My research focuses on AI Safety and Security, with a long-term vision of securing AI systems and using AI to secure the systems around them. As AI evolves from standalone models into autonomous agents capable of reasoning, planning, and acting in real-world environments, I aim to develop methods that make these systems both secure and security-capable. My research spans two complementary directions:

• Security of AI: Understanding and mitigating risks in AI models and agents, including backdoors, adversarial attacks, hidden policies, sandbagging, reward hacking, and scheming. I am particularly interested in frontier AI preparedness, focusing on dangerous capabilities, emerging misalignment, and reliable monitoring and control.
• AI for Security: Developing autonomous, reliable, and verifiable AI agents for safety and security applications, including cybersecurity, threat detection, risk monitoring, and automated remediation, with an emphasis on long-horizon reasoning, tool use, and verification.

Across both directions, my research emphasizes post-training and control to advance secure and security-capable AI systems.