Artificial intelligence (AI) and Internet-of-Things (IoT) technologies are increasingly integrated into oral healthcare workflows, enabling continuous patient monitoring through smart toothbrushes, intraoral sensors, and AI-driven radiographic analysis. However, the cybersecurity implications of this integration remain critically understudied, despite oral-health AI systems processing sensitive protected health information (PHI) across potentially insecure wireless channels. This paper presents the first systematic adversarial threat model specifically targeting AI-IoT oral healthcare pipelines, and proposes a multi-layered defense framework evaluated under realistic attack scenarios. We constructed a five-layer reference architecture encompassing patient-side IoT sensors, edge gateways, cloud-based AI inference engines, clinician interfaces, and governance controls. Seven threat categories were identified through structured threat modeling (STRIDE-ML). Adversarial robustness was evaluated on a dental radiograph classification task using fast gradient sign method (FGSM), projected gradient descent (PGD), and Carlini-Wagner (C&W) attacks. Defense mechanisms including adversarial training, randomized smoothing, differential privacy stochastic gradient descent (DP-SGD), and Byzantine-robust federated aggregation were benchmarked against these attacks. An undefended baseline convolutional neural network achieved 94.2% clean accuracy but only 18.7% robust accuracy under PGD-7 attack. The proposed full defense stack reduced attack success rate from 81.3% to 21.1% (73.9% relative reduction) while retaining 90.8% clean accuracy and AUC-ROC of 0.951. Byzantine-robust federated aggregation maintained global model accuracy above 87% at a Byzantine client fraction of 30%, compared with 41.8% for standard federated averaging.