The emergence of Large Language Models as general-purpose sequence reasoners has opened new opportunities in time series analysis. TIME-LLM demonstrated that a frozen pre-trained LLM, when equipped with a learnable reprogramming layer that aligns patch-level time series representations with the model's token embedding space, can be repurposed for time series forecasting without modifying any of its weights. This thesis extends that insight to a fundamentally different task: semi-supervised anomaly detection on univariate time series. The proposed model follows a reconstruction-based paradigm. It is trained exclusively on anomaly-free segments of each time series, learning to predict the immediately succeeding time step; at inference, anomaly scores are derived from per-step reconstruction errors.