CHANGES

  *) Fix ASN1 checks. Check for overflow by comparing with LONG_MAX
     and get fix the header length calculation.
     [Florian Weimer <Weimer@CERT.Uni-Stuttgart.DE>,
	Alon Kantor <alonk@checkpoint.com> (and others),
	Steve Henson]

  *) Use proper error handling instead of 'assertions' in buffer
     overflow checks added in 0.9.6e.  This prevents DoS (the
     assertions could call abort()).
     [Arne Ansper <arne@ats.cyber.ee>, Bodo Moeller]

 Changes between 0.9.6d and 0.9.6e  [30 Jul 2002]

  *) Add various sanity checks to asn1_get_length() to reject
     the ASN1 length bytes if they exceed sizeof(long), will appear
     negative or the content length exceeds the length of the
     supplied buffer.
     [Steve Henson, Adi Stav <stav@mercury.co.il>, James Yonan <jim@ntlp.com>]

  *) Fix cipher selection routines: ciphers without encryption had no flags
     for the cipher strength set and where therefore not handled correctly
     by the selection routines (PR #130).
     [Lutz Jaenicke]

  *) Fix EVP_dsa_sha macro.
     [Nils Larsch]

  *) New option
          SSL_OP_DONT_INSERT_EMPTY_FRAGMENTS
     for disabling the SSL 3.0/TLS 1.0 CBC vulnerability countermeasure
     that was added in OpenSSL 0.9.6d.

     As the countermeasure turned out to be incompatible with some
     broken SSL implementations, the new option is part of SSL_OP_ALL.
     SSL_OP_ALL is usually employed when compatibility with weird SSL
     implementations is desired (e.g. '-bugs' option to 's_client' and
     's_server'), so the new option is automatically set in many
     applications.
     [Bodo Moeller]

  *) Changes in security patch:

     Changes marked "(CHATS)" were sponsored by the Defense Advanced
     Research Projects Agency (DARPA) and Air Force Research Laboratory,
     Air Force Materiel Command, USAF, under agreement number
     F30602-01-2-0537.

  *) Add various sanity checks to asn1_get_length() to reject
     the ASN1 length bytes if they exceed sizeof(long), will appear
     negative or the content length exceeds the length of the
     supplied buffer. (CVE-2002-0659)
     [Steve Henson, Adi Stav <stav@mercury.co.il>, James Yonan <jim@ntlp.com>]

  *) Assertions for various potential buffer overflows, not known to
     happen in practice.
     [Ben Laurie (CHATS)]

  *) Various temporary buffers to hold ASCII versions of integers were
     too small for 64 bit platforms. (CVE-2002-0655)
     [Matthew Byng-Maddick <mbm@aldigital.co.uk> and Ben Laurie (CHATS)>

  *) Remote buffer overflow in SSL3 protocol - an attacker could
     supply an oversized session ID to a client. (CVE-2002-0656)
     [Ben Laurie (CHATS)]

  *) Remote buffer overflow in SSL2 protocol - an attacker could
     supply an oversized client master key. (CVE-2002-0656)
     [Ben Laurie (CHATS)]

 Changes between 0.9.6c and 0.9.6d  [9 May 2002]

  *) Fix crypto/asn1/a_sign.c so that 'parameters' is omitted (not
     encoded as NULL) with id-dsa-with-sha1.
     [Nils Larsch <nla@trustcenter.de>; problem pointed out by Bodo Moeller]

  *) Check various X509_...() return values in apps/req.c.
     [Nils Larsch <nla@trustcenter.de>]

  *) Fix BASE64 decode (EVP_DecodeUpdate) for data with CR/LF ended lines:
     an end-of-file condition would erronously be flagged, when the CRLF
     was just at the end of a processed block. The bug was discovered when
     processing data through a buffering memory BIO handing the data to a
     BASE64-decoding BIO. Bug fund and patch submitted by Pavel Tsekov
     <ptsekov@syntrex.com> and Nedelcho Stanev.
     [Lutz Jaenicke]

  *) Implement a countermeasure against a vulnerability recently found
     in CBC ciphersuites in SSL 3.0/TLS 1.0: Send an empty fragment
     before application data chunks to avoid the use of known IVs
     with data potentially chosen by the attacker.
     [Bodo Moeller]

  *) Fix length checks in ssl3_get_client_hello().
     [Bodo Moeller]

  *) TLS/SSL library bugfix: use s->s3->in_read_app_data differently
     to prevent ssl3_read_internal() from incorrectly assuming that
     ssl3_read_bytes() found application data while handshake
     processing was enabled when in fact s->s3->in_read_app_data was
     merely automatically cleared during the initial handshake.
     [Bodo Moeller; problem pointed out by Arne Ansper <arne@ats.cyber.ee>]

  *) Fix object definitions for Private and Enterprise: they were not
     recognized in their shortname (=lowercase) representation. Extend
     obj_dat.pl to issue an error when using undefined keywords instead
     of silently ignoring the problem (Svenning Sorensen
     <sss@sss.dnsalias.net>).
     [Lutz Jaenicke]

  *) Fix DH_generate_parameters() so that it works for 'non-standard'
     generators, i.e. generators other than 2 and 5.  (Previously, the
     code did not properly initialise the 'add' and 'rem' values to
     BN_generate_prime().)

