Security has been the focus of grid systems recently. As a kind of tool, grid security infrastructure (GSI) provides the authentication and authorization services and so on. These mechanisms mostly belong to the objective factors, which have not met the needs of security. As the subjective factor, trust model plays an important role in security field. A new two-level reputation trust architecture for grid is given to reduce the costs of system management largely, in which trust relationships amongst virtual organizations (VOs) are built on domain trust managers (DTMs) rather than resource nodes (RNs). Taking inter-domain trust propagation for example, trust model is improved by integrating global reputation and the subjective trust concept of each recommender into synthesizing final trust value. Moreover, before the grid starts to interact with the trustworthy entities, the pre-measure scheme is presented to filter distrustful entities further, which is based on accuracy and honesty. Experimental results indicate that the model can prevent from the malicious attacks better. Security has been the focus of grid systems recently. As a kind of tool, grid security infrastructure (GSI) provides the authentication and authorization services and so on. These mechanisms have belong to the objective factors, which have not met the needs of security. As the subjective factor, trust model plays an important role in security field. A new two-level reputation trust architecture for grid is given to reduce the costs of system management largely, in which trust relationships among virtual organizations (VOs) are built on domain Trust managers (DTMs) rather than resource nodes (RNs). Taking inter-domain trust propagation for example, trust model is improved by integrating global reputation and the subjective trust concept of each recommender into synthesizing final trust value. Moreover, before the grid starts to interact with the trustworthy entities, the pre-measure scheme is presented to filter distrustful entities further, which is based on accuracy and honesty. E xperimental results indicate that the model can prevent from the malicious attacks better.