Aiming at solving the difficulties of interaction synergism and insecurity of the multi-agent system, a trusted evaluation access control model based on dynamic Bayesian network is proposed in this paper. The model establishes trust domains according to the interaction records of agent itself and the history. The influences of time and interactive behavior on trust evaluation are also considered. Historical interactive records, trust decay control factor and punish factor etc. are introduced. And the sensitivity and accuracy of dynamic interactions are thus improved as well as the safety of the access process is enhanced.