Summary: Anomaly detection has been popular for a long time due to its ability to detect novel attacks. However, its practical deployment has been limited due to false positives. Taint-based techniques, on the other hand, can avoid false positives for many common exploits (e.g., code or script injection), but their applicability to a broader range of attacks (non-control data attacks, path traversals, race condition attacks, and other unknown attacks) is limited by the need for accurate policies on the use of tainted data. In this paper, we develop a new approach that combines the strengths of these approaches. Our combination is very effective, detecting attack types that have been problematic for taint-based techniques, while significantly cutting down the false positives experienced by anomaly detection. The intuitive justification for this result is that a successful attack involves unusual program behaviors that are exercised by an attacker. Anomaly detection identifies unusual behaviors, while fine-grained taint can filter out behaviors that do not seem controlled by attacker-provided data.