What best defines Training Data Leakage?

• A. Data leaked/stolen from the training environment due to weak access controls.
• B. Model outputs reveal the training data through memorization.
• C. Data poisoning during training that alters model behavior.
• D. Unauthorized model updates that bypass security controls.

Answer: (A)

Training Data Leakage happens when the data used to train a model ends up exposed to unauthorized people. The core issue is the training dataset itself leaving the secure boundary—typically due to weak or poorly implemented access controls, misconfigurations, or gaps in data governance. When those protections fail, the confidential training data can be leaked or stolen, creating privacy and compliance risks for the organization.

This is why it’s the best fit: it directly describes the exposure of the training data from its environment, rather than issues tied to how the model behaves or is manipulated. For comparison, memorization of training data in model outputs is a privacy risk related to what the model might reveal after training, not a leakage of the data from the training environment. Data poisoning and unauthorized model updates describe integrity or governance problems rather than direct data exposure from the training data itself.