless. Different databases receive different levels of scrutiny from security researchers. To date, Microsoft SQL Server and Oracle have probably received the most, which accounts for the large number of issues documented for each of those databases. Some databases have been around for many years, and others are relatively recent. Different databases have different kinds of flaws; some databases are not vulnerable to whole classes of problems that might plague others. Even defining "database" is problematic. Oracle bundles an entire application environment with its database server, with many samples and prebuilt applications. Should these applications be considered a part of the database? Is Microsoft's MSDE a different database than SQL Server? They are certainly used in different ways and have a number of differing components, but they were both subject to the UDP Resolution Service bug that was the basis for the "Slammer" worm.

Even if we were able to determine some weighted metric that accounted for age, stability, scrutiny, scope, and severity of published vulnerabilities, we would still be considering only "patchable" issues, rather than the inherent security features provided by the database. Is it fair to directly compare the comprehensive audit capabilities of Oracle with the rather more limited capabilities of MySQL, for instance? Should a database that supports securable views be considered "more secure" than a database that doesn't implement that abstraction? By default, PostgreSQL is possibly the most security-aware database available - but you can't connect to it over the network unless you explicitly enable that functionality. Should we take default configurations into account? The list of criteria is almost endless, and drawing any firm conclusions from it is extremely dangerous.

Ultimately, the more you know about a system, the better you will be able to secure it - up to a limit imposed by the features of that system. It isn't true to say, however, that the system with the most features is the most secure because the more functionality a system has, the more target surface there is for an attacker to abuse. The point of this book is to demonstrate the strengths and weaknesses of the various database systems we're discussing, not - most emphatically not - to determine which is the "most secure."

In the end, the most secure database is the one that you know the most about.

The State of Database Security Research

Before we can discuss the state of database security research, we should first define what we mean by the term. In general, when we use the phrase "database security research" we tend to mean research into specific, practical flaws in the security of database systems. We do not mean research into individual security incidents or discussions of marketing-led accreditation or certification efforts. We don't even mean academic research into the underlying abstractions of database security, such as field-, row-, and object-level security, or encryption, or formal protocol security analysis - though the research we are talking about may certainly touch on those subjects. We mean research relating to discoveries of real flaws in real systems.

So with that definition in mind, we will take a brief tour of recent - and not so recent - discoveries, and attempt to classify them appropriately.

Classes of Database Security Flaws

If you read about specific security flaws for any length of time, you begin to see patterns emerge, with very similar bugs being found in entirely different products. In this section, we attempt to classify the majority of known database security issues into the following categories:

* Unauthenticated Flaws in Network Protocols

* Authenticated Flaws in Network Protocols

* Flaws in Authentication Protocols

* Unauthenticated Access to Functionality

* Arbitrary Code Execution in Intrinsic SQL Elements

* Arbitrary Code Execution in Securable SQL Elements

* Privilege Elevation via SQL Injection

* Local Privilege Elevation Issues

So we begin with arguably the most dangerous class of all - unauthenticated flaws in network protocols. By this we mean buffer overflows, format string bugs, and so on, in the underlying network protocols used by database systems.

Unauthenticated Flaws in Network Protocols

Arguably the most famous bug in this class is the bug exploited by the SQL Server "Slammer" worm. The SQL Server Resolution Service operates over a UDP protocol, by default on port 1434. It exposes a number of functions, two of which were vulnerable to buffer overflow issues (CAN-2002-0649). These bugs were discovered by David Litchfield of NGS. Another SQL Server problem in the same category was the "hello" bug (CAN-2002-1123) discovered by Dave Aitel of Immunity, Inc., which exploited a flaw in the initial session setup code on TCP port 1433.

Oracle has not been immune to this category - most recently, David Litchfield found an issue with environment variable expansion in Oracle's "extproc" mechanism that can be exploited without a username or password (CAN-2004-1363). Chris Anley of NGS discovered an earlier flaw in Oracle's extproc mechanism (CAN-2003-0634) that allowed for a remote, unauthenticated buffer overflow. Mark Litchfield of NGS discovered a flaw in Oracle's authentication handling code whereby an overly long username would trigger an exploitable stack overflow (CAN-2003-0095).

David Litchfield also found a flaw in DB2's JDBC Applet Server (no CVE, but bugtraq id 11401) that allows a remote, unauthenticated user to trigger a buffer overflow.

In general, the best way to defend yourself against this class of problem is first, to patch. Second, you should attempt to ensure that only trusted hosts can connect to your database servers, possibly enforcing that trust through some other authentication mechanism such as SSH or IPSec. Depending on the role that your database server is fulfilling, this may be tricky.

Another possibility for defense is to implement an Intrusion Detection System (IDS) or an Intrusion Prevention System (IPS). These kinds of systems have been widely discussed in security literature, and are of debatable value. Although an IDS can (sometimes) tell you that you have been compromised, it won't normally prevent that compromise from happening. Signature-based IDS systems are only as strong as their signature databases and in most cases signatures aren't written by people who are capable of writing exploits, so many loopholes in the signatures get missed.

"True anomaly" IDS systems are harder to bypass, but as long as you stick to a protocol that's already in use, and keep the exploit small, you can usually slip by. Although some IDS systems are better than others, in general you need an IDS like you need someone telling you you've got a hole in the head. IDS systems will certainly stop dumber attackers, or brighter attackers who were unlucky, so they may be worthwhile provided they complement - and don't replace - skilled staff, good lockdown, and good procedures.

