Advanced SQL Injection Techniques
Advanced SQL Injection Techniques
Advanced SQL Injection Techniques
Mike Shema
SQL Injection attacks target the core of a web application: its database. Their most significant impact enables an attacker to retrieve, modify, or delete arbitrary data. It is a serious threat to any application with a database back-end and a threat that should be fully understood in order to develop adequate countermeasures. Every web server administrator must acknowledge techniques that can be used to identify an SQL Injection vulnerability (see Tobias Glemser's Article SQL Injection Attacks with PHP and MySQL, hakin9 03/2005) and assess the scope of its risk. The basic methodology for an SQL Injection attack is to identify a potential vector, then exploit that vector with customized SQL queries - all through the web browser. Identification of the potential for a vulnerability is important, but even more important is the ability to evaluate its impact. In some cases, a SQL Injection vector may offer nothing more than the capability to generate some syntax errors, such as trying to convert strings to numeric values. In other cases, the vector may enable the attacker to fully compromise a database's information. Although the examples refer to MySQL databases, the techniques apply to any database platform and, in most cases, can be applied without modification. The core of these techniques targets the SQL language. Certain database extensions merely make these techniques easier to accomplish.
attack the syntax of the query - insert common SQL characters with the intention of generating errors to identify potential attack vectors, attack the syntax of the language - target the SQL language itself in order to generate database errors or perform simple queries by manipulating language constructs and semantic identities, attack the logic of the query - rewrite the query to retrieve arbitrary data from tables to which developers did not intend access.
These techniques can be combined to assess a web application and determine its vulnerability to SQL Injection attacks. In the next sections the SQL Injection payloads are presented without the entire URL as an example. This makes it easier to understand the techniques without cumbersome parameters and text. This is also because the injection of these payloads is quite simple. Given a URL of the form http://site/page.cgi?a=foo&b=bar, a SQL Injection attack replaces the vulnerable parameter's value with its payload: http://site/page.cgi?a=&b=bar. As a further reminder, one has to remember to encode spaces and other characters in the payload so that they do not disrupt the syntax of the URL.
a a a a
= = = =
While the most common example is the single quote character (ASCII 0x27), many characters can be used to disrupt the syntax including:
Validation filters that only prohibit single quote characters (or some small set of characters) might prevent full exploitation of a vulnerability, but such filters are often inadequate. They may simply obscure more fundamental problems with the application's database connection architecture. Quotes vs. slashes PHP developers face several challenges and potentially confusing recommendations when creating strong input validation filters. PHP's magic_quotes() function automatically escapes all single quotes with a backslash character; however, if this feature is combined with a call to the strip_slashes() function, then the escape characters have been removed:
SELECT foo FROM bar WHERE a = '\''; - single quote escaped, SELECT foo FROM bar WHERE a = '''; - backslash stripped, query illformed.
The other danger of focusing on the single quote character is that developers may not be aware of the full range of characters and techniques available to an attacker for exploiting a SQL query. The attacker can combine SQL functions to generate errors in the syntax of a query. You can also use inherent SQL functions to generate errors. The SQL CHAR() function prints the ASCII equivalent of the argument. An attacker may be able to inject quote characters by using odd or even amounts of CHAR(0x27) strings (hexadecimal 0x27 represents the ASCII code for the single quote). This is important, because the attack consists of alphanumeric characters plus the parentheses. Consequently, monitoring input for quote characters will not catch or block the attack.
Variables may vary Database-related errors can also be generated by attacking variable types. This is most effective against numeric values, but is also successful against date or time variables. For example, here is a list of different values that you may try against parameters that expect decimal numbers:
8-, 16-, 32- and 64-bit values - 256, 65536, etc., integer overflows - 2^8 + 1, 2^16 +1, 2^32 + 1, or 2^64 + 1, unsigned vs. signed values - inserting negative values, floating-point overflows - for example 3.40282346638528860e+38, 1.79769313486231570e+308, alternate presentation - binary, octal, hexadecimal, or scientific notation.
