Code review and testing for the proposed fix is welcome: https://github.com/gavinandresen/bitcoin-git/tree/encryptionbug
Here's how I tested:
I dumped private keys from an unencrypted wallet (using bitcointools).
I wrote a little tool that took a list of private keys and a filename and reported whether or not the 32-bytes of any of the private keys appears anywhere in the file.
I also hacked a copy of bitcoin to dump any newly-generated-private keys into debug.log before they are encrypted and written to disk.
I verified that new keys for an encrypted wallet are never written to any of the Berkeley DB database files by encrypting the wallet, invalidating all of the existing keypool keys, generating new, encrypted keypool keys, sending bitcoins to a new, encrypted key, and checking all of the files for any unencrypted copies of the new private key.
So the problem became how to deal with old, previously-unencrypted keys that were ending up in the wallet.dat file.
I hoped that telling Berkeley DB to 'compact' the file would remove the 'slack' space that contained the old data (the root of the problem is BDB doesn't actually completely delete/overwrite data when you delete a key/value pair; if I missed a setting that makes it do that, please let me know). Doing that reduced the number of old private keys found by my tool, but didn't eliminate the problem.
Pieter worked on a different solution in parallel-- completely rewriting the database to a new file when encryption is turned on, then, assuming the rewrite succeeds (we trust BDB to do what a database is designed to do-- reliably write data to the disk), replace the old wallet.dat with the new wallet.dat.rewrite. During that process, all keypool keys are marked as 'used' (which actually means just not writing a 'pool' entry for them in wallet.dat.rewrite).
That works-- no old private keys made it into the new wallet.dat.
But during testing I ran into two issues:
1) The wallet.dat file isn't the only place the old private keys were found; the database transaction log (database/log.000..N) and some of the __db.00N files also contained them. Adding a call to remove() the database environment at shutdown fixed the __db.00N issue; I had to write code to remove the database/log.000..N file on clean shutdown.
2) If more wallet database operations were performed after the rewrite (I had written code to top-up the keypool with new, secure keys before locking the wallet), old private keys could end up in the new wallet.dat.
To fix both of those issues, we modified the code so that a shutdown happens after wallet re-encryption.
... which reminds me of something else I need to test: if you startup with a wallet encrypted by bitcoin versions 0.4.0 or 0.5.0, the wallet is re-encrypted and all of your old keypool keys are considered used. (I believe it does not shutdown after doing that, though, but it should)
We've never made any guarantees that unencrypted private keys will not end up on your hard disk; if you have a virus on your system that can directly read blocks off your hard disk then it can almost certainly also read system memory, and could steal your wallet passphrase the first time you sent bitcoins.
Filling up free space on your disk (as suggested in this thread) might work, but it is a much better idea to send all of your bitcoins to new, 'born-encrypted' keys. Once fixed binaries are available, that will be easy-- just upgrade and then send all of your bitcoins to yourself, using newly-generated addresses.
Suggestions for more radical solutions, like storing private keys in a separate file, are out of scope. We got what we got; I personally think that a really good "deterministic private keys" solution, where the private keys for your wallet are derived from a passphrase that is only in your head (and maybe written down and stored in your safe deposit box) would be a better use of developer time rather than reworking how we store private keys on disk.
Even better would be further work on multisignature/multidevice solutions so even if your private keys are compromised an attacker can't spend your coins.