IPS systems, on the other hand, do prevent some classes of exploit from working but again, every IPS system the authors have examined can be bypassed with a little work, so your security largely depends on the attacker not knowing which commercial IPS you're using. Someone may bring out an IPS that prevents all arbitrary code execution attacks at some point, which would be a truly wonderful thing. Don't hold your breath waiting for it, though.

Authenticated Flaws in Network Protocols

There are substantially fewer bugs in this category. This may reflect a reduced focus on remote, authenticated bugs versus remote, unauthenticated bugs among the security research community, or it may be sheer coincidence.

David Litchfield found a flaw in DB2 for Windows (CAN-2004-0795) whereby a remote user could connect to the DB2REMOTECMD named pipe (subject to Windows authentication) and would then be able to execute arbitrary commands with the privilege of the db2admin user, which is normally an "Administrator" account.

David discovered another flaw in DB2 in this category recently, relating to an attacker specifying an overly long locale LC_TYPE. The database applies this after the user authenticates, triggering the overflow.

There are several other bugs that debatably fall into this category, normally relating to web application server components; because we're focusing on the databases themselves we'll gloss over them.

In general the best way to protect yourself against this category of bugs is to carefully control the users that have access to your databases; a strong password policy will help - as long as you're not using plaintext authentication protocols (we discuss this more later). Auditing authenticated users is also a good idea for a number of reasons; it might give you a heads-up if someone is trying to guess or brute-force a password, and if you do have an incident, at least you have somewhere to start looking.

Flaws in Authentication Protocols

Several database systems have plaintext authentication protocols, by which we mean authentication protocols in which the password is passed "on the wire" in a plaintext or easily decrypted format. In a default configuration (that Sybase warns against, but which we have still seen in use) Sybase passes passwords in plaintext. By default, Microsoft SQL Server obfuscates passwords by swapping the nibbles (4-bit halves of a byte) and XORing with 0xA5. In both of these cases, the vendors warn against using the plaintext versions of their authentication protocols and provide strong, encrypted mechanisms that are relatively easy to deploy - but the defaults are still there, and still dangerous.

MySQL has historically had a number of serious problems with its authentication protocol. Although the protocol isn't plaintext, the mathematical basis of the authentication algorithm prior to version 4.1 was called into question by Ariel Waissbein, Emiliano Kargieman, Carlos Sarraute, Gerardo Richarte, and Agustin Azubel of CORE SDI (CVE-2000-0981). Their paper describes an attack in which an attacker that can observe multiple authentications is quickly able to determine the password hash.

A further conceptual problem with the authentication protocol in MySQL prior to version 4.1 is that the protocol only tests knowledge of the password hash, not the password itself. This leads to serious problems if a user is able to somehow determine another user's password hash - and MySQL has been subject to a number of issues in which that was possible.

Robert van der Meulen found an issue (CVE-2000-0148) in MySQL versions prior to 3.23.11 whereby an attacker could authenticate using only a single byte of the expected response to the server's challenge, leading to a situation whereby if you knew a user's username, you could authenticate as that user in around 32 attempts.

Chris Anley recently found a very similar problem in MySQL (CAN- 2004-0627) whereby a user could authenticate using an empty response to the server's challenge, provided he or she passed certain flags to the remote server.

This category of bugs is almost as dangerous as the "unauthenticated flaws in network protocols" category, because in many cases the traffic simply looks like a normal authentication. Attackers don't need to exploit an overflow or do anything clever, they simply authenticate without necessarily needing the password-or if they've been able to sniff the password, they just authenticate.

The best defense against this kind of bug is to ensure that your database patches are up-to-date, and that you don't have any plaintext authentication mechanisms exposed on your databases. If your DBMS cannot support encrypted authentication in your environment, you could use IPSec or SSH to provide an encrypted tunnel. MySQL provides explicit guidelines on how to do this in its documentation, though recent versions of MySQL allow authentication to take place over an SSL-encrypted channel.

Unauthenticated Access to Functionality

Some components associated with databases permit unauthenticated access to functionality that should really be authenticated. As an example of this, David Litchfield found a problem with the Oracle 8 and 9i TNS Listener, whereby a remote, unauthenticated user could load and execute an arbitrary function via the "extproc" mechanism (CVE-2002-0567). The function can have any prototype, so the obvious mode of attack is to load the libc or msvcrt library (depending upon the target platform) and execute the "system" function that allows an attacker to execute an arbitrary command line. The command will be executed with the privileges of the user that the database is running as - "oracle" on UNIX systems, or the local system user on Windows.

Recently, David Litchfield disclosed an issue that allows any local user to execute commands in the security context of the user that Oracle is running as (CAN-2004-1365). This bug works in exactly the same way as the bug listed earlier (CVE-2002-0567), except that it takes advantage of the implicit trust that extproc places in the local host. Oracle does not consider this to be a security issue (!) but we would caution you not to allow users to have shells on Oracle servers without seriously considering the security ramifications. Clearly, allowing a user to have a shell on a database server is dangerous anyway, but in this particular case there is a known, documented vector for attack that the vendor will not fix.

(Continues...)

---

Excerpted from The Database Hacker's Handbook by David Litchfield Excerpted by permission.
All rights reserved. No part of this excerpt may be reproduced or reprinted without permission in writing from the publisher.
Excerpts are provided by Dial-A-Book Inc. solely for the personal use of visitors to this web site.

---