In the AIxCC competition, while fully committing to the challenge of leveraging Large Language Models (LLMs), we also integrated traditional techniques to create a more robust bug-finding system with fewer blind spots.
The competition provided a baseline fuzzer (Jazzer for Java projects), but coverage-guided fuzzing in general often struggles with the complex validation logic that guards deep code paths. To address this, concolic execution is a well-known solution for exploring these paths by solving their input conditions. Our main challenge, therefore, was how to effectively leverage this powerful technique for the competition’s bug-finding goals.
