CORE SDI S.A. Security Advisory February 7th, 2001 SSH protocol 1.5 session key recovery vulnerability Date Published: 2001-02-07 Advisory ID: CORE-20010116 Bugtraq ID: 2344 CVE CAN: Not currently assigned. Title: Session Key recovery in SSH protocol 1.5 Class: Design/implementation error Remotely Exploitable: Yes Locally Exploitable: Yes Release Mode: USER RELEASE Vulnerability Description: SSH is a widely used client-server application for authentication and encryption of network communications. In order to ensure that all data exchanged between client and server is kept confidential a symmetric algorithm is used with a key obtained from the key exchange and authentication process done upon connection from the client to an SSH server. A would be attacker could obtain and store all the encrypted packets belonging to a specific client-server connection but that would provide no real value unless she is able to: . Decrypt them without having the session key used for the encryption This is equivalent to breaking the crypto algorithm used. or . Exploit some design or implementation problem on either client or server to obtain the session key and the proceed to decrypt the stored session using any implementation of the crypto algorithm used. This advisory describes a vulnerability in the SSH 1.5 protocol that allows an attacker to do the later. The key exchange in SSH protocol 1.5 uses PKCS#1_1.5 public key encryption standard to make the key exchange between client and server upon connection. An attack (see [1] and [2]) discovered by David Bleichenbacher on PKCS#1_1.5 can be exploited to recover arbitrary session keys. Combining Bleichenbacher's attack with a timing attack designed to obtain information about crypto operations performed on a SSH server it is possible to obtain a session key for an SSH session and therefore decrypt it or even alter it if it is still active. Vulnerable Packages/Systems: All versions of SSH supporting the protocol 1.5 key exchange. This vulnerability applies to SSH servers only. See the following section for vendor specific information. Solution/Vendor Information/Workaround: OpenSSH The vulnerability is present in OpenSSH up to version 2.3.0, although it is not possible to exploit it due to limits imposed on the number of simultaneous connections the server is allowed to handle, Nonetheless, Markus Friedl of OpenSSH.com has produced a patch that sets a random session key if RSA operations on the session key sent by the client fail. This effectively solves the problem by closing the oracle that leaks information. The patch was integrated to the OpenSSH source tree on January 29, 2001 AppGate The default configuration of the AppGate server is not vulnerable since it has SSH-1 support disabled. However it is possible for administrators to enable SSH-1 backwards compatibility to be able to use legacy clients. Those customers should apply the patches we have prepared. Patches can be downloaded from the AppGate support web or requested from support@appgate.com Mindbright The Mindbright ssh1-server is only an experimental product and we are not aware of anybody actually using it, it has never been sold or available as a separate entity. Since it is written in java it will need a really extreme machine to be able to handle the load needed to exploit this vulnerability. Anybody who feels that they need a patch for it is welcome to contact mindbright@mindbright.se. SSH.com ssh-1 up to version 1.2.31 is vulnerable. The official response from SSH.com follows: -SSH1 is deprecated and SSH.com does not support it anymore, the official response is upgrade to SSH2 -The SSH1 compatibility code built into SSH-2.4.0 always executes a fresh copy of SSHD1, which causes the server key to be regenerated for every connection. Thus, the attack is not at all feasible when using SSH1 with an SSH2 server in compatibility mode. Ssh-2.4.0 also includes code for limiting the maximum number of simultaneous connections. The maximum is controlled by the MaxConnections flag in /etc/ssh2/sshd2_config or with the --with-ssh-connection-limit= compile-time configure option. However, as noted, the limit is not required for protection when using SSH1 with SSHD2 in compatibility mode. -The following unsupported and untested patch can be applied to ssh-1.2.31 and earlier. It addresses the problem by regenerating the server key when the RSA operations fail. This is done at a rate of at most one key regeneration per minute to avoid possible DoS attacks. -------------- cut here ---------------------------------------------- --- rsaglue.c 1999/12/10 23:27:25 1.8 +++ rsaglue.c 2001/02/03 09:42:05 @@ -264,7 +268,15 @@ mpz_clear(&aux); if (value[0] != 0 || value[1] != 2) - fatal("Bad result from rsa_private_decrypt"); + { + static time_t last_kill_time = 0; + if (time(NULL) - last_kill_time > 60 && getppid() != 1) + { + last_kill_time = time(NULL); + kill(SIGALRM, getppid()); + } + fatal("Bad result from rsa_private_decrypt"); + } for (i = 2; i < len && value[i]; i++) ; -------------- cut here --------------------------------------------- LSH Not vulnerable. Does not support protocol version 1 Cisco Systems, F-Secure, other SSH server vendors No information provided. Additionally, advisories and information on security issues in SSH can be obtained from: http://www.core-sdi.com/advisories/buffer_over_ing.htm http://www.core-sdi.com/advisories/ssh-advisory.htm