The widespread application of artificial intelligence (AI) technology faces critical challenges from inherent security threats, hindering the development of trusted AI ecosystems. This paper presents a systematic review of AI’s intrinsic security, proposing a three-dimensional analytical framework covering data, model, and system layers to analyse core risks such as data contamination, adversarial sample attacks, and API abuse, alongside their cross-layer coupling effects. We then design a comprehensive dynamic defense system that integrates prevention, detection, and response and empirically validates the system across medical, financial, and autonomous driving domains, demonstrating significant risk reduction. Furthermore, the study outlines a phased technical roadmap for integrating quantum computing and edge AI to address the ‘impossible triangle’ of privacy, efficiency, and real-time performance. Finally, it advocates for interdisciplinary collaborative governance—combining legal, ethical, and technological approaches—to build a resilient and trustworthy AI ecosystem, identifying future research priorities for adaptive defense and full lifecycle verification.