If an application encounters a fatal protocol error and then calls
SSL_shutdown() twice (once to send a close_notify, and once to receive one) then
OpenSSL can respond differently to the calling application if a 0 byte record is
received with invalid padding compared to if a 0 byte record is received with an
invalid MAC. If the application then behaves differently based on that in a way
that is detectable to the remote peer, then this amounts to a padding oracle
that could be used to decrypt data.
In order for this to be exploitable "non-stitched" ciphersuites must be in use.
Stitched ciphersuites are optimised implementations of certain commonly used
ciphersuites. Also the application must call SSL_shutdown() twice even if a
protocol error has occurred (applications should not do this but some do
anyway). AEAD ciphersuites are not impacted.
This issue does not impact OpenSSL 1.1.1 or 1.1.0.
OpenSSL 1.0.2 users should upgrade to 1.0.2r.
This issue was discovered by Juraj Somorovsky, Robert Merget and Nimrod Aviram,
with additional investigation by Steven Collison and Andrew Hourselt. It was
reported to OpenSSL on 10th December 2018.
Note: Advisory updated to make it clearer that AEAD ciphersuites are not impacted.
*** This bug has been marked as a duplicate of bug 678564 ***