Prompt Leakage

Compromising AI systems via libraries, models, or datasets.

Embedding signals to prove model ownership.

Detecting unauthorized model outputs or data leaks.

Increasing model capacity via compute.

Logged record of model inputs, outputs, and decisions.

Legal or policy requirement to explain AI decisions.

Recovering training data from gradients.

Inferring sensitive features of training data.