http://www.core-sdi.com/bid/1949 http://www.core-sdi.com/bid/1426 http://www.core-sdi.com/bid/1323 http://www.core-sdi.com/bid/1006 http://www.core-sdi.com/bid/843 http://www.core-sdi.com/bid/660 Vendor notified on: 2001-01-16 Credits: This vulnerability was found and researched by Ariel Waissbein and Agustin Azubel of CORE SDI, Buenos Aires, Argentina. This advisory was drafted with the help of the SecurityFocus.com Vulnerability Help Team. For more information or assistance drafting advisories please mail vulnhelp@securityfocus.com. This and other CORE SDI security advisories can be obtained from: http://www.core-sdi.com/english/publications.html Technical Description - Exploit/Concept Code: In Section 1 we introduce the SSH1 key exchange, in Section 2 we introduce the attack, finally in Section 3 we discuss the attack's feasibility and argue why it is insecure to continue using this protocol. 1) SSH1 KEY-EXCHANGE PROTOCOL DESCRIPTION: 1.1.- The keys. Each host has a host unique permanent RSA key set which identifies it. A host is a SSH server (referenced as server), which runs the 'sshd' daemon or a SSH client (referenced as client) which runs the 'ssh' client program. The length of the host key is by default 1024 bits. Each server has its own server RSA key set which is automatically generated after a specified timeout (1 hour by default). This key set is never saved in any file. The length of this key is by default 768 bits. In every client-to-server connection, a 256 bits session key is generated by the client using pseudo-random data provided by the same client. This session key will be used in a symmetric algorithm (e.g. DES, Blowfish, 3DES) to encrypt the data flow on the connected channel after the key exchange is completed. To send the session key over an insecure channel to the server, it is encrypted by the client with the server key and the server host key together with other data using an asymmetric encryption algorithm (RSA-PKCS #1 1.5) as we explain in Subsection 1.4. The purpose of the two separate server keys is to make it impossible to decrypt a captured session by breaking into the server machine and getting access to the server key at a later time; one hour after the connection start not even the server machine can decipher the session key! 1.2.- Initiating a connection. Whenever a client connects to the server, the daemon forks. The parent stays in a loop waiting to accept more connections and the child manages the accepted connection. Before authenticating both endpoints, they do an identification exchange. 1.3.- The identification exchange. First, the server sends a formatted string to the client in plaintext, specifying the protocol supported versions and the server version. This string looks like "SSH-1.99-OpenSSH_2.3.0", where "1" denotes the protocol version major number, "99" the protocol version minor number and "OpenSSH_2.3.0" is the software version of the server. If the client does not support the received protocol, it closes the connection. If the protocol is supported by the client, it responds with a formatted string of the same plaintext format. The server then checks the client's response. If the versions do not match or the client version is not valid, the server closes the connection. If the versions do match, the key exchange is started. 1.4.- The key exchange. The server will send both of its public keys. First the server will fetch 64 bits from a PRNG, that will be used as a cookie to prevent IP spoofing attacks and TCP sequence number prediction. This only affects rhosts authentication. The client must send back this cookie when the session key is sent. This only works against somebody doing IP spoofing from a remote network; any machine on the local network can still see outgoing packets and catch the random cookie. The server then builds a packet of type SSH_SMSG_PUBLIC_KEY, concatenating the cookie, the size of the 'n' component of the RSA server key, the 'e' public exponent of the RSA server key and the modulus 'n' of the RSA server key (the public RSA server key), the size of the 'n' component of the RSA host key, the 'e' public exponent of the RSA host key and the modulus 'n' of the RSA host key the public RSA host key), the SSH protocol flags, the supported symmetric ciphers, and the supported authentication methods. Once the client has received the SSH_SMSG_PUBLIC_KEY packet, it computes a session ID in the same way the server does: [mpaux.c:compute_session_id()] The session ID is equal to a MD5 hash of the concatenation of the modulus of the host key of the server, the modulus of the server key and the server generated cookie. session_id := MD5(HostKey_RSAModulus||ServerKey_RSAModulus||Cookie) The length of a session_id is the same as the output of the MD5 function: 128 bits. The client generates a session key of 256 bits fetching data from a PRNG. This key will be the used in a symmetric algorithm to encrypt all the future flow of this SSH session. Before this key is encrypted and sent, the first 128 bits of this key, are XORed with the session_id. The client then uses the RSA algorithm (PKCS1 1.5) to encrypt consecutively the XORed session key and session_id with the server key and host key. Encryption is made using the smaller key first. Finally the client builds a packet containing the symmetric algorithm to use, the received cookie, the encrypted session key and the SSH protocol flags and sends it to the server. The server receives this packet and retrieves the symmetric algorithm chosen by the client and checks its compliance sending a "Warning: client selects unsupported cipher." message if it is not. It then checks that the received cookie matches the old cookie sent, sending another error message if it is not. It retrieves the encrypted key, processes the SSH algorithm flags and decrypts the session key (OpenSSH does an integrity check on the packet lenght before this). We explain this in detail since it is of great interest for our attack. To do this we introduce the PKCS #1 1.5 encoding. 