A model vulnerability is a weakness in an AI system that can be exploited to change how the model behaves, what it decides, or how reliably it performs. The weakness may come from the model itself, its training data, its inputs, or the way it is deployed. Common attack paths include data poisoning, adversarial inputs, prompt injection, and model extraction.
In cyber security, model vulnerabilities matter because AI is often used to support filtering, detection, triage, and operational decisions. If an attacker can influence the model, they may cause missed alerts, false approvals, or unstable outputs. Defenders reduce this risk with input validation, access control, logging, model testing, continuous monitoring, and human oversight for high-impact actions. In critical environments, fallback procedures are essential when AI output becomes unreliable.