As enterprises rush to integrate Generative AI, security focus remains largely on data privacy and model alignment. However, a critical vulnerability lies deeper: the AI infrastructure stack. This session explores the emerging threat vector of compiler-level exploits in GPU-accelerated environments. Drawing on six years of experience in GPU compute verification and compiler performance at NVIDIA and Intel, I will demonstrate how hardware-software interface vulnerabilities can lead to "silent" model corruption or unauthorized resource access. Attendees will learn practical strategies for implementing "Secure-by-Design" principles at the infrastructure layer, ensuring that the next generation of AI remains resilient against autonomous threats targeting the very silicon they run on.