Georgia Tech researchers presented findings from DARPA's AI Cyber Challenge at USENIX Security 2026, a two-year competition where seven teams built AI systems to identify and patch software vulnerabilities. Their paper, selected as a runner-up for the conference's distinguished paper award, revealed that reliability and human oversight proved as critical as AI capability, with 38-46% of AI-generated patches containing semantic errors that could break other features.
The competition tasked teams with developing Cyber Reasoning Systems to operate with minimal human assistance on 53 software projects in 143 hours. Georgia Tech's winning Team Atlanta combined multiple AI agents with traditional security tools. Researchers found that AI excelled at complex reasoning while traditional tools remained superior for common bugs, underscoring that human experts remain essential for verifying the safety of automated fixes before deployment.
Quwwaa is your AI news butler — a personalized brief and assistant for the stories you care about.
See more on Quwwaa — create a free account