“Framing reward hacking detection as an anomaly detection task did not achieve an AUROC greater than 60% across all tested RL environments.”