These numeric attacks often succeed in generating errors because the variables used to track these values are not strongly typed. In PHP the parameter type of all $_REQUEST variables is a string. This means that, although you can perform arithmetic operations on variables ($a = 1; $a++), the actual type of the variable may be considered a numeric string. The variable may even be silently promoted from a number to a numeric string when the value would normally result in an overflow, inf (infinity), or NaN (not a number) equivalent. For example, PHP's is_numeric("1e308") function returns true (it is a number), but is_numeric("1e309") returns false - neither a number or numeric string because it is beyond the double float type that PHP supports. A variable must be set to numeric explicitly using the settype() function, but beware that large values may return a value of inf which can also lead to errors in the query if it is expecting numerals. Fighting the synonyms Robust input validation filters can be an effective countermeasures to these techniques, but they are not sufficient. Database errors and other exceptions should be trapped and prevented from being sent to the browser. Verbose error information tends to provide useful information for malicious users targeting a database. As we will see a bit later, input validation filters may be inadequate. For example, we have already seen that the value 1e309 is not a number (for most languages and SQL databases) and will generate an error in less secure applications. Yet 1e309 does not contain any characters that are normally malicious. It is a purely alphanumeric value. Note that SQL is a rich language that provides an attacker to create many synonomous permutations. For example, CHAR(0x27) is equivalent to ASCII(0x27) which can also be written as x'27. We focus on using the CHAR(0x27) string to avoid raw quotes in the payload, but the specifics of each test are highly mutable. This also implies that syntaxbased filtering - such as application-layer firewalls - must be very robust in order to prevent these attacks. In fact, the combination of alternate encoding schemes (URL encoding, Unicode) and creative SQL will bypass most pattern-matching filters. Remember, CHAR(0x27) is the same as cH%41r(0x68-0x41).
In SQL, Shakespeare's observation of roses might look like the decidedly unpoetic: SELECT name FROM roses WHERE scent='sweet'; Whether a rose might be called shoe, bumblebee, or clock, its sweet-smelling attribute remains unchanged. SQL provides a rich set of functions that can be used to create semanticly equivalent queries that look quite different textually. This capability enables an attacker to identify and exploit injection vulnerabilities even when the server does not reveal error information or similar output. While it is useful to break queries in order to find potential vulnerabilities, it is also profitable to attack the query using the semantics of built-in SQL functions. Thus, instead of attacking the parser of the application language (PHP, JSP, etc.), the attack focuses on the SQL language itself. This has the added benefit of not only identifying attack vectors, but also provides more information about the input validation filters used by the application. Another byproduct of this technique is the ability to perform blind SQL Injection attacks, or attacks that do not rely on error generation in order to identify or exploit. Numeric data types Numeric data types are the easiest candidates to test with this technique. Figure 1 shows the original example URL, while Figures 2 and 3 present modified addresses We are using an older, insecure version of FreznoShop online shopping system - releases newer than 1.4 branch are quite invulnerable. Consider the following list of name/value pairs:
= = = =
111, 0x6f, 0157 (octal representation), 110+1 (use 110%2b1 in practice because the + stands for a space
Figure 3. The same string modified with usage of MOD() function From a database's point of view, each one of these requests results in the same value: 111. Also notice that none of these rely on the single quote character. The first three look like numeric or alphanumeric strings, the next two have apparently innocuous characters for the addition and subtraction symbols, and the final three include parentheses and a comma. If input validation were to focus on stripping the single quote, then a vulnerable application would gain no benefit from such a countermeasure. Raw parameters This technique, which uses semantic doppelgangers, enables the user to identify SQL Injection vectors. If the result of each request is identical, then it can be assumed that the application engine has parsed the raw parameter value and inserted it into the underlying SQL query. For example, consider this query for a rowid: SELECT foo FROM table WHERE rowid = 110+1; The database calculates 110+1 = 111 before resolving the rest of the query, according to its order of operations. This bears the same result as the original query: SELECT foo FROM table WHERE rowid = 111;
Before we explain how to extend this attack to extract arbitrary data, let us first examine some other cases that can be used for error generation. Even though this technique does not require us to generate database errors, such information is useful to determine versions and names of tables or columns. If the application's input validation filters have stripped quote characters, but not trapped database errors, then we can target incorrect SQL function syntax. For example:
BIN(-1),
LIMIT a (this is useful because it does not require parentheses), MOD(0,a).
