ZDR
How it works
A request is processed in memory: the model generates a response that is returned to the client, after which input and output content is not persisted. Customer content is excluded from abuse-monitoring logs that are otherwise retained for a limited period. Only derived data/metadata needed for safety (e.g. safety classifier results) may be retained. The mode is enabled after provider approval, scoped per API organization or project, and is distinct from the default retention policy. Providers also offer an intermediate variant (OpenAI: Modified Abuse Monitoring; Azure: modified abuse monitoring) that drops storage and human review but keeps real-time automated review.
Problem solved
Default provider logging of prompts and responses (e.g. up to 30 days for abuse monitoring) and the resulting data-at-rest risk block AI adoption in environments with strict confidentiality and compliance requirements (HIPAA, GDPR, professional secrecy). ZDR removes persistent content storage, enabling processing of sensitive data.
Components
Inputs and outputs are not persisted after the request is processed.
Customer content is excluded from abuse-monitoring logs that are otherwise retained for a limited period (OpenAI: up to 30 days).
Derived data may be retained for safety, e.g. User Safety classifier results (Anthropic).
ZDR is granted after provider approval, scoped to an API organization or project.
Hyperparameters (configurable axes)
How long input/output content is retained.
Level of content logging and review for safety.
Unit to which ZDR applies.