This paper tackles the problem of detecting topical communities from within an organization by mining readily available network access pattern information. A Bayesian generative process is used to model the behavior of user's network access pattern and thereby her consumption of online content. The idea is that users within same topical interest group tend to share similar online access patterns. By leveraging this pattern, along with side information of domain-names and keywords within the accessed websites, one is able to model these observations under the framework of a mixed membership statistical model. Hence the access patterns of users-to-websites, as measured at the edge of an organization's network boundary, can be decomposed into constituent topical communities without any human effort in selecting specific features. Experimental results on real-world network flow trace demonstrate that the proposed method can effectively detect topically meaningful community structures. Besides better detection accuracy of communities compared with other community detection methods, the proposed method can detect interesting but non-evident hidden communities which cannot readily be detected by other known methods.