Disclosed are systems and methods for limiting access of a user profile to dangerous content in a social network service. The described system produces a social graph for a given user profile in the social network service, and identifies clusters of objects (e.g., other user profiles, contents) within the social graph. The described system analyzes whether certain objects in the social graph should be characterized as suspicious based on their clustering and on a database of known forbidden objects. The described system may further learn and add unknown objects to the database of forbidden objects.