Of course, numeric values should also be tested for boundary conditions as mentioned in the previous section. Premature termination characters This technique lends itself to the creation of custom SQL queries. Such queries often do not require quote characters, but often require premature termination characters. Thus, a request might employ /* or -- in order to truncate additional, undesired statements. A string
Once again, we have consciously chosen to avoid using quote characters because they set off alarms or may be blocked. Yet this doesn't prevent us from creating complex strings. The REVERSE(), LEAST(), and GREATEST() functions only need parentheses and commas. The following examples are all semantically identical:
Countermeasures The best countermeasures for these attacks use input validation filters and strong data types when assigning user-supplied values to query parameters. Even though 0x27 is a valid
hexadecimal value, it should be prohibited by the application because the raw value contains a non-numeric character (or possibly silently coaxed into 27 decimal). Likewise, octal 0157 should either be denied because of the leading zero, or the leading zero could be stripped so the value becomes 157 decimal, which is merely a different row number. At the very least, developers should be aware of alternate bases and understand where they are interpreted: either in the application language or in the database. It's very easy to handle all user-supplied data as strings, but if the data are to be inserted into a query, then they should be explicitly assigned (cast) to the appropriate data type. For interpreted languages such as PHP, Perl, C#, or Visual Basic the assignment should be safe or generate a conversion error. If the web application uses a compiled language such as C or C+ +, then the type casting should be handled carefully and checked for exceptions (think of format-string attacks).
Then, one can flip the query and ensure that it succeeds in order to verify the injection vector - /*!32302+AND+1+*//* (it may be necessary to terminate the query). UNION SELECT Once a parameter has been identified as a vector for SQL Injection attacks, the next step is to determine the amount to which the database is exposed. This is accomplished by manipulating the logic of the original query. Most basic queries are of the form SELECT foo FROM bar WHERE a=b; in which the b of a=b clause is the parameter that can be manipulated. Consequently, the new query must consider the previous SELECT. The quickest technique is to use the UNION keyword. The UNION statement combines multiple SELECT statements and is supported by most databases. The basic form looks like SELECT foo FROM bar WHERE a=b UNION SELECT foo2 FROM bar2 WHERE c=d;. One useful UNION clause is to display the user name under which the database connection has been established. On MySQL you would do this with SELECT USER(). Inside a UNION clause the request might look like
SELECT text FROM articles WHERE id=0 UNION SELECT USER(); Several challenges present themselves when using UNION statements for SQL Injection attacks:
the UNION clause should terminate the query to ensure valid syntax - any additional logic must be truncated, UNION statements require matching column counts in each SELECT clause.
The first challenge is relatively easy to accomplish. Simply use one of the common terminators described in the previous section. This can be a comment delimiter (#, /*, --) in combination - if necessary - with a semicolon or single quote. Columns and bears The second challenge is not difficult to overcome, but requires a few iterative steps remniscent of Goldilocks and the three bears. The injected UNION clause will either have too few columns or too many of them - what you need is a number that is just right! If you can observe the database's error messages, then you'll see something like The used SELECT statements have a different number of columns. Column undercounts can be fixed by adding extra columns or column place-holders to the SELECT statement (see Figure 4). For example, consider the following statements:
user FROM mysql.user, 1,user FROM mysql.user, 1,1,user FROM mysql.user, user,user,user,user FROM mysql.user.
