Automating LLM-based harness synthesis for unfuzzed projects

The results are pretty interesting and impressive:

"The goal of our efforts is to enable continuous fuzzing for arbitrary open source projects. We ran our approach on a benchmark of C/C++ projects and have captured a subset of the successful integrations in order to integrate them to OSS-Fuzz or upstream ClusterFuzzLite.

During the initial runs of the fuzz harnesses three memory corruption issues were reported and also a couple of memory leakages. We manually debugged the issues to create fixes and report the issues upstream. The memory corruption issues were all heap-based read buffer overflows, where the heap-aspect comes from the harnesses allocating data on the heap. An interesting aspect for one of the discovered issues is that the project already had CodeQL running as part of its continuous integration workflow, however, CodeQL did not find the issue as reported by the automatically generated fuzzing set up."

I do wonder why they chose C/C++ projects as the benchmark.