Whilst the first training sets would have used Stack overflow and open source projects.

I have my suspicions that the AI companies have been using reverse engineering to gain more training data.

Ethics don't really concern these companies so...

When you have essentially infinite compute and are constantly looking for more training data. To me the next logical step is to start reversing compiled programs.

It could also explain why these AI models are so good at it too...

How do you copyright stolen material?