The Alquist Coder team, composed of students from the Czech Institute of Informatics, Robotics and Cybernetics (CIIRC) and the Faculty of Electrical Engineering (FEE) at the Czech Technical University (CTU) in Prague, secured an impressive second place in the international Amazon Nova AI Challenge. Competing against more than ninety elite university teams from around the world, they stood out as one of the best so-called defending teams in developing trusted AI tools for assisted programming.
The competition, held from January to June 2025, focused on one of today’s key challenges – the safety of generative artificial intelligence. The final round took the form of a realistic head-to-head tournament, where teams building defensive mechanisms clashed with those trying to break them. The goal for defending teams was to develop LLM models that generate high-quality Python code while minimizing the risk of misuse for harmful purposes. The announcement of the results took place in Seattle, USA, at the end of July.
“The competition was very tough, and second place was a huge success for us,” commented team lead Ondřej Kobza from CIIRC CTU. “I think we managed to build a great team, and thanks to our collaboration, we achieved a medal-worthy result. Discussions with other teams and Amazon organizers showed us that we can inspire each other with interesting solution approaches. We definitely see a lot of potential for further work in this area,” added Kobza.
Competition Winners:
Defending Teams:
🥇 University of Illinois Urbana-Champaign
🥈 Czech Technical University in Prague (Alquist Coder)
Attacking Teams:
🥇 Purdue University
🥈 Nova University Lisbon
Alquist Coder consists of researchers and students from the Czech Institute of Informatics, Robotics and Cybernetics (CIIRC) and the Artificial Intelligence Center of the Faculty of Electrical Engineering at CTU (AIC FEE). The development team included students Adam Černý (FEE) and Ivan Dostál (FIT), together with researchers Marie Rigaki and Muris Sladić from AIC FEE. The student team was led by Ondřej Kobza (CIIRC). Expert support was provided by Jan Šedivý (CIIRC) and Sebastian Garcia (AIC FEL).
Each of the ten finalists received $250,000 from Amazon to support development – including a financial grant and access to Amazon Web Services computing resources. The winning teams received an additional $250,000, and second places were awarded $100,000 to be shared among the team members.
Practical Experience
Access to Amazon’s advanced technologies allowed the Czech team to experiment with LLM models, including the use of the latest training methods and research in training data. The knowledge and approaches gained are not limited to the competition but have a direct impact on the development of secure AI systems for industrial and public applications.
“Participating in the competition and training LLM models on Amazon’s infrastructure worth hundreds of thousands of dollars gave us invaluable practical experience. Training LLM models for a specific domain is crucial know-how that we will apply in the development of real-world applications for our industry,” said Jan Šedivý, expert supervisor of the team from CIIRC CTU.
Security and Usability: Striking the Balance
Team evaluations were not based solely on their ability to prevent errors or attacks. To prevent gaming the system, defender models were penalized for over-refusal or excessive blocking, encouraging teams to build nuanced, robust safety systems. Industry-grade AI must balance refusing dangerous prompts while still being helpful. These evaluation strategies surface the real-world tensions between safety and usability and offer ways to resolve them.
The Alquist Coder team used their own defensive LLM model in the competition and also designed an offensive LLM agent that identified its vulnerabilities and generated additional data for fine-tuning the model.
“We were thrilled to participate because we could apply our expertise in cybersecurity. We designed a red-teaming agent based on an LLM that attacked our defensive model and looked for vulnerabilities. It was one of the most fun parts of the competition – we combined knowledge of cybersecurity with LLM technologies and improved the model’s evaluation. It helped us not only to find weaknesses but also to generate more data for further training,” added Muris Sladić, a PhD student from AIC FEE.
More information can be found in the Amazon Science report.
