Inference risk is the possibility that a machine learning model can derive sensitive information from data it sees, even when that information was never explicitly provided. The model may reveal clues through its outputs, confidence scores, embeddings, logs, or metadata, and it may also memorize patterns from training data that later expose private details.
In cyber security, inference risk matters because it turns model behavior into a privacy and leakage problem. Attackers can use prompt probing, model inversion, membership inference, or side-channel analysis to infer whether a person, record, or secret was present in training or in surrounding system data. Defenders reduce this risk by minimizing inputs, restricting access to raw datasets and logs, using anonymization or pseudonymization where possible, limiting retention, and testing models for leakage before deployment. Good governance treats inference risk as part of the full AI lifecycle, not just a problem at data collection time.