Data leakage occurs when information submitted to a system appears in outputs to other users, when training on customer data allows extraction of that data, or when logging exposes confidential content.
Leakage may occur through model memorization, retrieval misconfiguration, or inadequate tenant isolation.
Alternative Names:
Information Leakage, Model Data Leakage
Why it Matters?
This is the confidentiality risk that matters most for legal AI, because a leak involving privileged material is a serious ethical event rather than a technical one. The controls that address it are no-training commitments preventing memorization, tenant isolation extending to retrieval indexes, and zero retention eliminating stored copies. Verifying that isolation covers embeddings and search indexes, not just document storage, is the specific diligence step.
Frequently Confused with
Related terms
Frequently asked questions
How does data leakage occur in AI systems?
What controls address it?
