Amazon.com Inc. today announced the winners of its inaugural Nova AI Challenge, a global competition that tasked university teams with hardening and attacking artificial intelligence coding assistants in a live adversarial tournament.

Held in Santa Clara, California, the challenge brought together top student researchers to probe the security of large language model-powered development tools in real-world scenarios. The winners of the defending track, Team PurpCorn-PLAN from the University of Illinois Urbana-Champaign, built a coding assistant using a custom 8 billion-parameter model provided by Amazon.

On the attacking track, Team PurCL from Purdue University took top honors by successfully jailbreaking rival models using multi-turn prompts and novel red-teaming techniques. Runners-up included Team AlquistCoder from Czech Technical University and Team RedTWIZ from Nova University Lisbon.

The Nova AI Challenge was focused on a critical question facing the industry: Can we build AI coding assistants that are both helpful and secure? Rather than relying on static benchmarks, the tournament featured live interactions between attacker and defender bots. Red teams built automated systems to exploit vulnerabilities in AI coding models, while defenders applied reasoning-based guardrails, policy optimization and synthetic training data to block unsafe outputs without degrading utility.

Amazon evaluated the models put forward by the participating teams using both automated tools such as CodeGuru and expert human reviewers, assessing metrics such as functional code generation, diversity of attack, and over-refusal behavior. Teams that struck the right balance between safety and usability stood out, as overzealous blocking was penalized.

Each team received $250,000 in sponsorship and Amazon Web Services Inc. credits, with the winners gaining an additional $250,000 in prize money and the runners-up receiving an additional $100,000.

“Each of the teams brought fresh perspectives to complex problems that will help accelerate the field of secure, trustworthy AI-assisted software development and advance how we secure AI systems at Amazon,” said Eric Docktor, chief information security officer of specialized businesses at Amazon. “What makes this tournament format particularly valuable is seeing how security concepts hold up under real adversarial pressure, which is essential for building secure, trustworthy AI coding systems that developers can rely on.”

Rohit Prasad, senior vice president of Amazon AGI, added that the competition revealed promising methods for combining safety and utility, noting that adversarial evaluation will become essential for building trustworthy foundation models.

Following the finals, participants presented their research at the Amazon Nova AI Summit in Seattle, highlighting how adversarial testing frameworks can be expanded beyond coding assistants to domains such as healthcare and misinformation.

Photo: SiliconANGLE/Reve