How Data Leakage Happens in LLMs

Data leakage happens when models reveal sensitive information they’ve seen during training or via logs and prompts. Attackers can probe models to reconstruct secrets, or accidentally exposed logs can leak user data.

Typical leakage paths:

• Training on raw production data that contains PII or secrets.
• Storing prompts and outputs in logs without redaction.
• Third‑party providers retaining data longer than expected.

The AI Security pillar page explains how to combine technical and governance controls to reduce leakage risk.