Compare EFS customer-owned storage with CometAPI response storage controls

Teams evaluating AI privacy controls should not treat customer-owned storage and no conversation storage as interchangeable promises. The boundary is the route and its contract.

Anthropic’s Enterprise Frontier Safeguards (EFS) announcement describes a specific architecture: monitoring activity data can live in a customer’s own cloud account under customer-controlled keys, access policies, and audit logging. Anthropic says rollout is phased. This does not show that an intermediary inherits the arrangement.

CometAPI’s privacy policy says it does not collect, store, or record conversation content, while its Responses API contract documents store as defaulting to true and storing generated responses for later retrieval. That is a scope question, not a conclusion: does the policy cover this path, and what happens to prompts, outputs, metadata, and upstream copies?

For a governed CometAPI Responses integration, record the exact route, account, model, request settings, response-retention behavior, and upstream terms separately. Test this CometAPI route with non-sensitive fixtures and require route-specific clarification before asserting EFS or ZDR equivalence. A provider’s customer-owned storage design cannot be imported into another API path by assumption.

Sources: Anthropic EFS announcement, CometAPI privacy policy, and CometAPI Responses documentation.

The practical control I would add is an asset-level retention map, not just a provider-level promise. For an image workflow, prompts and masks, uploaded reference images, generated previews, final exports, response objects, and request metadata can have different sensitivity and retention needs. A route may avoid saving conversation text while still retaining an image or response object for retrieval, or exposing metadata needed to reconstruct the request.

I would therefore test the exact Responses path with clearly labelled, non-sensitive fixtures and record the lifecycle of each artifact separately: what is sent upstream, what store changes, whether an image or response ID remains retrievable, and what the customer can delete or audit. Until those observations and the applicable terms line up, “customer-owned storage,” “no conversation storage,” and “ZDR” should remain separate labels in the production asset register rather than being collapsed into one control.