Model extraction is an attack in which an adversary probes an AI system to copy its behavior, infer its parameters, or learn details about its training data. Instead of breaking into the server, the attacker sends many carefully chosen queries and studies the outputs until the model can be approximated or reverse engineered. This matters because modern AI models are valuable assets: they may contain proprietary logic, security-sensitive capabilities, or patterns learned from private data.
In cyber security, model extraction can undermine hosted detection tools, fraud filters, and assistant systems by letting attackers clone them or study their decision boundaries. That can help bypass protections or build competing models at lower cost. Defenses include strict authentication, query rate limits, output throttling, anomaly detection, watermarking, and techniques such as differential privacy or model obfuscation. Security teams also monitor for repeated probing, since extraction often looks like ordinary use at first.