     In the new general case, we do not insist that 'generator' is
     actually a primitive root: This requirement is rather pointless;
     a generator of the order-q subgroup is just as good, if not
     better.
     [Bodo Moeller]
 
  *) Map new X509 verification errors to alerts. Discovered and submitted by
     Tom Wu <tom@arcot.com>.
     [Lutz Jaenicke]

  *) Fix ssl3_pending() (ssl/s3_lib.c) to prevent SSL_pending() from
     returning non-zero before the data has been completely received
     when using non-blocking I/O.
     [Bodo Moeller; problem pointed out by John Hughes]

  *) Some of the ciphers missed the strength entry (SSL_LOW etc).
     [Ben Laurie, Lutz Jaenicke]

  *) Fix bug in SSL_clear(): bad sessions were not removed (found by
     Yoram Zahavi <YoramZ@gilian.com>).
     [Lutz Jaenicke]

  *) Add information about CygWin 1.3 and on, and preserve proper
     configuration for the versions before that.
     [Corinna Vinschen <vinschen@redhat.com> and Richard Levitte]

  *) Make removal from session cache (SSL_CTX_remove_session()) more robust:
     check whether we deal with a copy of a session and do not delete from
     the cache in this case. Problem reported by "Izhar Shoshani Levi"
     <izhar@checkpoint.com>.
     [Lutz Jaenicke]

  *) Do not store session data into the internal session cache, if it
     is never intended to be looked up (SSL_SESS_CACHE_NO_INTERNAL_LOOKUP
     flag is set). Proposed by Aslam <aslam@funk.com>.
     [Lutz Jaenicke]

  *) Have ASN1_BIT_STRING_set_bit() really clear a bit when the requested
     value is 0.
     [Richard Levitte]

  *) [In 0.9.6d-engine release:]
     Fix a crashbug and a logic bug in hwcrhk_load_pubkey().
     [Toomas Kiisk <vix@cyber.ee> via Richard Levitte]

  *) Add the configuration target linux-s390x.
     [Neale Ferguson <Neale.Ferguson@SoftwareAG-USA.com> via Richard Levitte]

  *) The earlier bugfix for the SSL3_ST_SW_HELLO_REQ_C case of
     ssl3_accept (ssl/s3_srvr.c) incorrectly used a local flag
     variable as an indication that a ClientHello message has been
     received.  As the flag value will be lost between multiple
     invocations of ssl3_accept when using non-blocking I/O, the
     function may not be aware that a handshake has actually taken
     place, thus preventing a new session from being added to the
     session cache.

     To avoid this problem, we now set s->new_session to 2 instead of
     using a local variable.
     [Lutz Jaenicke, Bodo Moeller]

  *) Bugfix: Return -1 from ssl3_get_server_done (ssl3/s3_clnt.c)
     if the SSL_R_LENGTH_MISMATCH error is detected.
     [Geoff Thorpe, Bodo Moeller]

  *) New 'shared_ldflag' column in Configure platform table.
     [Richard Levitte]

  *) Fix EVP_CIPHER_mode macro.
     ["Dan S. Camper" <dan@bti.net>]

  *) Fix ssl3_read_bytes (ssl/s3_pkt.c): To ignore messages of unknown
     type, we must throw them away by setting rr->length to 0.
     [D P Chang <dpc@qualys.com>]

 Changes between 0.9.6b and 0.9.6c  [21 dec 2001]

  *) Fix BN_rand_range bug pointed out by Dominikus Scherkl
     <Dominikus.Scherkl@biodata.com>.  (The previous implementation
     worked incorrectly for those cases where  range = 10..._2  and
     3*range  is two bits longer than  range.)
     [Bodo Moeller]

  *) Only add signing time to PKCS7 structures if it is not already
     present.
     [Steve Henson]

  *) Fix crypto/objects/objects.h: "ld-ce" should be "id-ce",
     OBJ_ld_ce should be OBJ_id_ce.
     Also some ip-pda OIDs in crypto/objects/objects.txt were
     incorrect (cf. RFC 3039).
     [Matt Cooper, Frederic Giudicelli, Bodo Moeller]

  *) Release CRYPTO_LOCK_DYNLOCK when CRYPTO_destroy_dynlockid()
     returns early because it has nothing to do.
     [Andy Schneider <andy.schneider@bjss.co.uk>]

  *) [In 0.9.6c-engine release:]
     Fix mutex callback return values in crypto/engine/hw_ncipher.c.
     [Andy Schneider <andy.schneider@bjss.co.uk>]

  *) [In 0.9.6c-engine release:]
     Add support for Cryptographic Appliance's keyserver technology.
     (Use engine 'keyclient')
     [Cryptographic Appliances and Geoff Thorpe]

  *) Add a configuration entry for OS/390 Unix.  The C compiler 'c89'
     is called via tools/c89.sh because arguments have to be
     rearranged (all '-L' options must appear before the first object
     modules).
     [Richard Shapiro <rshapiro@abinitio.com>]

  *) [In 0.9.6c-engine release:]
     Add support for Broadcom crypto accelerator cards, backported
     from 0.9.7.
     [Broadcom, Nalin Dahyabhai <nalin@redhat.com>, Mark Cox]

  *) [In 0.9.6c-engine release:]
     Add support for SureWare crypto accelerator cards from 
     Baltimore Technologies.  (Use engine 'sureware')
     [Baltimore Technologies and Mark Cox]

