What are the risks with Project Lombok?

I'm coming up with performance goals for the new year, and I thought I'd be fun to put a goal to reduce the size of the code base, especially boilerplate. One action I've come up with to address this is to use Project Lombok to make beans as short as they should be. But I have a habit of overlooking downsides to new software and approaches, so I'm relying on the Stack Overflow community: Can anyone tell me why Lombok is a bad idea?

A limitation of Lombok is the fact that it is closely tied to the java compiler. Since the annotation processor API only allows creation of new files during the compilation (and not the modification of the existing files) lombok uses that API as a entry point to modify the java compiler. Unfortunately these modifications of the compiler make heavy usage of non-public APIs. Using lombok may be a good idea but you must be aware that upgrading your compiler may break your code. The probability is low but I always feel uncomfortable using non-public APIs.

In my opinion source code in "Java+Lombok" is not Java source code anymore. I see it as something similar Borland company made many years ago in their Borland C++ Builder IDE for VCL - they introduced "properties" in C++ code effectively introducing some kind of a new programming language that wasn't C++ anymore (not C++ in sense of standard of C++ language). Sources using "Java+Lombok" are not valid sources in sense of Java language specification. Moreover I think annotations were not designed to influence language semantic.