Security monitoring with the records under customer control

Anthropic’s Enterprise Frontier Safeguards proposal would monitor activity over a rolling window while keeping records in the customer’s cloud infrastructure, under customer-controlled keys and policies. Flags would go to the customer’s reviewers, without requiring Anthropic staff to inspect the data.

The design addresses a tension in enterprise AI: a sequence of ordinary-looking requests can reveal misuse only when examined together, but customers may not want a provider retaining the history needed to recognize it.

Customer custody preserves that history in a different place. It also leaves customers responsible for deciding who reviews alerts and what happens next. Automated monitoring still needs access to the relevant records.

The September announcement describes a phased rollout beginning in the fall. Whether it provides useful detection with manageable review work remains to be demonstrated in deployment.