德勤：2025年筑牢AI网络安全防线：新风险环境下的韧性构建研究报告（英文版）

As organisations begin to adopt and scale AI, malicious actors are evolving to target these systems. Some techniques such as prompt injections, jailbreaks and model extraction are relatively new and arise from the way AI models process data and instructions. Others, including supply chain compromise or the exploitation of vulnerable components, build on long standing cyber-attack methods. Nevertheless, the impacts are amplified by AI’s reliance on external models, open source tools and com