1.5 - PKCS#1 - 1.5 (from rsaglue.c in ssh-1.2.30) To send a message m using a RSA public exponent e, with a public modulus n, the encrypter encodes the message m as M := 0x00 || 0x02 || P || 0x00 || m where 0x00 and 0x02 are the value of the first 2 bytes in hexa, and P is an hexadecimal padding string containing no zero octets. The ciphertext is: c := M^e mod (n) . (i.e. M to the e-th power modulo n) To recover m, the decrypter calculates c^d, where d is the private exponent, checks whether the first two bytes are 0x00 and 0x02 and calls the function fatal() in log-server.c closing the connection if the check failed. Otherwise it sets all the data after the second zero as the message (in case the format is correct this will return m). OpenSSH uses OpenSSL which behaves different, see the RSA_padding_check_PCKS1_type2() function for more details. The cleartext for this session key is recovered by checking which is the bigger public modulus and decrypting first with the key corresponding to the bigger modulus and secondly with the smaller one (in case of a tie the server key goes first). This is done using the rsa_private_decrypt() function. When this is done the server computes the session key, and does a XOR of the decrypted data with the computed session id to obtain the session key generated by the client. Finally the server sets the symmetric encryption scheme and key to the ones chosen by the client, and sends a packet describing the success to the client. This packet is the first encrypted packet of the flow secured by the symmetric algorithm. 2) ATTACK DESCRIPTION. 2.1.- Bleichenbacher's attack. Daniel Bleichenbacher presented an adaptive ciphertext attack to RSA encryption standard PKCS1_1.5 at the Crypto 98 Conference ([1]), which on input of a ciphertext c, outputs the cleartext m corresponding to this ciphertext. To carry out this attack the attacker needs to make use of a decryption oracle. As we shall see, this is automatically provided by the RSA functions used in SSH1 ( or in the OpenSSL library used in OpenSSH). This is the protocol flaw that enables the attack we present. Specifically, an attacker needs only to access an oracle that will answer if a string c' calculated by her is or is not PKCS#1_1.5-format compliant, even less, she only needs to know if it holds true that the hexadecimal representation of the string (c')^d mod (n) starts with the octets 0x00 and 0x02 (here d is the private secret exponent and n the public modulus). To decrypt a ciphertext without the private key, the attacker needs to access to this oracle 2^{20} times (average-time complexity). This estimation holds true for a 1024 bit key size. We shall not explain the attack in detail. To decrypt a ciphertext c an attacker will need to access the oracle with messages of the type c.s^e mod (n) where e and n are the public exponent and public modulus, and s is chosen by the attacker algorithm following certain rules. We refer to the paper [1] for further details. In each step of the attack, the attacker finds a collection of intervals in which the cleartext is contained, first starting with a big interval of size 2^{1018}=2^{1024-16} and reducing it until a single interval of size one - whose only member is the cleartext- is left. 2.2.- The attack on SSH-1 Suppose that we are sniffing a connection between a client and the server. We can then easily detect when this connection starts and get the packet containing the encrypted session key. We can then work in parallel, saving all successive packets exchanged between server and client, and at the same time attempt a session key decryption with the attack we present. Once the session key is decrypted all the saved encrypted packets sent between this client and the server can be decrypted in a straight-forward manner. To obtain the session key we will make use of Bleichenbacher's attack together with a simple timing attack technique. Let c := E_{K1}(E_{K2}(K)) denote the captured ciphertext, where K1 and K2 are the server and host key (the order of these keys does not alter the way in which the attack is made, since the order can be easily deduced as we explain in the following section, we suppose without loss of generality that K1 is the host key and K2 is the server key), K is the session key or rather the plaintext string containing the session key, and E_{A}(B) denotes RSA-PKCS1_1.5 encryption of the cleartext B using the Public key A. The attack is divided in two main steps, Firstly the attacker will first attempt to recover E_{K2}(K) from c using a plain Bleichenbacher attack, and secondly K is calculated by the attacker from E_{K2}(K) using a reduction we explain in the next subsection together with Bleichenbacher's attack. Notice that the calls to the