Recommenders have become a fundamental tool to navigate the huge amount of information available on the web. However, their ubiquitous presence comes with the risk of exposing sensitive user information. This paper explores this problem in the context of user-based collaborative filtering. We consider an active attacker equipped with externally available knowledge about the interests of users. The attacker creates fake identities based on this external knowledge and exploits the recommendations it receives to identify the items appreciated by a user. Our experiment on a real data trace shows that while the attack is effective, the inherent similarity between real users may be enough to protect at least part of their interests.