- author = "Petr Kaska",
- title = "Evolving Robust Defenses: A Genetic Programming Approach to {LLM} Jailbreak Protection",
- booktitle = "Student Conference on Innovation, Technology and Science in IT",
- year = "2026",
- editor = "Petr Veigend",
- address = "Brno, Czech Republic",
- month = "5 " # may,
- keywords = "genetic algorithms, genetic programming, Gemma3:12b",
-
URL = "
https://excel.fit.vutbr.cz/submissions/2026/002/2.pdf",
- size = "3 pages",
- abstract = "Large language models remain vulnerable to jailbreak prompts, while many existing defences require access to model internals and are therefore difficult to deploy in black-box settings. We propose a prompt-level defence based on genetic programming that learns a transferable rule for inserting small character-level perturbations into the input prompt before it reaches the target model. The defence is evaluated using a judge-based 0-10 scoring protocol on a broad benchmark covering multiple jailbreak families, open-weight target models, and benign queries; the selected configuration reduces harmful compliance while preserving most legitimate behaviour. The main contribution is a lightweight and deployable defence that improves robustness without retraining or modifying the defended model.",
-
notes = "Studentska Konference Inovaci, Technologii a Vedy v IT
Excel@FIT http://excel.fit.vutbr.cz/
Faculty of Information Technology, Brno University of Technology",