  *) [In 0.9.6c-engine release:]
     Add support for crypto accelerator cards from Accelerated
     Encryption Processing, www.aep.ie.  (Use engine 'aep')
     [AEP Inc. and Mark Cox]

  *) Add a configuration entry for gcc on UnixWare.
     [Gary Benson <gbenson@redhat.com>]

  *) Change ssl/s2_clnt.c and ssl/s2_srvr.c so that received handshake
     messages are stored in a single piece (fixed-length part and
     variable-length part combined) and fix various bugs found on the way.
     [Bodo Moeller]

  *) Disable caching in BIO_gethostbyname(), directly use gethostbyname()
     instead.  BIO_gethostbyname() does not know what timeouts are
     appropriate, so entries would stay in cache even when they have
     become invalid.
     [Bodo Moeller; problem pointed out by Rich Salz <rsalz@zolera.com>

  *) Change ssl23_get_client_hello (ssl/s23_srvr.c) behaviour when
     faced with a pathologically small ClientHello fragment that does
     not contain client_version: Instead of aborting with an error,
     simply choose the highest available protocol version (i.e.,
     TLS 1.0 unless it is disabled).  In practice, ClientHello
     messages are never sent like this, but this change gives us
     strictly correct behaviour at least for TLS.
     [Bodo Moeller]

  *) Fix SSL handshake functions and SSL_clear() such that SSL_clear()
     never resets s->method to s->ctx->method when called from within
     one of the SSL handshake functions.
     [Bodo Moeller; problem pointed out by Niko Baric]

  *) In ssl3_get_client_hello (ssl/s3_srvr.c), generate a fatal alert
     (sent using the client's version number) if client_version is
     smaller than the protocol version in use.  Also change
     ssl23_get_client_hello (ssl/s23_srvr.c) to select TLS 1.0 if
     the client demanded SSL 3.0 but only TLS 1.0 is enabled; then
     the client will at least see that alert.
     [Bodo Moeller]

  *) Fix ssl3_get_message (ssl/s3_both.c) to handle message fragmentation
     correctly.
     [Bodo Moeller]

  *) Avoid infinite loop in ssl3_get_message (ssl/s3_both.c) if a
     client receives HelloRequest while in a handshake.
     [Bodo Moeller; bug noticed by Andy Schneider <andy.schneider@bjss.co.uk>]

  *) Bugfix in ssl3_accept (ssl/s3_srvr.c): Case SSL3_ST_SW_HELLO_REQ_C
     should end in 'break', not 'goto end' which circuments various
     cleanups done in state SSL_ST_OK.   But session related stuff
     must be disabled for SSL_ST_OK in the case that we just sent a
     HelloRequest.

     Also avoid some overhead by not calling ssl_init_wbio_buffer()
     before just sending a HelloRequest.
     [Bodo Moeller, Eric Rescorla <ekr@rtfm.com>]

  *) Fix ssl/s3_enc.c, ssl/t1_enc.c and ssl/s3_pkt.c so that we don't
     reveal whether illegal block cipher padding was found or a MAC
     verification error occured.  (Neither SSLerr() codes nor alerts
     are directly visible to potential attackers, but the information
     may leak via logfiles.)

     Similar changes are not required for the SSL 2.0 implementation
     because the number of padding bytes is sent in clear for SSL 2.0,
     and the extra bytes are just ignored.  However ssl/s2_pkt.c
     failed to verify that the purported number of padding bytes is in
     the legal range.
     [Bodo Moeller]

  *) Add OpenUNIX-8 support including shared libraries
     (Boyd Lynn Gerber <gerberb@zenez.com>).
     [Lutz Jaenicke]

  *) Improve RSA_padding_check_PKCS1_OAEP() check again to avoid
     'wristwatch attack' using huge encoding parameters (cf.
     James H. Manger's CRYPTO 2001 paper).  Note that the
     RSA_PKCS1_OAEP_PADDING case of RSA_private_decrypt() does not use
     encoding parameters and hence was not vulnerable.
     [Bodo Moeller]

  *) BN_sqr() bug fix.
     [Ulf Möller, reported by Jim Ellis <jim.ellis@cavium.com>]

  *) Rabin-Miller test analyses assume uniformly distributed witnesses,
     so use BN_pseudo_rand_range() instead of using BN_pseudo_rand()
     followed by modular reduction.
     [Bodo Moeller; pointed out by Adam Young <AYoung1@NCSUS.JNJ.COM>]

  *) Add BN_pseudo_rand_range() with obvious functionality: BN_rand_range()
     equivalent based on BN_pseudo_rand() instead of BN_rand().
     [Bodo Moeller]

  *) s3_srvr.c: allow sending of large client certificate lists (> 16 kB).
     This function was broken, as the check for a new client hello message
     to handle SGC did not allow these large messages.
     (Tracked down by "Douglas E. Engert" <deengert@anl.gov>.)
     [Lutz Jaenicke]

  *) Add alert descriptions for TLSv1 to SSL_alert_desc_string[_long]().
     [Lutz Jaenicke]

  *) Fix buggy behaviour of BIO_get_num_renegotiates() and BIO_ctrl()
     for BIO_C_GET_WRITE_BUF_SIZE ("Stephen Hinton" <shinton@netopia.com>).
     [Lutz Jaenicke]

