At Amazon, responsible AI development includes partnering with leading universities to foster breakthrough research. Recognizing that many academic institutions lack the resources for large-scale studies, we’re transforming the landscape with the Amazon Nova AI Challenge. While the Amazon Nova AI Challenge will explore various facets of generative AI (Gen AI), this year’s challenge is centered on “Trusted AI: advancing secure, AI-assisted software development to build safer, more reliable applications.”
“This challenge is truly unique,” says Dr. Ruoxi Jia, an assistant professor and faculty advisor for the Virginia Tech team. “Cutting-edge AI research typically demands access to massively trained models or open-weight models. But, unfortunately, open-weights model often don’t have the right level of performance. Amazon is dramatically lowering the research barrier by giving academia unprecedented access to resources and real-world scenarios.”
AI is revolutionizing software development by automating tedious yet essential tasks, such as updating software, which frees up teams to focus on innovation. For example, by integrating Amazon Q’s code transformation capability into Amazon’s internal systems, the team was able to reduce the time it takes to upgrade a Java application to Java 17 from what is typically 50 developer-days to just a few hours. This saved an estimated equivalent of 4,500 developer-years of work and generated an estimated $260M in annualized efficiency gains. However, as AI becomes more integrated into coding processes, it will inevitably bring new security challenges. By proactively addressing these risks, trust and safety are prioritized from the start.
In a “tournament-style” format, ten university teams—five model developer (defense) teams and five red (attack) teams—are competing in four sequential tournaments to strengthen AI-based secure software development. Each defense team’s code-generating model will face all five red teams, which will probe for vulnerabilities and flaws using automated techniques. The Amazon Nova team built a custom model on AWS Trainium hardware for the challenge for enabling open, collaborative research in secure software development.
The first tournament kicked off in January 2025, with a final round being held live in June 2025. All teams will publish research papers detailing their methodologies and findings, ultimately enhancing user experience, preventing misuse, and enabling more secure application of AI to software development. The advances from the challenge will contribute to the broader field of responsible AI development in code generation and beyond.
“Research is often very solitary,” says Atharva Naik, the student lead of the Carnegie Mellon University team. “Here we’re actually competing with other research teams and trying to keep up with their advancements and out-do them in real time.” The tournament format also drives teams to implement and test their strategies quickly without getting bogged down in one solution, stresses Naik.
The selected teams in each category:
Model Developer teams
- Carnegie Mellon University
- Columbia University
- Czech Technical University, Prague, Czech Republic
- University of Illinois Urbana-Champaign (UIUC)
- Virginia Tech
Red teams
- NOVA School of Science and Technology, Lisbon, Portugal
- Purdue University
- University of California, Davis
- University of Texas at Dallas
- University of Wisconsin Madison
Organizers and participants agree that the tournament format has proved to be highly motivating. The attack and defense systems must be effective against five opposing teams, and each team will get better as they face off against different opponents in each tournament.
“We brought together the best and brightest from academia to not only compete, but collectively tackle one of the most important problems in the real-world application of generative AI – safe and secure software development,” says Rohit Prasad, SVP of Amazon Artificial General Intelligence. “We’ve designed this challenge as a unique, fast-paced tournament for accelerating academic research for practical use. In Amazon tradition, I’m looking forward to the competing teams working hard, having fun, and making some history along the way to the finals.”
“This challenge exemplifies our commitment to advancing responsible AI development and security,” said Steve Schmidt, Chief Security Officer, Amazon. “By partnering with universities, we’re tapping into a wellspring of fresh ideas and cultivating future AI security leaders. This initiative goes beyond theoretical research—it’s about developing new ways to identify security vulnerabilities and protect against threats that can be directly applied to generative AI coding assistants. I can’t wait to see what the students invent, and to share their research.”
Each team will receive $250,000 in sponsorship, monthly AWS credits, and the chance to compete for top prizes. The winning red team and model developer team will each receive $250,000 (to be split among students), with second-place teams earning $100,000. Including stipends, $700,000 in prizes, and AWS credits, the total investment in teams exceeds $5 million.
Bringing together the best
Amazon organizers reviewed over 90 proposals to choose the final ten teams to compete in the challenge. According to Michael Johnston, an applied science leader at Amazon overseeing the science and engineering behind the challenge, it was a tough decision, and those selected had to bring a wide range of unique and practical ideas to the table. Because each team will compete against multiple opponents, they need to be ready with multiple strategies. And because their opponents will be constantly adjusting, those ideas have to show creativity and adaptability.
The challenge is inherently multi-disciplinary – situated at the intersection of responsible AI, Gen AI, security, conversational AI, and automated software development. As such it has brought together teams with expertise across multiple fields of study bringing different talents and perspectives to the competition.
From theory to practical solutions
The Amazon Nova AI Challenge encourages teams to approach problems through a more pragmatic lens than what’s often used in academic research. Virginia Tech’s Jia points out that academic literature tends to focus on theoretical problems, with a bias toward complicated solutions. That’s not what is needed here. The challenge helps frame the problem in a way that is beneficial to real people, says Jia. In her conversations with Amazon researchers, she says it’s clear that they’re not impressed by solutions that are overly complex. “They tell me they want the solution that is the simplest: one that is robust and easy to troubleshoot,” she says. “That can be eye-opening and reshapes my research philosophy a little bit.”
The teams are also accessing a level of resources not often available in academia. “We are trying new things that otherwise would be very hard on an academic budget,” says Naik. His team has generally worked with smaller datasets, and hasn’t had an opportunity to train LLMs or run larger experiments. Kobza was similarly impressed during bootcamp. “We got instant access to AWS Trainium chips, a family of AI chips purpose built by AWS for AI training and inference to deliver high performance while reducing costs. That’s really amazing, especially for a team from a small university in the Czech Republic where we don’t have the opportunity to access this powerful hardware.”
“The challenge aligns with our research interests really well,” says Professor Xiangyu Zhang of Purdue, an advisor to his school’s team. “And we do need the extra step of competition. We are attacking from all corners.”
Stay tuned for updates on the teams’ progress and coverage of the finals in June 2025.
