UGen: An Agentic Framework for Generating Microarchitectural Attack PoCs (arxiv.org)

Researchers have unveiled uGen, an innovative framework that leverages large language models (LLMs) to automate the generation of microarchitectural attack proof-of-concepts (PoCs). This advancement addresses the significant challenges faced in assessing system vulnerabilities, particularly as microarchitectural attacks evolve and exploit novel vectors in modern processors. Traditional approaches to developing attack implementations are not only labor-intensive but also require specialized knowledge and often lack portability. uGen aims to streamline this process by systematically identifying and addressing gaps in the knowledge of existing LLMs, thereby facilitating the creation of functionally correct attack code tailored to defense needs. The significance of uGen lies in its potential to enhance the efficacy of security assessments through automation. In evaluation tests, uGen achieved a remarkable success rate of up to 100% for the Spectre-v1 attack using Claude Sonnet-4 and 80% for the Prime+Probe attack with Qwen3-Coder. Notably, it can generate these PoCs at a minimal cost of just $1.25 in under four minutes. This framework not only improves the accessibility of microarchitectural attack assessments but also represents a pivotal step toward scalable and systematic vulnerability testing in the AI/ML domain, thereby contributing to more robust cybersecurity measures.