A Hyperproperty Lens on Security against Prompt Injection Attacks
Elaine Li, Barnard College, Columbia University, Assistant Professor of Computer Science
Prompt injection attacks exploit an AI agent’s interactions with untrusted external data to cause it to perform dangerous actions, such as exfiltrate user data or complete unauthorized financial transactions. Existing work has shown that deterministic, model-level defenses in the literature achieve one of two extreme ends of the security-utility tradeoff: overdefended, or insecure. In this talk, I will argue that insusceptibility against prompt injection attacks is a hyperproperty, yet is treated by existing defenses as a trace property. I will discuss several hyperproperty-based monitoring approaches to preventing prompt injection attacks in a single-agent setting, and control flow hijacking in a multi-agent setting. I will also discuss present challenges of evaluating security and utility of defenses.
About the speaker
I am an Assistant Professor of Computer Science at Barnard College, Columbia University. My research interests are in logic, automata theory, and theoretical formal methods. Previously, I was a postdoctoral associate at Cornell University, working with Professor Alexandra Silva, and a Faculty Fellow/Assistant Professor at the Courant Institute, New York University. I received my Ph.D. in Computer Science from New York University in August 2025, advised by Professor Thomas Wies. A summary of my dissertation was invited to appear in the EATCS Bulletin. I received my B.S. (Honours) from Yale-NUS College in May 2019.