Open Source Summit + Embedded Linux Conference North America 2026

This session explores a dual-phase strategy for hardening the QEMU Virtual Machine Monitor (VMM) through advanced fuzzing and AI-driven automation. We begin by detailing a manual hardening effort that expanded QEMU’s testing surface from 18 to 60 active targets, increasing device line coverage by more than 30%. While effective, manual target creation is a resource-intensive process that struggles to scale across the hundreds of virtualized devices supported by QEMU.

To address these scaling challenges, we introduce an AI-driven agentic pipeline designed to automate the generation and validation of fuzzing targets. This system leverages Large Language Models (LLMs) to analyze device source code and memory regions, generating candidate C++ targets for the QEMU fuzzing engine.

We will discuss the implementation of a self-correcting feedback loop where the agent captures compilation and runtime errors to iteratively refine its output until a stable target is produced. Attendees will see how this approach aims to reach >80% device line coverage by automating the remaining hardware targets that currently lack dedicated fuzzing.