After six intense months of designing, testing, and engaging in adversarial models, Team RedTWIZ – representing the Faculty of Science and Technology of Universidade Nova de Lisboa (NOVA FCT) and featuring CMU Portugal Ph.D. students – Diogo Silva (team leader) and Diogo Tavares, along with Rafael Ferreira, João Pedro Soares, Daniel Pina, Iago Paulo, Artur Horal, and Henrique Paz – secured 2nd place at the 2025 Amazon Nova AI Challenge Trusted AI, a worldwide university AI competition.
Launched in January 2025, the finals of this inaugural edition took place on June 26 and 27 in Santa Clara, California (USA), bringing together ten elite university teams from across the globe to face the challenge of building and breaking secure, trustworthy AI systems.
Team RedTWIZ — the only Portuguese team in the competition — was led by David Semedo and João Magalhães, who is also the CMU Portugal National Co-Director. Both are researchers in the Multimodal Systems group at NOVA LINCS, within the Department of Computer Science at NOVA.
Team RedTWIZ’s innovative techniques, multi-turn attack strategies, and deep understanding of LLM behavior earned them 2nd place overall, along with a $100,000 prize.
“Securing second place in the Amazon Nova AI Challenge is an outstanding result. While we always aim for the top, it’s great to have the team’s effort recognized, and this result is a testament to the effort and innovation that we brought to the competition with our work and research. I am particularly proud of the grad students. Their hard work, dedication, and perseverance, along with all the great ideas that they brought to fruition, from novel attack strategies to evaluation and dynamic planning, are what enabled us to stand out. ”, shares Diogo Silva.
Of the ten participating teams, only two were from outside the United States: NOVA FCT (Portugal) and Czech Technical University (Czech Republic). The U.S. was represented by leading institutions including Carnegie Mellon University, University of Illinois Urbana-Champaign, University of California Davis, Virginia Tech, University of Texas at Dallas, Columbia University, University of Wisconsin, and Purdue University.
This year’s theme, “Trusted AI: Advancing Secure, AI-Assisted Software Development,” challenged participants to develop and attack models that could generate code securely by design. The competition featured five Model Developer teams (defense) and five Red Teams (attack), with NOVA FCT’s team acting as one of the latter. Over four rounds, each model was tested against increasingly sophisticated adversarial strategies.
“This challenge pushed the limits of AI safety and trustworthiness—with cutting-edge techniques, adversarial attacks, and multi-turn evaluations—and our team was right at the core of it. We tackled complex problems, built innovative defences, and grew immensely through the process”, shared João Magalhães.
First place in the competition was shared by two teams: Team PurpCorn-PLAN from the University of Illinois Urbana-Champaign (UIUC) and Team PurCL from Purdue University. The two runner-up teams were Team RedTWIZ from NOVA FCT and Team AlquistCoder from the Czech Technical University in Prague.
More information about the competition here.
News article “CMU Portugal team selected for the Amazon Nova AI Challenge”