Unveiling Champions of the AI Cybersecurity Challenge at DEFCON 33
After a rigorous two-year competition, the champions of the AI Cybersecurity Challenge (AIxCC) were announced during the DEFCON 33 hacking conference held on August 9. This prestigious event highlighted the remarkable intersection of artificial intelligence and cybersecurity.
Claiming the top position was Team Atlanta, a formidable consortium of specialists hailing from the Georgia Institute of Technology (Georgia Tech), Samsung Research, the Korea Advanced Institute of Science & Technology, and the Pohang University of Science and Technology. Their impressive efforts earned them a remarkable prize of $4 million.
Trail of Bits, a cybersecurity firm based in New York known for its innovative security research, secured the second position, winning $3 million in this competitive arena of AI-driven cybersecurity.
Completing the podium was Theori, a collaborative group of AI researchers and cybersecurity experts from both the United States and South Korea. They finished third in the Defense Advanced Research Projects Agency’s (DARPA) showcase, receiving a prize of $1.5 million.
Remarkably, the cyber reasoning systems developed by these three teams form part of a suite of four models that have been open-sourced, with immediate availability for public use.
“The remaining three models will be released in the coming weeks,” stated DARPA Director Stephen Winchell during the awards ceremony at DEFCON 33.
The Genesis of AIxCC: A Two-Year Journey
Conceptualized at Black Hat 2023 by Perri Adams, a program manager at DARPA, AIxCC aimed to bring together computer scientists, AI enthusiasts, software developers, and cybersecurity professionals to forge a new era of AI-enhanced cybersecurity tools. These tools are designed to protect critical infrastructure and government operations in the United States.
The initiative received financial backing from DARPA and the Advanced Research Projects Agency for Health (ARPA-H) to assess the potential of AI in identifying and rectifying software vulnerabilities, thereby paving the way for a future where cyber threats could be neutralized as swiftly as they are detected.
Seven finalists — Team Atlanta, Trail of Bits, Theori, All You Need Is A Fuzzing Brain, Shellphish, 42-b3yond-6ug, and Lacrosse — were announced at DEFCON 32 in August 2024, with each team receiving $2 million for their contributions.
Support from technology giants such as Google, Microsoft, Anthropic, and OpenAI, each contributing over $1 million in AI model credits, ensured that the participants possessed the computational prowess necessary to confront significant cybersecurity challenges.
Prior to the winners’ announcement, Jim O’Neill, Deputy Secretary for the U.S. Department of Health and Human Services (HHS), declared an additional $1.4 million injection to supplement the originally planned $29.5 million in prize money.
Refining Cybersecurity Tools with Additional Funding
During a press conference following the announcement, Andrew Carney, the program manager for AIxCC, divulged that the supplementary funding would aid finalists in fine-tuning their tools for practical applications.
The allocation of these additional resources will proceed in stages, conditional upon the winning teams demonstrating tangible adoption of their innovations by key infrastructure entities.
Accelerated Vulnerability Patching at $152 Per Fix
In the culminating phase of AIxCC, held over the past year, participating teams were required to implement their systems in a controlled, simulated environment laden with flaws specifically introduced by the organizers.
The seven finalist teams successfully uncovered 54 out of 70 synthetic vulnerabilities intentionally embedded in the challenge, achieving a detection rate of 77%. This marks a substantial advancement when compared to the previous year’s semifinal, where only 37% of vulnerabilities were identified.
Teams managed to patch 43 of these 54 vulnerabilities and also discerned 18 real-world flaws not engineered by the organizers, successfully addressing 11 of those.
These zero-day discoveries serve to underscore the models’ capacity to identify critical weaknesses in scenarios beyond controlled environments.
“We are currently engaged in the disclosure of [these real-world zero-day vulnerabilities] to the relevant maintainers,” Carney announced on stage.
Both speed and efficiency emerged as defining characteristics, with the AI systems managing to rectify vulnerabilities in an average of 45 minutes — a stark contrast to traditional manual approaches.
Jennifer Roberts, the director of resilient systems at ARPA-H, remarked to reporters that such capabilities are particularly vital within the healthcare sector, where the average time to remedy a vulnerability stretches to 491 days, as opposed to 60 to 90 days in other fields.
Moreover, the economic viability of task completion during the competition was calculated at $152 per fix, presenting a notable cost advantage over conventional human labor expenditures.
“This establishes a new baseline — rapid advancements are imminent. To enhance our safety, we must elevate collective security. This is the pathway forward,” asserted Carney.
Winchell added, “Presently, we inhabit a digital landscape supported by antiquated frameworks. Many of our codebases, programming languages, and business methodologies are burdened by substantial technical debt accrued over time.”
Prize Funds Ignite Future AI Security Innovations
Team Atlanta, having established a track record of victories in various hacking competitions and academic forums, primarily utilized traditional vulnerability discovery methodologies — such as dynamic analysis and fuzzing — merged with OpenAI’s large language models (LLMs), including o4-mini, GPT-4o, and o3, to achieve their success in AIxCC.
Their team excelled in nearly every category, uncovering the highest number of real-world vulnerabilities among the competitors.
Regarding the prize, Taesoo Kim, the chief leader of the team and a professor at Georgia Tech, stated that a substantial portion of the winnings would be allocated to the institute to support future AI-driven vulnerability research initiatives.
The runner-up, Trail of Bits, is comprised of a nimble group of 10 engineers specializing in pioneering software security tools, including their own cyber reasoning platform, Buttercup.
Notably, the company collaborates with the UK’s AI Security Institute to bolster its capabilities.
In the AIxCC challenge, Trail of Bits successfully combined Buttercup and conventional vulnerability discovery techniques with LLMs such as Anthropic’s Claude Sonnet 4, GPT-4.1, and GPT-4.1 mini. Their noteworthy achievements included identifying the largest variety of unique vulnerability categories, known as Common Weakness Enumeration categories (CWEs).
The third-place contestant, Theori, boasts an impressive legacy of triumphs in security competitions, including eight victories at DEFCON’s capture the flag finals.
Source link: Infosecurity-magazine.com.
