Program Repair

I am very excited about this research problem! Feels like many pieces required to make it workable, but all the pieces are out there.

Imagine, Friday you identify a bug, write a test to exemplify the bug. Some ML/AI-based approach generates a ballpark fix for a bug, maybe it doesn't compile or actually fix the bug, but at least it contains all the ingredients required to fix. Then some trial-and-error (e.g. Genetic Programming) algorithm is used to search through many variants to find a combination that is a correct fix. All of this deployed as part of CI/CD. Morning morning you have a nice patch to review. And also, the patch is also the most performant out of a number of fixes found.

Anything about software that is measurable can be optimized, possibly even security considerations as mentioned in another comment.

There’s a whole website that collates research related to automated program repair: https://program-repair.org/

People interested in program repair and related topics may also check out this episode of the Haskell Interlude podcast: https://haskell.foundation/podcast/18/

Full disclaimer, Matthias is a colleague of mine :)

Laudable goals. It's dual use for finding and/or inducing vulnerabilities (and not fixing them) until the technology is perfected though, right?

Is this going to be a really wild security ride for a decade or so? Maybe it's already started and I just haven't heard..

I'm curious to know if compiler exposes internal API for external client to listen to its events ? The purpose might be useful for data mining and learning.