Título – An In-depth Analysis of Jailbreaking Through Domain Characterization of LLM Training Sets
Autores – Carlos Peláez-González, Andrés Herrera-Poyatos, Francisco Herrera-Triguero
Resumen – Research on large language models (LLMs) is a prominent field in open-world machine learning. Despite their significant capabilities in natural language processing, LLMs face several challenges that must be overcome, namely, consistency, hallucinations and jailbreaking. In this work we focus on the latter. Jailbreak attacks involve crafting prompts designed to bypass the model alignment safeguards of LLMs, leading to harmful outputs that compromise the safety of the LLM model. Our main contribution is a fundamental vision of jailbreaking in terms of the different domains of language that arise when training and aligning LLMs. This theoretical perspective showcases the limitations of current methods and allows us to classify jailbreak attacks in terms of which deficiency of the model they exploit. This contrasts to current classifications that are based on how the prompt is constructed, such as prompt templating. We conclude that a deeper understanding of the behavior of LLMs is essential to prevent jailbreak attacks.
Publicado en – H. Quinti´an et al. (Eds.): HAIS 2024, LNAI 14858, pp. 116-127, 2025