CS Colloquium: Jingxuan He

Speaker:  Jingxuan He

Date:  Wednesday, December 10, 2025

Time:  3:30 PM - 4:30 PM

Location:  Harold Frank Hall 1132

Host:  Wenbo Guo

Title:  Security: A Next Frontier in AI Coding

Abstract:  AI is revolutionizing programming, but is it also creating a new generation of security debt? This talk confronts this critical question through a comprehensive analysis of AI’s impact on the software security landscape. I will begin by introducing two novel benchmarks to quantify the risks: CyberGym, to assess AI’s offensive capabilities in vulnerability reproduction and discovery, and BaxBench, to measure AI’s propensity to introduce security flaws when generating code. Drawing on insights from these evaluations, I will then present two proactive approaches to make AI-generated code secure by design. I will detail our work in fine-tuning models on curated datasets of secure code patterns to minimize the generation of vulnerabilities. Furthermore, I will describe an inference-time constraining mechanism that enforces type safety in generated code, probably eliminating entire classes of bugs.

Speaker Bio: Jingxuan HE is a Postdoctoral Researcher at UC Berkeley, working with Prof. Dawn Song. He completed his PhD at ETH Zurich, where he was advised by Prof. Martin Vechev. His research lies at the intersection of security, AI, and programming languages, with a current focus on evaluating AI’s impact on cybersecurity and mitigating associated risks. His work has been recognized with an ACM CCS Distinguished Paper Award and an ETH Medal for Outstanding Doctoral Thesis, and is adopted for model evaluations by leading AI labs like Anthropic. For more information, please visit https://jxhe.info/.