აბსტრაქტი. The proliferation of 5G networks introduces a new threat landscape characterized by ultra-low latency services, massive device density, and software-driven network functions. These properties make static signature-based intrusion detection systems insufficient for detecting evolving and previously unseen attacks. This paper presents an adaptive hybrid intrusion detection system (IDS) for 5G networks combining an incremental Adaptive Random Forest (ARF) classifier, a bottleneck autoencoder, and a Conditional Tabular GAN (CTGAN) augmentation module. Experiments use the public 5G-NIDD dataset — 1,215,016 flows after cleaning, nine classes — treated throughout as the 5G Network Intrusion Detection Dataset (not the 3GPP Non-IP Data Delivery service). Two strictly separate evaluation tasks are defined: (i) multiclass known-attack classification and (ii) binary anomaly detection for zero-day simulation. The Dn(x) discriminator score used in the fusion function is extracted offline from the pre-trained CTGAN discriminator and embedded as a static scoring function at inference; the GAN generator is not used at runtime. Under the multiclass task the hybrid model achieves weighted F1 = 0.98 (macro F1 = 0.97 ± 0.01 over five random seeds). Under the binary zero-day task, using a full leave-one-attack-type-out protocol across all eight attack classes, macro F1 = 0.86 ± 0.02. Per-class results, a condensed confusion matrix, fusion-weight ablation, and latency percentiles (mean / p95 / p99) are reported. The architecture maps onto 3GPP NWDAF logical functions and NEF northbound APIs.