The private keys for these certificates were generated on Gavin's main work machine, following the certificate authoritys' recommendations for generating certificate signing requests.
For OSX, the private key was generated by Keychain.app on Gavin's main work machine. The key and certificate is in a separate, passphrase-protected keychain file that is unlocked to sign the Bitcoin-Qt.app bundle.
For Windows, the private key was generated by Firefox running on Gavin's main work machine. The key and certificate were exported into a separate, passphrase-protected PKCS#12 file, and then deleted from Firefox's keystore. The exported file is used to sign the Windows setup.exe.
Gavin is a single point of failure. He could be coerced to divulge the secret signing keys, allowing somebody to distribute a Bitcoin-Qt.app or bitcoin-qt-setup.exe with a valid signature but containing a malicious binary.
Or the machine Gavin uses to sign the binaries could be compromised, either remotely or by breaking in to his office, allowing the attacker to get the private key files and then install a keylogger to get the passphrase that protects them.
"Air gapping" the machine used to do the signing will not work, because the signing process needs to access a timestamp server over the network. And it would not prevent the "rubber hose cryptography" threat (coercing Gavin to sign a bad binary or divulge the private keys).
Windows binaries are reproducibly 'gitian-built', and the setup.exe file created by the NSIS installer system is a 7zip archive, so you could check to make sure that the bitcoin-qt.exe file inside the installer had not been tampered with. However, an attacker could modify the installer's code, so when the setup.exe was run it compromised users' systems. A volunteer to write an auditing tool that checks the setup.exe for tampering, and checks the files in it against the list of gitian signatures, is needed.
The long-term solution is something like the 'gitian downloader' system, which uses signatures from multiple developers to determine whether or not a binary should be trusted. However, that just pushes the problem to "how will non-technical users securely get the gitian downloader code to start?"