Federated Learning is a machine learning methodology that emphasizes data privacy, involving minimal interaction with each other’s systems, primarily exchanging model parameters. However, this approach can introduce challenges in system development and operation because it inherently faces statistical and system heterogeneity issues. The diverse data storage formats and system environments across clients limit the feasibility of training with a uniform code. To distribute a new code to each environment, active participation of Federated Learning collaborators is necessary, incurring time and cost. Moreover, it impedes adopting modern automated development and deployment paradigms such as DevOps or MLOps. This study investigates how Large Language Models (LLMs) can automatically tailor a single code to individual client environments in heterogeneous scenarios without human intervention. Moreover, to enable the automatic adaptation of the deployed code for conducting new experiments within the system, it is imperative to assess the presence of potentially malicious code that could jeopardize data security. To address this challenge, we introduce a novel prompt engineering technique to enhance LLMs’ detection capabilities, thereby bolstering our ability to detect malicious code effectively.