Abstract: With the in-depth advancement of informatization, artificial intelligence(AI) technology has demonstrated its potential to drive the energy revolution and the intelligentization of power grids, especially in core scenarios such as load forecasting, fault diagnosis, and energy dispatching, becoming a key engine for the construction of new power systems. However, the security threats it triggers exhibit an intertwined nature of technical and non-technical characteristics, necessitating the establishment of a multi-dimensional risk analysis framework. Technical threats mainly stem from the inherent vulnerabilities of algorithms, including data poisoning attacks on long short-term memory (LSTM) load forecasting models and adversarial sample interference in graph neural network (GNN) grid topology modeling. For instance, attackers can use generative AI to forge false data with a similarity of 93% to real load curves, causing the prediction error rate to increase by 2.8 times. Non-technical threats are manifested as external attacks such as social engineering and supply chain infiltration, for example imitating dispatchers' voice instructions through deepfake technology or implanting backdoor models by exploiting vulnerabilities in open-source AI frameworks. This article proposes a full life-cycle protection architecture for power systems in response to the above two types of threats.