Multimodal Security in Artificial Intelligence: A Dual-Track Audit of MNIST Classification Integrity and Qwen-0.5B Adversarial Robustness

As Artificial Intelligence (AI) becomes a core component of critical digital infrastructure, securing multimodal pipelines combining vision and language is a high-stakes challenge. This article investigates the threat landscape of integrated AI systems, specifically investigating data poisoning in MNISTbased handwritten digit classification and adversarial evasion attacks in the Qwen 2.5 0.5B Large Language Model. We propose a Secure AI Infrastructure Framework that bridges technical cybersecurity controls with robust automated audits. Using the Adversarial Robustness Toolbox (ART), we demonstrate that infrastructure-level defenses can reduce visual Attack Success Rates (ASR) by 16 percent. Concurrently, using the Promptfoo security suite, we evaluate the 0.5B model's resistance to "DANstyle" jailbreaks, achieving a 100 percent resistance rate with a total audit latency of 19 seconds on a CPU baseline. Our findings provide a strategic roadmap for deploying resilient, multimodal AI systems in resource-constrained and high-risk environments.