Rapid advancements in connected and autonomous vehicles (CAVs) are fueled by breakthroughs in machine learning, yet they encounter significant risks from adversarial attacks. This study explores the vulnerabilities of machine learning-based intrusion detection systems (IDSs) within in-vehicle networks (IVNs) to adversarial attacks, shifting focus from the common research on manipulating CAV perception models. Considering the relatively simple nature of IVN data, we assess the susceptibility of IVN-based IDSs to manipulation—a crucial examination, as adversarial attacks typically exploit complexity. We propose an adversarial attack method using a substitute IDS trained with data from the onboard diagnostic port. In conducting these attacks under black-box conditions while adhering to realistic IVN traffic constraints, our method seeks to deceive the IDS into misclassifying both normal-to-malicious and malicious-to-normal cases. Evaluations on two IDS models—a baseline IDS and a state-of-the-art model, MTH-IDS—demonstrated substantial vulnerability, decreasing the F1 scores from 95% to 38% and from 97% to 79%, respectively. Notably, inducing false alarms proved particularly effective as an adversarial strategy, undermining user trust in the defense mechanism. Despite the simplicity of IVN-based IDSs, our findings reveal critical vulnerabilities that could threaten vehicle safety and necessitate careful consideration in the development of IVN-based IDSs and in formulating responses to the IDSs’ alarms.