The Defense Advanced Research Projects Agency (DARPA) announced the winners of its groundbreaking AI Cyber Challenge (AIxCC) at DEF CON 33 in Las Vegas on August 8, 2025, marking a significant milestone in automated vulnerability discovery and patching technology.
Competition Winners and Prize Distribution
Team Atlanta claimed the top prize of $4 million, consisting of experts from the Georgia Institute of Technology, Samsung Research, the Korea Advanced Institute of Science & Technology, and the Pohang University of Science and Technology. Trail of Bits, a New York City-based small business, secured second place with $3 million, while Theori, a team of AI researchers and security professionals from the U.S. and South Korea, earned third place and $1.5 million.
The competition featured seven finalist teams total, including All You Need Is A Fuzzing Brain (Texas A&M University), Shellphish (Arizona State University), 42-b3yond-6ug (Northwestern University), and LACROSSE (Minneapolis).
Impressive Performance Metrics
The challenge yielded remarkable results that demonstrate the viability of AI-powered cybersecurity automation. The finalist teams collectively discovered 54 of 70 synthetic vulnerabilities planted by DARPA, achieving a 77% success rate, and successfully patched 43 vulnerabilities for a 61% patch success rate. Perhaps most impressively, teams completed patches in an average of just 45 minutes.
Beyond the planned synthetic vulnerabilities, the teams uncovered 18 previously unknown real-world software flaws that weren’t planted in advance, which are now being disclosed to software maintainers.
Diverse Technical Approaches
Traditional Tools Enhanced with AI
Trail of Bits, Shellphish, and LACROSSE built systems rooted in traditional security methods like fuzzing and static analysis, then enhanced them with large language models (LLMs). Trail of Bits used LLMs to generate seed inputs for fuzzing tools to improve code coverage, while Shellphish deployed “Grammar Guy” to generate progressive grammars based on feedback loops analyzing uncovered code paths.
AI-First with Traditional Validation
All You Need Is A Fuzzing Brain and Theori employed LLMs as their primary reasoning engines with traditional tools for validation. The Texas A&M team took the most AI-forward approach, using LLMs for vulnerability analysis, system architecture, and strategic decision-making, with about 90% of their codebase written using AI assistance.
Hybrid Approaches
Team Atlanta distinguished itself by using fine-tuned custom models based on Llama 7B, specialized specifically for C programming language analysis. 42-b3yond-6ug developed “super patches” capable of fixing multiple different bugs simultaneously, even when those bugs appear unrelated.
Open Source Commitment and Future Impact
All seven finalist teams are open-sourcing their AI tools for global use. Four tools were released immediately following the announcement, with the remaining three becoming available in subsequent weeks. Trail of Bits has already released their system called “Buttercup,” a fully automated, AI-driven system for discovering and patching vulnerabilities that can run on typical laptops.
Addressing Critical Infrastructure Vulnerabilities
The challenge specifically targeted vulnerabilities in open-source software that underpins critical infrastructure such as hospitals, power plants, and water systems. DARPA Director Stephen Winchell emphasized the urgency, stating that current digital infrastructure represents “ancient digital scaffolding” with accumulated technical debt that creates “a problem that is beyond human scale”.
DARPA and ARPA-H are providing an additional $1.4 million to help teams integrate their programs into software specifically designed to protect critical infrastructure. The agencies view this as a pivotal moment for cyber defense, with Andrew Carney, the program manager, declaring: “There is no excuse not to leverage this flavor of automation. And it will only get better. This is the new floor”.
