Vehicle-to-Vehicle communication can improve traffic safety and efficiency. This technology, however, increases the attack surface, making new attacks possible. To cope with these threats, researchers have made a great effort to identify and explore the potential of cyberattacks and also proposed various intrusion or misbehaviour detection systems, in particular machine learning-based solutions. Simulations have become essential to design and evaluate such detection systems as there are no real publicly available Vehicular Ad-Hoc Network (VANET) datasets containing a variety of attacks. The drawback is that simulations require a significant amount of computational resources and time for configuration. In this paper, we present an attack simulation and generation framework that allows training the attack generator with either simulated or real VANET attacks. We outline the structure of our proposed framework and describe the setup of a standard-compliant attack simulator that generates valid standardised CAM and DENM messages specified by ETSI in the Cooperative Intelligent Transport Systems (C-ITS) standards. Based on the introduced framework, we demonstrate the feasibility of using deep learning for the generation of VANET attacks, which ultimately allows us to test and verify prototypes without running resource-demanding simulations.