  *) Rework the configuration and shared library support for Tru64 Unix.
     The configuration part makes use of modern compiler features and
     still retains old compiler behavior for those that run older versions
     of the OS.  The shared library support part includes a variant that
     uses the RPATH feature, and is available through the special
     configuration target "alpha-cc-rpath", which will never be selected
     automatically.
     [Tim Mooney <mooney@dogbert.cc.ndsu.NoDak.edu> via Richard Levitte]

  *) In ssl3_get_key_exchange (ssl/s3_clnt.c), call ssl3_get_message()
     with the same message size as in ssl3_get_certificate_request().
     Otherwise, if no ServerKeyExchange message occurs, CertificateRequest
     messages might inadvertently be reject as too long.
     [Petr Lampa <lampa@fee.vutbr.cz>]

  *) Enhanced support for IA-64 Unix platforms (well, Linux and HP-UX).
     [Andy Polyakov]

  *) Modified SSL library such that the verify_callback that has been set
     specificly for an SSL object with SSL_set_verify() is actually being
     used. Before the change, a verify_callback set with this function was
     ignored and the verify_callback() set in the SSL_CTX at the time of
     the call was used. New function X509_STORE_CTX_set_verify_cb() introduced
     to allow the necessary settings.
     [Lutz Jaenicke]

  *) Initialize static variable in crypto/dsa/dsa_lib.c and crypto/dh/dh_lib.c
     explicitly to NULL, as at least on Solaris 8 this seems not always to be
     done automatically (in contradiction to the requirements of the C
     standard). This made problems when used from OpenSSH.
     [Lutz Jaenicke]

  *) In OpenSSL 0.9.6a and 0.9.6b, crypto/dh/dh_key.c ignored
     dh->length and always used

          BN_rand_range(priv_key, dh->p).

     BN_rand_range() is not necessary for Diffie-Hellman, and this
     specific range makes Diffie-Hellman unnecessarily inefficient if
     dh->length (recommended exponent length) is much smaller than the
     length of dh->p.  We could use BN_rand_range() if the order of
     the subgroup was stored in the DH structure, but we only have
     dh->length.

     So switch back to

          BN_rand(priv_key, l, ...)

     where 'l' is dh->length if this is defined, or BN_num_bits(dh->p)-1
     otherwise.
     [Bodo Moeller]

  *) In

          RSA_eay_public_encrypt
          RSA_eay_private_decrypt
          RSA_eay_private_encrypt (signing)
          RSA_eay_public_decrypt (signature verification)

     (default implementations for RSA_public_encrypt,
     RSA_private_decrypt, RSA_private_encrypt, RSA_public_decrypt),
     always reject numbers >= n.
     [Bodo Moeller]

  *) In crypto/rand/md_rand.c, use a new short-time lock CRYPTO_LOCK_RAND2
     to synchronize access to 'locking_thread'.  This is necessary on
     systems where access to 'locking_thread' (an 'unsigned long'
     variable) is not atomic.
     [Bodo Moeller]

  *) In crypto/rand/md_rand.c, set 'locking_thread' to current thread's ID
     *before* setting the 'crypto_lock_rand' flag.  The previous code had
     a race condition if 0 is a valid thread ID.
     [Travis Vitek <vitek@roguewave.com>]

  *) Add support for shared libraries under Irix.
     [Albert Chin-A-Young <china@thewrittenword.com>]

  *) Add configuration option to build on Linux on both big-endian and
     little-endian MIPS.
     [Ralf Baechle <ralf@uni-koblenz.de>]

  *) Add the possibility to create shared libraries on HP-UX.
     [Richard Levitte]

 Changes between 0.9.6a and 0.9.6b  [9 Jul 2001]

  *) Change ssleay_rand_bytes (crypto/rand/md_rand.c)
     to avoid a SSLeay/OpenSSL PRNG weakness pointed out by
     Markku-Juhani O. Saarinen <markku-juhani.saarinen@nokia.com>:
     PRNG state recovery was possible based on the output of
     one PRNG request appropriately sized to gain knowledge on
     'md' followed by enough consecutive 1-byte PRNG requests
     to traverse all of 'state'.

     1. When updating 'md_local' (the current thread's copy of 'md')
        during PRNG output generation, hash all of the previous
        'md_local' value, not just the half used for PRNG output.

     2. Make the number of bytes from 'state' included into the hash
        independent from the number of PRNG bytes requested.

     The first measure alone would be sufficient to avoid
     Markku-Juhani's attack.  (Actually it had never occurred
     to me that the half of 'md_local' used for chaining was the
     half from which PRNG output bytes were taken -- I had always
     assumed that the secret half would be used.)  The second
     measure makes sure that additional data from 'state' is never
     mixed into 'md_local' in small portions; this heuristically
     further strengthens the PRNG.
     [Bodo Moeller]

  *) Fix crypto/bn/asm/mips3.s.
     [Andy Polyakov]

  *) When only the key is given to "enc", the IV is undefined. Print out
     an error message in this case.
     [Lutz Jaenicke]

  *) Handle special case when X509_NAME is empty in X509 printing routines.
     [Steve Henson]

  *) In dsa_do_verify (crypto/dsa/dsa_ossl.c), verify that r and s are
     positive and less than q.
     [Bodo Moeller]

  *) Don't change *pointer in CRYPTO_add_lock() is add_lock_callback is
     used: it isn't thread safe and the add_lock_callback should handle
     that itself.
     [Paul Rose <Paul.Rose@bridge.com>]

  *) Verify that incoming data obeys the block size in
     ssl3_enc (ssl/s3_enc.c) and tls1_enc (ssl/t1_enc.c).
     [Bodo Moeller]

  *) Fix OAEP check.
     [Ulf Möller, Bodo Möller]

  *) The countermeasure against Bleichbacher's attack on PKCS #1 v1.5
     RSA encryption was accidentally removed in s3_srvr.c in OpenSSL 0.9.5
     when fixing the server behaviour for backwards-compatible 'client
     hello' messages.  (Note that the attack is impractical against
     SSL 3.0 and TLS 1.0 anyway because length and version checking
     means that the probability of guessing a valid ciphertext is
     around 2^-40; see section 5 in Bleichenbacher's CRYPTO '98
     paper.)

