Anthropic previews customer-controlled Enterprise Frontier Safeguards

Anthropic announced Enterprise Frontier Safeguards, or EFS, on September 1. The company’s description says customers can retain activity data in their own cloud accounts while Anthropic’s automated systems analyze a rolling window for serious misuse. Flags go to the customer, and Anthropic human review is not required. Rollout was scheduled to begin in phases later in the fall.

The architecture splits custody from detection

EFS is presented as a response to a real conflict: cross-session misuse detection benefits from retained activity data, while regulated customers may need that data to remain under their own keys, access policies, and audit controls. Anthropic says customer-owned storage, customer-managed encryption keys, and automated review are individually opt-in. It names S3, Azure Blob Storage, and Google Cloud Storage as possible storage destinations.

That division creates an integration contract worth examining carefully. Security teams need to know which events leave the model-serving path, what fields enter the rolling window, how flags are delivered, and which party owns triage, retention, deletion, and incident response.

What to verify before planning around EFS

The announcement covers a phased future rollout rather than general availability. Teams should confirm eligibility, supported product surface, region, cloud account topology, encryption-key behavior, failure modes, and audit-log completeness at procurement time. They should also model cloud storage, reads, writes, and egress because Anthropic says the safeguard itself has no fee while the customer’s cloud provider bills those operations.

A useful proof of concept should test both privacy and detection paths: verify that activity data stays in the designated account, that access is attributable, and that a simulated flag reaches the correct response queue without exposing content to unintended reviewers.