DEFCON Conference Releases New Video on AI Cybersecurity Challenge

In this video, the DEFCON conference team interviews a team participating in the AICC (AI Cybersecurity Challenge), a competition organized by DARPA to explore the use of artificial intelligence techniques to solve security problems, including the automatic detection and correction of bugs. The team discusses their approach, the challenges they faced, and the tools they used to build their solution.

The team decided to participate in the challenge after hearing about the competition at the Black Hat 2023 conference. Intrigued by the idea of using AI to solve security problems, they formed a team composed of academic members and former advisors. Their goal was to explore new and innovative techniques to automate bug detection and correction.

Initially, the team considered using traditional fuzzing and static analysis tools, which are well-established in the security field. However, they quickly realized that to succeed in this challenge, they needed to integrate machine learning techniques and large language models. Their initial solution involved using agents based on large models to orchestrate traditional tools and generate results. Over time, they adjusted their approach to overcome the limitations of traditional tools and large language models, particularly in dealing with hallucinations generated by the latter.

The team used a combination of classical techniques and AI to build their system. For example, for bug correction, they primarily used large language models to generate patches, while using traditional tools to verify the validity of these patches. For bug detection, they introduced a model called the "delta model," which analyzes code changes to determine if they introduce new bugs. They also used machine learning techniques to improve the dependency graphs generated by classical tools, eliminating false positives and focusing on potential targets.

Over the two years of the competition, the team observed a rapid evolution of large language models. They admitted to underestimating the capabilities of these models initially, which led them to rely more on classical techniques. However, they recognized that the teams that performed better in the competition were those that fully embraced large language models.

Reflecting on their experience, the team emphasized the importance of integration and engineering to ensure the proper functioning of their system. They also mentioned that minor but critical bugs in their integration affected their performance, particularly in bug correction. They advised against making last-minute code changes and encouraged further exploration of the potential of large language models for system integration.

In conclusion, this video provides a fascinating glimpse into the challenges and innovations in the field of cybersecurity and AI. It shows how teams must constantly adapt and innovate to succeed in high-level competitions like the AICC.