The rapid evolution of generative artificial intelligence (GenAI) has given rise to new safety challenges, making it urgent to develop risk assessment frameworks that can support precise governance. Drawing on the stressor‒exposure‒effect analytical paradigm grounded in ecological risk theory, this study conceptualizes GenAI as a stressor within a sociotechnical ecosystem and proposes a safety risk assessment framework comprising three core dimensions: algorithmic fidelity, technological dominance, and administrative governance capacity. It further develops a preliminary system of assessment indicators, providing a basis for translating the theoretical framework into an operational measurement tool. Within this framework, algorithmic fidelity is subdivided into cognitive fidelity (including both backward-looking and forward-looking fidelity) and affective fidelity. Technological dominance considers the penetration of infrastructure and the degree of user acceptance or reliance, while the administrative governance capacity is used to assess the macro-level regulatory capacity of governing authorities based on the policy arrangement approach. Taking clinical decision support in healthcare as an example, this study conducts an exploratory application of the proposed GenAI safety risk assessment framework. The analysis preliminarily illustrates the importance of backward-looking generative fidelity and informal penetration risk in specific contexts, as well as the framework's application logic and potential applicability. The study further discusses directions for methodological innovation and breakthroughs in future GenAI safety risk governance, including the construction of dynamic and computable models of risk evolution, development of intelligent and adaptive assessment and validation technologies, and promotion of scenario-oriented modular and differentiated assessment. Overall, the proposed safety risk assessment framework helps characterize the formation mechanisms and transmission pathways of GenAI risks, providing theoretical support and modular assessment tools for differentiated risk research across diverse application scenarios.