2025 – Active
Since 2025, I have led research on evaluating and advancing the capabilities of autonomous AI agents in hardware and systems security. While modern agentic AI models demonstrate remarkable potential in software engineering and code generation, discovering and exploiting hardware vulnerabilities—such as microarchitectural side channels, transient execution flaws, and memory controller bugs—requires complex cross-layer reasoning spanning instruction set architectures, microcode, register specifications, and physical hardware interfaces.
This project systematically assesses where state-of-the-art agentic systems succeed and where they stumble when tasked with hardware-focused vulnerability discovery, testing, and exploitation. We design specialized benchmarking environments to evaluate agents on real-world security workflows—including silicon specification auditing, automated hardware fuzzing, side-channel signal analysis, and Proof-of-Concept (PoC) synthesis. By characterizing these frontier capabilities and identifying current reasoning bottlenecks, our goal is both to understand emerging offensive security risks and to develop next-generation autonomous defensive agents capable of identifying silicon and microarchitectural flaws before hardware is fabricated.