Ensuring deep neural network (DNN) safety is critical in cyber-physical systems (CPS), where incorrect predictions can have severe consequences. Since conventional metrics like accuracy cannot be monitored in CPS, detecting Out-of-Distribution (OOD) data is essential. This work presents an application-independent architecture leveraging latent space and uncertainty estimation for efficient OOD detection with low computational overhead. A multi-task approach enhances detection while a Variational Autoencoder (VAE) improves robustness by structuring the latent space more effectively. Experiments on OSR benchmarks and a CPS dataset show competitive performance with state-of-the-art methods.