     Before 0.9.5, the countermeasure (hide the error by generating a
     random 'decryption result') did not work properly because
     ERR_clear_error() was missing, meaning that SSL_get_error() would
     detect the supposedly ignored error.

     Both problems are now fixed.
     [Bodo Moeller]

  *) In crypto/bio/bf_buff.c, increase DEFAULT_BUFFER_SIZE to 4096
     (previously it was 1024).
     [Bodo Moeller]

  *) Fix for compatibility mode trust settings: ignore trust settings
     unless some valid trust or reject settings are present.
     [Steve Henson]

  *) Fix for blowfish EVP: its a variable length cipher.
     [Steve Henson]

  *) Fix various bugs related to DSA S/MIME verification. Handle missing
     parameters in DSA public key structures and return an error in the
     DSA routines if parameters are absent.
     [Steve Henson]

  *) In versions up to 0.9.6, RAND_file_name() resorted to file ".rnd"
     in the current directory if neither $RANDFILE nor $HOME was set.
     RAND_file_name() in 0.9.6a returned NULL in this case.  This has
     caused some confusion to Windows users who haven't defined $HOME.
     Thus RAND_file_name() is changed again: e_os.h can define a
     DEFAULT_HOME, which will be used if $HOME is not set.
     For Windows, we use "C:"; on other platforms, we still require
     environment variables.

  *) Move 'if (!initialized) RAND_poll()' into regions protected by
     CRYPTO_LOCK_RAND.  This is not strictly necessary, but avoids
     having multiple threads call RAND_poll() concurrently.
     [Bodo Moeller]

  *) In crypto/rand/md_rand.c, replace 'add_do_not_lock' flag by a
     combination of a flag and a thread ID variable.
     Otherwise while one thread is in ssleay_rand_bytes (which sets the
     flag), *other* threads can enter ssleay_add_bytes without obeying
     the CRYPTO_LOCK_RAND lock (and may even illegally release the lock
     that they do not hold after the first thread unsets add_do_not_lock).
     [Bodo Moeller]

  *) Change bctest again: '-x' expressions are not available in all
     versions of 'test'.
     [Bodo Moeller]

 Changes between 0.9.6 and 0.9.6a  [5 Apr 2001]

  *) Fix a couple of memory leaks in PKCS7_dataDecode()
     [Steve Henson, reported by Heyun Zheng <hzheng@atdsprint.com>]

  *) Change Configure and Makefiles to provide EXE_EXT, which will contain
     the default extension for executables, if any.  Also, make the perl
     scripts that use symlink() to test if it really exists and use "cp"
     if it doesn't.  All this made OpenSSL compilable and installable in
     CygWin.
     [Richard Levitte]

  *) Fix for asn1_GetSequence() for indefinite length constructed data.
     If SEQUENCE is length is indefinite just set c->slen to the total
     amount of data available.
     [Steve Henson, reported by shige@FreeBSD.org]
     [This change does not apply to 0.9.7.]

  *) Change bctest to avoid here-documents inside command substitution
     (workaround for FreeBSD /bin/sh bug).
     For compatibility with Ultrix, avoid shell functions (introduced
     in the bctest version that searches along $PATH).
     [Bodo Moeller]

  *) Rename 'des_encrypt' to 'des_encrypt1'.  This avoids the clashes
     with des_encrypt() defined on some operating systems, like Solaris
     and UnixWare.
     [Richard Levitte]

  *) Check the result of RSA-CRT (see D. Boneh, R. DeMillo, R. Lipton:
     On the Importance of Eliminating Errors in Cryptographic
     Computations, J. Cryptology 14 (2001) 2, 101-119,
     http://theory.stanford.edu/~dabo/papers/faults.ps.gz).
     [Ulf Moeller]
  
  *) MIPS assembler BIGNUM division bug fix. 
     [Andy Polyakov]

  *) Disabled incorrect Alpha assembler code.
     [Richard Levitte]

  *) Fix PKCS#7 decode routines so they correctly update the length
     after reading an EOC for the EXPLICIT tag.
     [Steve Henson]
     [This change does not apply to 0.9.7.]

