Detecting vulnerabilities in the source code of IoT systems is critical due to the limitations of resources, complex architecture, and security challenges of their devices. Automated discovery is useful for minimizing the risks introduced by the increasing size of the software. We propose a transformer-based identification of vulnerabilities in the IoT systems source code. We employ our method, which leverages different BERT modifications to process code representations and improve detection efficacy. Our contribution is to use BERT-based models on IoT security by introducing a pre-training scheme that is specially designed to improve vulnerability identification. Experiments on benchmark datasets prove that our approach is superior to traditional methods, providing better metrics and thus demonstrating its effectiveness in securing IoT systems. These results also demonstrate that IoT-specialised transformers not only reduce false positives but also generalise well to unseen code, thereby enhancing the automated security assessment of IoT systems.