A surrogate model is a stand-in system trained to approximate the behavior of a target model. In machine learning, it may be built legitimately to compress a large model or reduce compute costs. In cyber security, the same idea becomes sensitive when an attacker uses repeated queries and collected outputs from a public AI endpoint to train a copy that mimics the original service.
This matters because the target model’s value is not only in its weights, but also in its behavior. If responses can be harvested at scale, a competitor or attacker may reproduce capabilities without breaching the backend. Defenders treat this as model extraction risk and look for signs such as high-volume automated traffic, unusual account patterns, proxy use, and repeated probing of the same prompts. Controls like rate limiting, anomaly detection, and output monitoring help make surrogate-model training harder and more expensive.