Explain how sensitive information disclosure happens upstream of the final answer and identify the first three controls to add.
The teach Leakage starts at intake, not just at output. If you send the whole artifact to the model, you have already taken the first risk. Output guardrails help, but they arrive late. Better controls happen earlier: data minimization, field-level masking, scoped retrieval, and clear retention rules for prompts, traces, and evaluations. Ask these in order: Do we really need this field for the task? Can we send a placeholder, token, or summary instead? Who else stores this prompt after the model sees it? That sequence shifts the team from “catch leaks” to “remove leak opportunities.”
Sign up free — one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in