Autoencoders are widely adopted for feature extraction in security applications, yet we show that their compression objective is structurally incompatible with rare-event detection. In insider threat environments where malicious records constitute less than 5% of observations [31, 4], bottleneck compression suppresses the very behavioral deviations that distinguish insiders from normal users. We formalise this failure mode as the Recall-Compression Paradox: a monotonic degradation in minority-class recall as latent dimensionality decreases. On the Insider Threat Dataset for Classified Environments [1], our best autoencoder configuration (d=24) achieved a recall of only 0.7200, compared to 0.8964 for Extreme Gradient Boosting (XGBoost) under identical cross-validation conditions. This is not a tuning problem. It is an architectural failure.