Skip to content

AI Safety & Responsible AI

Model leakage

Unintended disclosure of protected model behavior, parameters, or training information.

Example

Testing reveals memorized fragments of confidential training material in model output.

Why people use it

Knowing how “Model leakage” works helps teams identify harms and choose proportionate safeguards.

What you'll hear

“We need to account for Model leakage before we ship this.”

Related terms