Catastrophic forgetting is the loss of previously learned knowledge when a model is trained on new tasks or fresh data. In machine learning, this usually happens in sequential or continual learning settings, where each update can overwrite earlier patterns instead of adding to them. The result is that a system becomes better at the newest task while silently degrading on older ones.
In cyber security, this matters for long-lived models such as intrusion detection, malware classification, and analyst-assist tools that must adapt as threats change. If a defender retrains too aggressively, the model may forget older attack signatures, tactics, or normal baseline behavior, creating blind spots. Attackers can also exploit this weakness by feeding a system biased or noisy data over time, pushing it toward poor future performance. Common defenses include replay buffers, regularization, and consolidation methods that preserve important knowledge while allowing updates.