Code generators are autoregressive (code->code) not labeled (text->image) so the attack wouldn't work. Also, you can actually tell if a code model works or not

Code generators are autoregressive (code->code) not labeled (text->image) so the attack wouldn't work. Also, you can actually tell if a code model works or not while you're training it, by running tests on the output. You could put up a lot of misleading code where the comments are wrong or there's bugs in it… seems bad for obvious reasons though.