  *) Fix bug in PKCS#12 key generation routines. This was triggered
     if a 3DES key was generated with a 0 initial byte. Include
     PKCS12_BROKEN_KEYGEN compilation option to retain the old
     (but broken) behaviour.
     [Steve Henson]

  *) Enhance bctest to search for a working bc along $PATH and print
     it when found.
     [Tim Rice <tim@multitalents.net> via Richard Levitte]

  *) Fix memory leaks in err.c: free err_data string if necessary;
     don't write to the wrong index in ERR_set_error_data.
     [Bodo Moeller]

  *) Implement ssl23_peek (analogous to ssl23_read), which previously
     did not exist.
     [Bodo Moeller]

  *) Replace rdtsc with _emit statements for VC++ version 5.
     [Jeremy Cooper <jeremy@baymoo.org>]

  *) Make it possible to reuse SSLv2 sessions.
     [Richard Levitte]

  *) In copy_email() check for >= 0 as a return value for
     X509_NAME_get_index_by_NID() since 0 is a valid index.
     [Steve Henson reported by Massimiliano Pala <madwolf@opensca.org>]

  *) Avoid coredump with unsupported or invalid public keys by checking if
     X509_get_pubkey() fails in PKCS7_verify(). Fix memory leak when
     PKCS7_verify() fails with non detached data.
     [Steve Henson]

  *) Don't use getenv in library functions when run as setuid/setgid.
     New function OPENSSL_issetugid().
     [Ulf Moeller]

  *) Avoid false positives in memory leak detection code (crypto/mem_dbg.c)
     due to incorrect handling of multi-threading:

     1. Fix timing glitch in the MemCheck_off() portion of CRYPTO_mem_ctrl().

     2. Fix logical glitch in is_MemCheck_on() aka CRYPTO_is_mem_check_on().

     3. Count how many times MemCheck_off() has been called so that
        nested use can be treated correctly.  This also avoids 
        inband-signalling in the previous code (which relied on the
        assumption that thread ID 0 is impossible).
     [Bodo Moeller]

  *) Add "-rand" option also to s_client and s_server.
     [Lutz Jaenicke]

  *) Fix CPU detection on Irix 6.x.
     [Kurt Hockenbury <khockenb@stevens-tech.edu> and
      "Bruce W. Forsberg" <bruce.forsberg@baesystems.com>]

  *) Fix X509_NAME bug which produced incorrect encoding if X509_NAME
     was empty.
     [Steve Henson]
     [This change does not apply to 0.9.7.]

  *) Use the cached encoding of an X509_NAME structure rather than
     copying it. This is apparently the reason for the libsafe "errors"
     but the code is actually correct.
     [Steve Henson]

  *) Add new function BN_rand_range(), and fix DSA_sign_setup() to prevent
     Bleichenbacher's DSA attack.
     Extend BN_[pseudo_]rand: As before, top=1 forces the highest two bits
     to be set and top=0 forces the highest bit to be set; top=-1 is new
     and leaves the highest bit random.
     [Ulf Moeller, Bodo Moeller]

  *) In the NCONF_...-based implementations for CONF_... queries
     (crypto/conf/conf_lib.c), if the input LHASH is NULL, avoid using
     a temporary CONF structure with the data component set to NULL
     (which gives segmentation faults in lh_retrieve).
     Instead, use NULL for the CONF pointer in CONF_get_string and
     CONF_get_number (which may use environment variables) and directly
     return NULL from CONF_get_section.
     [Bodo Moeller]

  *) Fix potential buffer overrun for EBCDIC.
     [Ulf Moeller]

  *) Tolerate nonRepudiation as being valid for S/MIME signing and certSign
     keyUsage if basicConstraints absent for a CA.
     [Steve Henson]

  *) Make SMIME_write_PKCS7() write mail header values with a format that
     is more generally accepted (no spaces before the semicolon), since
     some programs can't parse those values properly otherwise.  Also make
     sure BIO's that break lines after each write do not create invalid
     headers.
     [Richard Levitte]

  *) Make the CRL encoding routines work with empty SEQUENCE OF. The
     macros previously used would not encode an empty SEQUENCE OF
     and break the signature.
     [Steve Henson]
     [This change does not apply to 0.9.7.]

  *) Zero the premaster secret after deriving the master secret in
     DH ciphersuites.
     [Steve Henson]

  *) Add some EVP_add_digest_alias registrations (as found in
     OpenSSL_add_all_digests()) to SSL_library_init()
     aka OpenSSL_add_ssl_algorithms().  This provides improved
     compatibility with peers using X.509 certificates
     with unconventional AlgorithmIdentifier OIDs.
     [Bodo Moeller]

  *) Fix for Irix with NO_ASM.
     ["Bruce W. Forsberg" <bruce.forsberg@baesystems.com>]

  *) ./config script fixes.
     [Ulf Moeller, Richard Levitte]

  *) Fix 'openssl passwd -1'.
     [Bodo Moeller]

  *) Change PKCS12_key_gen_asc() so it can cope with non null
     terminated strings whose length is passed in the passlen
     parameter, for example from PEM callbacks. This was done
     by adding an extra length parameter to asc2uni().
     [Steve Henson, reported by <oddissey@samsung.co.kr>]

  *) Fix C code generated by 'openssl dsaparam -C': If a BN_bin2bn
     call failed, free the DSA structure.
     [Bodo Moeller]

  *) Fix to uni2asc() to cope with zero length Unicode strings.
     These are present in some PKCS#12 files.
     [Steve Henson]

  *) Increase s2->wbuf allocation by one byte in ssl2_new (ssl/s2_lib.c).
     Otherwise do_ssl_write (ssl/s2_pkt.c) will write beyond buffer limits
     when writing a 32767 byte record.
     [Bodo Moeller; problem reported by Eric Day <eday@concentric.net>]

  *) In RSA_eay_public_{en,ed}crypt and RSA_eay_mod_exp (rsa_eay.c),
     obtain lock CRYPTO_LOCK_RSA before setting rsa->_method_mod_{n,p,q}.

