Concept drift undermines the reliability of unsupervised anomaly detection in safety-critical systems, where sensor distributions evolve over time. We benchmark the drift robustness of two widely used detectors (Isolation Forest and One-Class SVM) alongside a streaming change detector (ADWIN), using a modified version of NASA’s C-MAPSS FD001 turbofan dataset. To emulate gradual system degradation, we inject 5% synthetic anomalies across sensors 1–5 and apply a multiplicative drift (1.0→1.5) from cycles 80 to 120. All models are trained on pre-drift data (< cycle 80) using Min-Max normalisation and evaluated across pre-, mid-, and post-drift phases. Isolation Forest achieves an F1 score of 0.857 pre-drift but degrades to 0.140 mid-drift and 0.080 post-drift; retraining at cycle 120 fails to recover performance. One-Class SVM exhibits weak detection throughout (pre-drift F1 = 0.364; post-drift = 0.080), while ADWIN, used naïvely as an anomaly signal on aggregate sensor values, returns F1 = 0 throughout. These findings illustrate three key points: static detectors collapse under moderate drift, simple retraining is ineffective when drift co-occurs with anomalies, and drift signals alone do not reliably flag anomalies. We release our simulation protocol and codebase to support future work in drift-resilient anomaly detection for streaming safety-critical environments.