Abstract

Traditional fuzzing techniques rely heavily on heuristic methods and code coverage, which can often lead to limitations in bug discovery. This presentation discusses the potential overlap between fuzzing and Reinforcement Learning (RL), especially in the context of recent transformer models. Using a framework that expands feedback mechanisms beyond code coverage, the talk aims to explore the possible pathways to build a more efficient bug-hunting methodology. Inspired by real-world case studies like Zenbleed, the discussion covers how an RL-guided system could better adapt to multiple feedback signals. The talk also opens a dialogue on modeling the broader landscape of bug hunting activities within an RL context, encouraging an exploration toward more adaptable and efficient security research methods.