Overview

Software vulnerabilities remain one of the most critical challenges in computing. Our research combines deep learning with program analysis to build systems that can automatically detect vulnerabilities, generate proofs of exploitability, and understand the limitations of current detection approaches.

We take a rigorous, evaluation-driven approach, building benchmarks that test whether models truly understand security semantics rather than memorizing surface patterns.

Key Directions

• Vulnerability Detection: Evaluating and improving how LLMs and deep learning models detect vulnerabilities, including causal approaches that go beyond spurious correlations.
• Proof-of-Vulnerability Generation: Automated systems that not only find bugs but generate concrete exploits to validate them.
• Red Teaming: Studying how AI repair agents can inadvertently introduce new vulnerabilities while fixing existing bugs.
• Fuzzing: Neural-guided fuzzing techniques that combine program analysis with learned models to discover deep bugs.
• Binary Analysis: Learning execution semantics from traces for tasks like binary similarity detection and memory dependence analysis.

Impact

Our work on deep learning-based vulnerability detection revealed critical shortcomings in existing approaches, reshaping how the community evaluates detection tools. Neuzz pioneered neural program smoothing for fuzzing and has been widely adopted. FaultLine demonstrates that LLM agents can automatically generate proof-of-vulnerability exploits.

Contributors

Selected Publications