     (RSA objects have a reference count access to which is protected
     by CRYPTO_LOCK_RSA [see rsa_lib.c, s3_srvr.c, ssl_cert.c, ssl_rsa.c],
     so they are meant to be shared between threads.)
     [Bodo Moeller, Geoff Thorpe; original patch submitted by
     "Reddie, Steven" <Steven.Reddie@ca.com>]

  *) Fix a deadlock in CRYPTO_mem_leaks().
     [Bodo Moeller]

  *) Use better test patterns in bntest.
     [Ulf Möller]

  *) rand_win.c fix for Borland C.
     [Ulf Möller]
 
  *) BN_rshift bugfix for n == 0.
     [Bodo Moeller]

  *) Add a 'bctest' script that checks for some known 'bc' bugs
     so that 'make test' does not abort just because 'bc' is broken.
     [Bodo Moeller]

  *) Store verify_result within SSL_SESSION also for client side to
     avoid potential security hole. (Re-used sessions on the client side
     always resulted in verify_result==X509_V_OK, not using the original
     result of the server certificate verification.)
     [Lutz Jaenicke]

  *) Fix ssl3_pending: If the record in s->s3->rrec is not of type
     SSL3_RT_APPLICATION_DATA, return 0.
     Similarly, change ssl2_pending to return 0 if SSL_in_init(s) is true.
     [Bodo Moeller]

  *) Fix SSL_peek:
     Both ssl2_peek and ssl3_peek, which were totally broken in earlier
     releases, have been re-implemented by renaming the previous
     implementations of ssl2_read and ssl3_read to ssl2_read_internal
     and ssl3_read_internal, respectively, and adding 'peek' parameters
     to them.  The new ssl[23]_{read,peek} functions are calls to
     ssl[23]_read_internal with the 'peek' flag set appropriately.
     A 'peek' parameter has also been added to ssl3_read_bytes, which
     does the actual work for ssl3_read_internal.
     [Bodo Moeller]

  *) Initialise "ex_data" member of RSA/DSA/DH structures prior to calling
     the method-specific "init()" handler. Also clean up ex_data after
     calling the method-specific "finish()" handler. Previously, this was
     happening the other way round.
     [Geoff Thorpe]

  *) Increase BN_CTX_NUM (the number of BIGNUMs in a BN_CTX) to 16.
     The previous value, 12, was not always sufficient for BN_mod_exp().
     [Bodo Moeller]

  *) Make sure that shared libraries get the internal name engine with
     the full version number and not just 0.  This should mark the
     shared libraries as not backward compatible.  Of course, this should
     be changed again when we can guarantee backward binary compatibility.
     [Richard Levitte]

  *) Fix typo in get_cert_by_subject() in by_dir.c
     [Jean-Marc Desperrier <jean-marc.desperrier@certplus.com>]

  *) Rework the system to generate shared libraries:

     - Make note of the expected extension for the shared libraries and
       if there is a need for symbolic links from for example libcrypto.so.0
       to libcrypto.so.0.9.7.  There is extended info in Configure for
       that.

     - Make as few rebuilds of the shared libraries as possible.

     - Still avoid linking the OpenSSL programs with the shared libraries.

     - When installing, install the shared libraries separately from the
       static ones.
     [Richard Levitte]

  *) Fix SSL_CTX_set_read_ahead macro to actually use its argument.

     Copy SSL_CTX's read_ahead flag to SSL object directly in SSL_new
     and not in SSL_clear because the latter is also used by the
     accept/connect functions; previously, the settings made by
     SSL_set_read_ahead would be lost during the handshake.
     [Bodo Moeller; problems reported by Anders Gertz <gertz@epact.se>]     

  *) Correct util/mkdef.pl to be selective about disabled algorithms.
     Previously, it would create entries for disableed algorithms no
     matter what.
     [Richard Levitte]

  *) Added several new manual pages for SSL_* function.
     [Lutz Jaenicke]

 Changes between 0.9.5a and 0.9.6  [24 Sep 2000]

  *) In ssl23_get_client_hello, generate an error message when faced
     with an initial SSL 3.0/TLS record that is too small to contain the
     first two bytes of the ClientHello message, i.e. client_version.
     (Note that this is a pathologic case that probably has never happened
     in real life.)  The previous approach was to use the version number
     from the record header as a substitute; but our protocol choice
     should not depend on that one because it is not authenticated
     by the Finished messages.
     [Bodo Moeller]

  *) More robust randomness gathering functions for Windows.
     [Jeffrey Altman <jaltman@columbia.edu>]

  *) For compatibility reasons if the flag X509_V_FLAG_ISSUER_CHECK is
     not set then we don't setup the error code for issuer check errors
     to avoid possibly overwriting other errors which the callback does
     handle. If an application does set the flag then we assume it knows
     what it is doing and can handle the new informational codes
     appropriately.
     [Steve Henson]

  *) Fix for a nasty bug in ASN1_TYPE handling. ASN1_TYPE is used for
     a general "ANY" type, as such it should be able to decode anything
     including tagged types. However it didn't check the class so it would
     wrongly interpret tagged types in the same way as their universal
     counterpart and unknown types were just rejected. Changed so that the
     tagged and unknown types are handled in the same way as a SEQUENCE:
     that is the encoding is stored intact. There is also a new type
     "V_ASN1_OTHER" which is used when the class is not universal, in this
     case we have no idea what the actual type is so we just lump them all
     together.
     [Steve Henson]

  *) On VMS, stdout may very well lead to a file that is written to
     in a record-oriented fashion.  That means that every write() will
     write a separate record, which will be read separately by the
     programs trying to read from it.  This can be very confusing.

