

I think this is definitely a concern, yes.
I think it’s very likely that these models are effectively being tricked into retrieving source code of projects from memory of seeing the actual source code in the biases. This is why some project are much easier to uncover than others.


Fair fair. Ih suppose what I mean is that it’s not really reverse engineering so much as it is recollecting from training samples.