Anthropic announced Enterprise Frontier Safeguards on September 1, proposing to put AI monitoring data inside cloud infrastructure the customer controls.
The company said it developed the approach with more than 100 customers and planned a phased rollout later in the fall. The announcement therefore described a coming service, with access dependent on eligibility and rollout, rather than a setting every Claude user could switch on that morning.
Where the records live
Under Anthropic's design, customers can use their own storage, encryption keys and access rules. Automated systems look for misuse across a period of activity; alerts go to the customer's people for review. Anthropic says its employees do not need to perform that human review.
That arrangement gives an IT department a specific system to evaluate. A team reviewing an alert needs enough information to distinguish stolen credentials from legitimate work, while employees and customers have an interest in limiting who can inspect the underlying material. Moving records into the company's account puts those decisions closer to its existing security staff. It also gives those staff another responsibility.
Privacy depends on the product and configuration
Anthropic's API documentation distinguishes its own role as data processor from the arrangements on Amazon Bedrock and Google's cloud platform. A policy that applies to one route into Claude should not be assumed to cover every other route.
Customer control depends on which account receives the logs, who can decrypt them and how alerts are reviewed, including false alarms.
Anthropic said the safeguards would carry no separate Anthropic charge, although customer-owned cloud storage and associated operations could generate cloud-provider fees. Businesses comparing proposals should ask for those costs alongside the model bill. An offer to keep records under the customer's control still needs a named owner, a retention schedule and an explanation staff can understand before they upload confidential work.
Retention and operating costs
An infrastructure team may administer the cloud account while another team investigates alerts. Those responsibilities include deciding who can read records, who can change that access and what a reviewer receives after an alert. The announcement does not establish that every eligible account already has identical controls.
Retention length affects storage volume. For a hypothetical workload producing 2GB of new records each day, 30 days would create about 60GB of records before compression, replication or deletion behavior. Ninety days would create about 180GB on the same basis. Those are volume calculations, not quoted provider prices.
A cost estimate also needs to specify which records, attachments and copies are retained, and whether reads or exports during an investigation incur charges. The retention policy needs the same scope so its deletion rules cover the material actually stored.
A no-training promise does not describe every stored copy
Anthropic's feature-by-feature retention table gives a concrete reason to examine the application, not only the model name. Its Files API retains uploaded files until deletion or their configured expiration and is not ZDR eligible. An application using that upload route has a different storage behavior from one sending eligible material directly with a request.
The documentation also distinguishes client-side memory, whose storage the customer controls, from server-side features with continuing state. An organization can therefore have several retention periods within one assistant: uploaded source documents, model traffic and its own operational records. A promise about one of those stores does not delete the others.
For example, a legal-document assistant might keep the original file in an upload service while its security team retains activity logs in its cloud account. Removing an activity record would leave the original upload unaffected unless the application's deletion process addresses it too. This is an architectural example, not a report about an actual customer's system.
Anthropic describes customer storage, customer-managed encryption keys and automated review as separately optional controls. That makes the configuration material to the comparison. An organization primarily seeking custody of logs may choose differently from one whose main concern is who can decrypt or review them. Its proposed EFS design also leaves model behavior, API rates and rate limits unchanged; changing log custody is not a model-performance upgrade.
