A trained AI model can retain information about the data it was trained on and, in some cases, reveal it through carefully crafted prompts or attacks. This matters for Redgate, where Test Data Management exists to keep production data out of places it shouldn't be. This talk looks at privacy from two angles. The first is the model itself: how do we train or fine-tune models without them revealing sensitive information they've learned? The second is the system around the model. In modern LLM and agentic applications, data can flow through prompts, vector stores, tools, and logs, creating new opportunities for sensitive information to leak. For each of these, I'll cover the practical defences that make a difference. On the model side, that includes sanitising training data, using synthetic data where appropriate, and techniques that reduce the risk of memorisation. On the systems side, it means validating and redacting inputs, limiting what models can access, protecting retrieval systems, and treating logs as sensitive data.