Any SSL connection that does not have RSA blinding turned on can be abused through a "timing attack" which can reveal the private RSA key. This RSA timing attack was proven to be effective against Stunnel (and other SSL software) by David Brumley and Dan Boneh. Their original announcement to the Stunnel list is at the end of this document. OpenSSL versions 0.9.6i and earlier, and 0.9.7a and earlier did not enable RSA blinding by default. The next versions of OpenSSL (0.9.6j and 0.9.7b) are expected to enable blinding by default, with a compile-time option to disable this feature. For those unable to upgrade OpenSSL currently, I provide the following patch to Stunnel-4.04 that forces RSA blinding on all RSA keys -- server, client (if using peer certificates) and emphemeral keys (if temporary keys are used, for example with exportable crypto). This patch does not enable blinding if you are using a version of OpenSSL greater than 0.9.7a. (0.9.7b has blinding on by default. It unfortunately also has a memory leak. Hopefully 0.9.7c will fix this.) It does not, however, differentiate between 0.9.6 variants - it will enable blinding for even non-vulnerable OpenSSL versions of the 0.9.6 series. If you compile your version of OpenSSL 0.9.7b or later with blinding not automatically enabled (which is not the default) it will not be enabled in Stunnel either for consistancy. I do not expect anyone will be compiling OpenSSL 0.9.7b or later without blinding on by default, so this should not be a problem. Enabling RSA blinding has a negligible performance impact (2% by some estimates) that is well worth the security of your private RSA keys. ----------------------------------------------------------------------- To: stunnel-users@mirt.net Date: 13 Mar 2003 16:09:17 -0800 From: David Brumley Subject: Timing attack against stunnel/OpenSSL Dan Boneh and I have been researching timing attacks against software crypto libraries. Timing attacks are usually used to attack weak computing devices such as smartcards. We've successfully developed and mounted timing attacks against software crypto libraries running on general purpose PC's. We found that we can recover an RSA secret from OpenSSL using anywhere from only 300,000 to 1.4 million queries. We demonstrated our attack was pratical by successfully launching an attack against Apache + mod_SSL and stunnel on the local network. Our results show that timing attacks are practical against widely-deploy servers running on the network. While OpenSSL definitely does provide for blinding, mod_SSL doesn't appear to use it. One reason is it appears difficult to enable blinding from the SSL API. This paper was submitted to Usenix security 03. The link to the paper is here: http://crypto.stanford.edu/~dabo/abstracts/ssl-timing.html We notified CERT about a month ago re: this attack, so it's possible you heard about this from them already. flames > /dev/null. Feel free to write with any questions. Cheers, -David Brumley