makomk reported a remote vulnerability that I pulled into the master bitcoin/bitcoin tree on December 20. If you are running git-HEAD code on the production network you should pull the latest code to get the bug fixed.
This affects only anybody who has pulled and compiled their own bitcoind/bitcoin-qt from the source tree in the last 5 days.
I made a mistake. I refactored the ConnectInputs() function into two pieces (FetchInputs() and ConnectInputs()), and should have duplicated a check in ConnectInputs for an out-of-range previous-transaction-output in the FetchInputs() method. The result was a new method I wrote to help prevent a possible OP_EVAL-related denial-of-service attack (AreInputsStandard()) could crash with an out-of-bounds memory access if given an invalid transaction.
The bug-fix puts a check in FetchInputs and an assertion in AreInputsStandard. This does not affect the back-ported "mining only" code I wrote that some miners and pools have started using.
The good news is this was found and reported before binaries with the vulnerability were released; the bad news is this was not found before the code was pulled and could have made it into the next release if makomk had not been testing some unrelated code.
Before releasing 0.6, I would like to have an "intelligent, bitcoin-specific fuzzing tool" that automatically finds this type of bug that we can run before every release. If anybody already has one, please speak up!