Research

RL-trained agents that discover and exploit vulnerabilities end-to-end—characterizing the frontier of AI cyber capabilities, building the benchmarks to measure them, and translating findings into actionable defense.

Designing realistic environments, experiments, and metrics that elicit and measure what AI systems can actually do in security contexts—from CTF-scale challenges to full-lifecycle attack simulations against defended infrastructure.

Understanding how autonomous agents behave under adversarial pressure—emergent misalignment from reward hacking, capability elicitation methodology, containment architecture, and the policy implications of AI systems reaching expert-level cyber capabilities.

Studying adversarial control of systems where software compromise produces physical effects—including robotics, drones, industrial systems, and critical infrastructure—with a focus on control planes, autonomy, sensing, communications, and safety-critical failure modes.

Investigating attack surfaces across phones, embedded devices, firmware, trusted hardware, and specialized computing platforms—spanning low-level software, hardware-software trust boundaries, device protocols, and exploitation below the application layer.

Discovering and exploiting vulnerabilities in modern software and infrastructure—from novel attack primitives and exploit chains to proof-of-concept development and systematic analysis of the conditions that make real-world compromise possible.