Explorando o uso de LLMs para fuzzing de código Lua

Abstract

Fuzzing is a crucial technique for finding vulnerabilities in software, but its effectiveness in scripting languages such as Lua is limited by the difficulty of generating semantically valid test inputs. This work proposes a methodology that uses Large Language Models (LLMs) to generate semantically rich mutations for fuzzing Lua scripts. Our approach involves developing a fuzzer prototype that leverages an LLM’s in-context learning capabilities to create syntactically and semantically plausible code variations. We will validate this methodology by evaluating the code coverage gain of mutations relative to original seeds, the effectiveness in bug identification, and the execution time of the process, aiming to demonstrate the feasibility of using LLMs for the security of systems utilizing Lua.