function fatal() can be used as the needed oracle. Successful negotiation of a session key will end with the reception of a SSH_SMSG_SUCCESS packet at the client. A failure will end with the connection being shutdown due to the calls to the fatal() function from within the rash_private_decrypt() function. An attacker can -prior to the attack- determine what is the time needed for the server to reach the connection shutdown call in the fatal() function if the first encryption is not format compliant, and what is the time needed for the server to reach it if the first encryption is format compliant and the second encryption is not. This is basically the way of retrieving answers from the oracle and it implies a timing attack as well as a few modifications to Bleichenbacher's attack. To carry out the attack and recover the session key the host key needs to remain the same during the attack, we suppose that this is the case and shall discuss the feasibility of this in the following section. Suppose now that E_{K2}(K) is already calculated and known to the attacker, and call c':=E_{K2}(K). The attacker then uses c' to recover K. To do this, instead of accessing the oracle with messages of the form c.s^e mod (n), she will access the oracle with messages of the form c'.s'^{e'} mod (n'), where c' is defined as c':=E_{K2}(K), and e' and n'are the second public exponent and modulus (corresponding to E_{K1}(-)), and s' is chosen following the same rules as defined by Bleichenbacher's attack. 3. Implementation and Feasibility The estimation for the number of times needed to access the oracle on a adaptive ciphertext Bleichenbacher's attack for a 1024 bits modulus is approximately 2^{20}, as we said before. This means that the server should handle about 2^{20} connections to make the first decryption, i.e. to get E_{K2}(K). After this is done, to recover K, another adaptive ciphertext attack of the same sort should be carried out, with presumably less accesses to the oracle --say 2^{19}-- since the second key is smaller than the first one, to recover K. Hence, to carry out the attack we present here, an attacker should perform around (2^{20}+2^{19}) connections to the server during the lifespan of a server key K (default is one hour) which implies a rate of oracle queries of around 400/sec. Limiting the number of simultaneous connections to the server will greatly reduce the feasibility of this attack, this is in fact a standard feature in at least the OpenSSH implementation of SSH-1. It is necessary to note that the attacker also needs to perform crypto operations (RSA encryptions with a small exponent) for each query during the attack but those are computationally cheaper the ones performed on the server side. This seems to make our attack infeasible for most cases. nonetheless, high end servers are still a possible target for this attack. It is also worth mentioning that the number of connections given is for the average case and specifics cases will fall below the average. We follow to discuss other vulnerable cases in which our attack becomes feasible. An issue to be taken into account is the order of the keys K1 and K2, that is whether K1 is the server key and K2 the host key, or the other way around. This issue, we deferred to this section, is of some importance to our attack. As we mentioned the order of the keys is changed to K2 for the host key, and K1 for the server key in case the size of K2 is strictly greater than K1. In that case, the attacker has limited time to recover E_{K2}(K)(because K1 has a default timeout of one hour), but has an indefinite amount of time to recover K from E_{K2}(K). This would make the attack easier since it reduces the initial recovery attack to 2^{20} oracle queries within an hour. The second phase could be done at a much slower connection rate. It might also happen that the public modulus n is much smaller than the specified values, and this lucky stroke would speed up the attack considerably. Another issue to be taken into account, is when the default settings for the server key timeout are changed increasing the key lifespan and thus the time window for the attack. It is not likely, however, that the default settings for the key size will be purposely reduced. There is also a technology or rather server efficiency issue to be taken into account. Although the average case of the attack we present seems infeasible today, this might not be the case for specific attacks that deviates from the average or for specific attack scenarios en the present or the near future. The conclusion of this report is that although the attack described might not be a direct threat to the wide audience that relies on SSH1 for secure network communications, there is, nonetheless an exploitable flaw in the SSH-1 key exchange protocol that should be either fixed or addressed during the deployment of SSH as a security component. References [1] Daniel Bleichenbacher, "Chosen ciphertext attacks on RSA encryption standard PKCS #1", Advances in Cryptology, CRYPTO 98. Springer. [2] Daniel Bleichenbacher, Burt Kaliski and Jessica Staddon, "Recent results on PKCS#1: RSA encryption standard ". RSA Laboratories' Bulletin 7. http://www.rsa.com/rsalabs DISCLAIMER: The content of this advisory are copyright (c) 2001 CORE SDI Inc. and may be distributed freely provided that no fee is charged for this distribution and proper credit is given.