Figure 4. A successful UNION SELECT attack Each one of these queries is designed to grab the user name (or names) from the default mysql.user table. The number of columns increases from one to four in each example. In practice, it is better to repeat the column name to ensure that the value is displayed in the application. The first placeholder works, but it's hard to tell which column the web application will display. Column overcounts can be addressed by using the CONCAT statement. Overcounts occur when the first SELECT statement expects fewer columns than your custom query. The CONCAT statement resolves this by concatenating each column into a single string. Thus, multiple columns are reduced to a single column. For example: SELECT foo FROM table WHERE a=b UNION SELECT CONCAT(*) FROM mysql.user; This can be combined with the undercount technique when necessary: SELECT foo,bar FROM table WHERE a=b
UNION SELECT 1,CONCAT(*) FROM mysql.user; The major caveat is that any NULL value in one of the column results will cast the CONCAT string to NULL. Aim at rows Once you have matched column counts for the query, the next step is often to specify an arbitrary row to retrieve from a table. When the query returns multiple rows, often only the first one is displayed. To some degree, a good WHERE clause can help target specific rows, but only if the table's general structure (column names) is known before-hand. A much easier method uses offsets within the LIMIT clause. You can limit the result to one row by using LIMIT 1, but you can control which row is returned by adding the optional offset beginning with 0. For example:
SELECT foo FROM table WHERE a=b CONCAT(*) FROM mysql.user LIMIT SELECT foo FROM table WHERE a=b CONCAT(*) FROM mysql.user LIMIT SELECT foo FROM table WHERE a=b CONCAT(*) FROM mysql.user LIMIT
You can progress through the offsets until the query returns a NULL row. Unlike the previous examples of simple queries, it is necessary to place parentheses around the clause that contains the LIMIT statement. Otherwise it will be incorrectly applied to the entire query. Defence by statements The use of prepared statements (also known as parameterized queries) or stored procedures are effective countermeasures to these techniques because they separate the logic of the query from the data of the query. Consequently, injection attacks can corrupt the original SQL query, but will not be able to rewrite it in such a manner that arbitrary tables or data can be accessed. A potential drawback of prepared statements is that they require additional set-up within the application. This could lead to a performance degradation; however, such an impact may be minimal. The security gains are definitely good.
Our core idea is to identify a SQL Injection vulnerability via creative use of SQL formatting characters (syntax) or SQL functions (semantics), then exploit the vulnerability by attacking the SQL logic. Although it primarily focuses on numeric and string manipulation, other functions can be used (or rather misused) to generate errors for vulnerability identification:
Enumeration is another important part of SQL Injection; one that is beyond our scope here. Nevertheless, here are some simple queries that can be used to further determine information about a database:
SELECT USER(), SELECT SESSION_USER(), SELECT CURRENT_USER(), SELECT SYSTEM_USER(), SELECT SUBSTRING_INDEX(USER(),'@',1), SHOW CHARACTER SET, SELECT CURDATE(), SELECT CURTIME().
A more fundamental problem of SQL Injection is the lack of separation between the query's logic and data. The logic is defined by the developer and is expected to remain static. The data are collected from the user. When the data and logic intermingle,
such as using string concatenation to build queries, then user-supplied data can manipulate the logic of the query. This is the higher risk compared to input validation, because a modified query provides access to arbitrary data in the database. A formatting character maliciously inserted into a stored procedure may merely produce a database error instead of exposing the actual data. This is not meant to imply that input validation is not important; however, any countermeasure to these types of attacks should focus equally on query construction and execution. Without a comprehensive understanding of the different techniques that attackers employ against web applications, developers will not create effective countermeasures. From an assessment perspective, auditors who do not adequately investigate the scope of a SQL Injection vulnerability present an inaccurate view of the application's risk - and if testing only relies on injecting single quote characters, then the assessment may be useless. SQL Injection attacks can be executed with many different characters.