     The solution is to put a BIO filter in the way that will buffer
     text until a linefeed is reached, and then write everything a
     line at a time, so every record written will be an actual line,
     not chunks of lines and not (usually doesn't happen, but I've
     seen it once) several lines in one record.  BIO_f_linebuffer() is
     the answer.

     Currently, it's a VMS-only method, because that's where it has
     been tested well enough.
     [Richard Levitte]

  *) Remove 'optimized' squaring variant in BN_mod_mul_montgomery,
     it can return incorrect results.
     (Note: The buggy variant was not enabled in OpenSSL 0.9.5a,
     but it was in 0.9.6-beta[12].)
     [Bodo Moeller]

  *) Disable the check for content being present when verifying detached
     signatures in pk7_smime.c. Some versions of Netscape (wrongly)
     include zero length content when signing messages.
     [Steve Henson]

  *) New BIO_shutdown_wr macro, which invokes the BIO_C_SHUTDOWN_WR
     BIO_ctrl (for BIO pairs).
     [Bodo Möller]

  *) Add DSO method for VMS.
     [Richard Levitte]

  *) Bug fix: Montgomery multiplication could produce results with the
     wrong sign.
     [Ulf Möller]

  *) Add RPM specification openssl.spec and modify it to build three
     packages.  The default package contains applications, application
     documentation and run-time libraries.  The devel package contains
     include files, static libraries and function documentation.  The
     doc package contains the contents of the doc directory.  The original
     openssl.spec was provided by Damien Miller <djm@mindrot.org>.
     [Richard Levitte]
     
  *) Add a large number of documentation files for many SSL routines.
     [Lutz Jaenicke <Lutz.Jaenicke@aet.TU-Cottbus.DE>]

  *) Add a configuration entry for Sony News 4.
     [NAKAJI Hiroyuki <nakaji@tutrp.tut.ac.jp>]

  *) Don't set the two most significant bits to one when generating a
     random number < q in the DSA library.
     [Ulf Möller]

  *) New SSL API mode 'SSL_MODE_AUTO_RETRY'.  This disables the default
     behaviour that SSL_read may result in SSL_ERROR_WANT_READ (even if
     the underlying transport is blocking) if a handshake took place.
     (The default behaviour is needed by applications such as s_client
     and s_server that use select() to determine when to use SSL_read;
     but for applications that know in advance when to expect data, it
     just makes things more complicated.)
     [Bodo Moeller]

  *) Add RAND_egd_bytes(), which gives control over the number of bytes read
     from EGD.
     [Ben Laurie]

  *) Add a few more EBCDIC conditionals that make `req' and `x509'
     work better on such systems.
     [Martin Kraemer <Martin.Kraemer@MchP.Siemens.De>]

  *) Add two demo programs for PKCS12_parse() and PKCS12_create().
     Update PKCS12_parse() so it copies the friendlyName and the
     keyid to the certificates aux info.
     [Steve Henson]

  *) Fix bug in PKCS7_verify() which caused an infinite loop
     if there was more than one signature.
     [Sven Uszpelkat <su@celocom.de>]

  *) Major change in util/mkdef.pl to include extra information
     about each symbol, as well as presentig variables as well
     as functions.  This change means that there's n more need
     to rebuild the .num files when some algorithms are excluded.
     [Richard Levitte]

  *) Allow the verify time to be set by an application,
     rather than always using the current time.
     [Steve Henson]
  
  *) Phase 2 verify code reorganisation. The certificate
     verify code now looks up an issuer certificate by a
     number of criteria: subject name, authority key id
     and key usage. It also verifies self signed certificates
     by the same criteria. The main comparison function is
     X509_check_issued() which performs these checks.
 
     Lot of changes were necessary in order to support this
     without completely rewriting the lookup code.
 
     Authority and subject key identifier are now cached.
 
     The LHASH 'certs' is X509_STORE has now been replaced
     by a STACK_OF(X509_OBJECT). This is mainly because an
     LHASH can't store or retrieve multiple objects with
     the same hash value.

     As a result various functions (which were all internal
     use only) have changed to handle the new X509_STORE
     structure. This will break anything that messed round
     with X509_STORE internally.
 
     The functions X509_STORE_add_cert() now checks for an
     exact match, rather than just subject name.
 
     The X509_STORE API doesn't directly support the retrieval
     of multiple certificates matching a given criteria, however
     this can be worked round by performing a lookup first
     (which will fill the cache with candidate certificates)
     and then examining the cache for matches. This is probably
     the best we can do without throwing out X509_LOOKUP
     entirely (maybe later...).
 
     The X509_VERIFY_CTX structure has been enhanced considerably.
 
     All certificate lookup operations now go via a get_issuer()
     callback. Although this currently uses an X509_STORE it
     can be replaced by custom lookups. This is a simple way
     to bypass the X509_STORE hackery necessary to make this
     work and makes it possible to use more efficient techniques
     in future. A very simple version which uses a simple
     STACK for its trusted certificate store is also provided
     using X509_STORE_CTX_trusted_stack().
 
     The verify_cb() and verify() callbacks now have equivalents
     in the X509_STORE_CTX structure.