Vehicular intrusion detection systems (IDS) are crucial to ensure the security of vehicular ad hoc networks (VANETs). However, most current IDS for vehicles have been developed using closed datasets, resulting in a limited detection range. Furthermore, in the real world, updates to IDS often fall behind the emergence of novel and unknown attacks, rendering these systems ineffective in defending against such attacks. To overcome this limitation and protect against network attacks in open scenarios, we propose a novel vehicular intrusion detection method that uses meta-recognition. This method utilizes a new neural network to extract joint features and calibrate the predicted values of a pre-trained model via extreme value theory (EVT). In addition, to adapt to the VANETs environment, we introduce temperature scaling and tail separation sampling methods to enhance the modeling effect and increase the prediction accuracy. Comprehensive experiments indicated that the proposed method can detect known attacks at a fine-grained level, identify unknown attacks, and outperform the current state-of-the-art schemes.