Safer AI,
for all users.
I am dedicated to uncovering security and privacy challenges arising from AI, understanding their impact on stakeholders, and developing solutions to address them.
- LLM/Agent SecurityLarge language models and autonomous AI agents are rapidly becoming part of everyday applications, where they interpret untrusted content, make decisions, and act on users’ behalf. These capabilities introduce security and privacy risks throughout the system: adversarial prompts may manipulate model behavior, sensitive information may be exposed through model interactions, and compromised tools or dependencies may cause agents to take harmful actions without users’ knowledge. My research examines these emerging threats and develops practical defenses that help LLM-powered systems remain secure, trustworthy, and under meaningful user control.
- Usable SecurityHuman–AI interaction is interesting and highly diverse. The effectiveness of AI often depends on how users interact with it. Some users develop effective interaction strategies that dramatically improve their productivity, while others lack the knowledge or skills to use AI smartly, resulting in wasted time, meaningless token consumption, or even additional security and privacy risks. A good AI system should not rely on users’ expertise. Instead, it should be intuitive, secure, and accessible to everyone, while even providing additional support for vulnerable and underrepresented populations.