The rapid adoption of Edge-AI in smart edge-IoT environments has dramatically led to an augmented vulnerability to cyber risks arising from distributed learning, data heterogeneity, and adversarial manipulation. This paper proposes a new risk-aware adaptive learning model that federated Edge-AI systems explicitly simulates cyber risk in the process of local training and global aggregation. The proposed solution combines stochastic optimization and adversarial risk bounding with adaptive gradient correction to develop strong learning in non-IID data distributions and malicious client behavior. Convergence guarantees are defined by the theoretical analysis in the case of limited adversarial perturbations. The proposed framework achieves up to $95\mathrm{\%}$ detection accuracy and demonstrates more than $20\mathrm{\%}$ improvement in robustness, where robustness is defined as the relative degradation in detection performance under adversarial perturbations. The performance is evaluated against state-of-the-art baselines, including HADA-FL and centralized training on the Edge-IIoTset dataset, with results reported as averages over multiple randomized runs. Furthermore, the model converges within $50$ communication rounds, which corresponds to a fixed training horizon rather than an early-stopping criterion. These findings demonstrate the usefulness of risk-sensitive adaptive learning in safe and trustworthy Edge-AI implementation in a new generation edge-IoT environment.