One Bit-flip Away from Collapse: Security and Safety in AI Models and Accelerators

The unprecedented progress in deep learning has led to the widespread deployment of advanced Artificial Intelligence (AI) models, such as deep neural networks (DNNs), convolutional neural networks (CNNs), transformer-based large language models (LLMs), vision-language models (VLMs), and emerging state-space models like Mamba, across domains including healthcare, autonomous systems, satellite communication, and social media. While these models have achieved remarkable performance, their ever-increasing scale and reliance on specialized Edge AI accelerators (used in real-time, battery-operated systems for on-device deep learning) have exposed critical safety and security vulnerabilities. In particular, both unintentional hardware faults (e.g., soft errors) and adversarial attacks (e.g., fault injection, bit-flips) can lead to severe degradation or complete failure of AI systems, with potentially devastating consequences in real-world edge applications. This paper provides a comprehensive review of bit-flip fault and attack vulnerabilities across the AI stack, spanning model architectures and hardware platforms. Integrating insights from existing literature with practical assessments, this study intends to be a valuable reference for those dedicated to creating AI systems that are safe, robust, and secure.