Summary: We propose anomalous taint detection, an approach that combines fine-grained taint tracking with learning-based anomaly detection. Anomaly detection is used to identify behavioral deviations that manifest when vulnerabilities are exercised. Fine-grained taint-tracking is used to target the anomaly detector on those aspects of program behavior that can be controlled by an attacker. Our preliminary results indicate that the combination increases detection accuracy over either technique, and promises to offer better resistance to mimicry attacks.