AIxCC Behind the Scenes: Team Lacrosse

In this behind-the-scenes interview from DEF CON 33, members of Team Lacrosse discuss their journey competing in the DARPA AI Cyber Challenge (AIxCC). Opening at 00:00, the team describes the engineering complexity of building automated Cyber Reasoning Systems (CRS) capable of autonomously discovering, validating, and patching zero-day vulnerabilities in critical open-source software libraries.

At 04:01, the team members detail their architectural division of labor, explaining how their pipeline integrates fuzzing harnesses, static analysis engines, and large language models into a unified vulnerability triage pipeline. At 08:00, they analyze the role of open-source static analysis frameworks compared to commercial scanners, evaluating tool accuracy and false positive rates under tight competition time constraints.

At 12:00, the team offers candid insights into integrating generative AI into automated cyber defense. They compare modern LLMs to enthusiastic junior interns, observing that while models excel at generating plausible patch candidates, they require rigorous automated verification and regression testing harnesses to catch hallucinated code and avoid introducing new vulnerabilities into production systems. (AI summary — Gemini Flash)