Ensuring both robustness and fairness in artificial intelligence models is a fundamental challenge, particularly in high-stakes domains such as finance, healthcare, and security. This study evaluates the adversarial robustness and algorithmic fairness of a diverse set of models, including convolutional neural networks (ResNet-50, EfficientNet), transformer-based architectures (BERT, GPT-4, T5, LLaMA 2), and traditional machine learning models (Random Forest, XGBoost, SVM). Experimental assessments involved adversarial attacks such as Projected Gradient Descent (PGD) and Fast Gradient Sign Method (FGSM), as well as fairness evaluation across biased datasets (COMPAS, Adult Census Income, German Credit). Results indicate vast adversarial robustness difference, with vision models experiencing severe accuracy drop (by as much as 45%) compared to NLP transformers (by 35%). Adversarial training corrected the vulnerabilities, recovering accuracy loss by as much as 23%. Regarding fairness, baseline models exhibited stark biases, particularly in recidivism prediction and income classification tasks. It was found that FairBoost and Adversarial Debiasing techniques reduce discriminatory errors, demonstrating the value of using such methodologies. Although there was a trade-off between fairness and accuracy, advanced adaptive learning processes were critical. The findings establish the value of hybrid approaches blending adversarial robustness with fairness-aware optimization for more resilient and equitable AI systems.