Artificial intelligence (AI) technologies are being deeply integrated into critical infrastructures, making AI resilience essential to ensuring the secure and stable operation of such systems. This study defines AI resilience in terms of four core dimensions—robustness, defensibility, recoverability, and evolvability—and reviews the current state of research in this area. Focusing on these four dimensions, we survey key technical advances both in China and abroad, with particular attention to new challenges and emerging solutions brought about by technologies such as large language models (LLMs). On this basis, we identify several prominent issues hindering the development of AI resilience, including the lack of top-level planning for capability building, absence of evaluation frameworks grounded in realistic application scenarios, and insufficient emphasis on the resilience of LLMs. To address these challenges, we recommend strengthening strategic guidance to establish a systematic resilience framework; developing high-fidelity, multi-dimensional, and reproducible evaluation systems; and exploring the potentials of LLMs to enhance multi-level resilience across the entire lifecycle of training, deployment, operation, and update, thereby enabling the construction of more reliable, trustworthy, and sustainable intelligent systems.