Probing AI Safety with Source Code