- Policy Response: Anthropic plans customer-controlled monitoring after enterprises objected to 30-day retention of prompts and outputs from its Covered Models.
- Current Rule: Four Covered Models still default to 30-day retention; Anthropic offers a temporary exception only to eligible Fable customers.
- Customer Control: Enterprise Frontier Safeguards would keep logs, keys, and review in customer clouds while automated systems continue safety scanning.
- Phased Rollout: The safeguards are not generally available; Anthropic targets a phased fall 2026 rollout, with eligibility and activation dates unpublished.
Anthropic is changing how eligible enterprise customers can satisfy safety-monitoring requirements after pushback against mandatory 30-day retention of prompts and outputs from its Covered Models. Its planned Enterprise Frontier Safeguards system would keep monitoring data and human review under customer control, reducing the need to place sensitive logs with Anthropic or another processor, but it is not yet generally available.
The default rule still applies to four named models. Anthropic is offering some eligible Fable customers a temporary zero-data-retention exception while it prepares the new system, which it says will arrive in phases later in fall 2026. Customers outside that exception continue to face the retention and review terms of the Anthropic or cloud-provider channel they use.
What the June Rule Changed
Anthropic introduced the 30-day retention requirement in June for organizations using Claude Fable 5 or Claude Mythos 5. The current Covered Models list also includes the just released Claude Fable 5.1 and Claude Mythos 5.1.
The rule changed the terms for organizations that had previously arranged zero data retention, or ZDR. It applied to ZDR workspaces in the Claude Console, Claude Code use through Claude Enterprise with ZDR, and ZDR access through Amazon Bedrock, Google Cloud Agent Platform, or Microsoft Foundry. Those organizations had to enable retention to use a Covered Model unless Anthropic expressly authorized an exception.
Free, Pro, and Max subscribers were not moved from ZDR by the June rule because Anthropic already retained inputs and outputs on those consumer services. Ordinary non-Covered Claude models remain under customers’ existing agreements and configured retention settings, so an organization eligible for ZDR can keep using earlier models without accepting the Covered Model rule.
Anthropic says the retained window lets its safeguards connect patterns across requests, sessions, and accounts, including repeated jailbreak attempts, stolen credentials, espionage, and data-extortion activity. For data processed by Anthropic, the current policy bars employees from reading retained conversations by default. Automated systems must first flag potential harm, after which a small approved group can use a controlled and logged review path.
The ordinary deletion window is 30 days, but content held for a safety investigation or legal obligation can remain longer. Anthropic also says it does not use retained enterprise prompts and outputs to train new models or for non-safety purposes without express permission.
How Customer-Controlled Monitoring Would Work
Enterprise Frontier Safeguards is designed to preserve automated misuse detection while changing who holds the monitoring data and who can review it. Activity records would stay in the customer’s Amazon S3, Azure Blob Storage, or Google Cloud Storage account under that customer’s encryption keys, access policies, and audit logs.
Anthropic’s automated safeguards would analyze a rolling traffic window for serious misuse and compromised credentials. When the system identifies a concern, it would send a flag to the customer, whose own cleared personnel could review and manage the underlying activity. The design does not require Anthropic employees to inspect customer content.
That is a change in custody and review authority, not an end to storage or monitoring. Anthropic describes the design as providing privacy equivalent to ZDR, even though monitoring data remains in the customer’s environment. The company has not published an independent technical assessment, production telemetry, false-positive rate, or audit showing that the design matches the old system’s safety performance or its claimed privacy outcome.
Anthropic intends to offer the controls through Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google Agent Platform, and Microsoft Foundry. The company says it will not charge for Enterprise Frontier Safeguards itself, but customers that choose their own cloud storage will pay their provider for storage, reads, writes, and data transfers.
Relief Before the Safeguards Arrive
Anthropic says availability will begin in phases later in fall 2026, with broader access also targeted for later that fall. It has not published exact activation dates, a public eligibility test, a production pilot roster, or service commitments.
The immediate relief is narrower. Eligible customers using Fable 5 or Fable 5.1 for internal business applications can receive a limited-time ZDR exception while they wait for Enterprise Frontier Safeguards. Anthropic can modify or withdraw that arrangement if it identifies misuse, and its public material does not state a comparable temporary exception for Mythos 5 or Mythos 5.1.
The timing is clearest on AWS. A September 1 AWS announcement says eligible internal-use customers can use Fable 5 and Fable 5.1 with ZDR through December 31, 2026. That date applies to the AWS bridge, not to every Anthropic or cloud-provider contract.
Cloud Rules Still Differ
The future system aims to give customers comparable controls across major platforms, but current handling is not uniform. For non-exempt Fable 5 and Fable 5.1 use on Bedrock, AWS documents its aws_review mode: AWS keeps prompts and outputs for up to 30 days, may conduct required human review within AWS, and does not share that content with Anthropic.
Google’s Advanced AI abuse-monitoring terms also describe up to 30 days of prompt and response logging, automated investigation, and possible review by authorized Google personnel. For Fable and Mythos use, the documentation says customers must also enable data sharing with Anthropic. Anthropic’s current API documentation groups Microsoft Foundry with Anthropic-processed paths. Microsoft’s exact reviewer responsibility and any live Enterprise Frontier Safeguards implementation remain unspecified.
What Anthropic Has Not Yet Shown
Anthropic says more than 100 customers and major cloud providers helped shape the design. It has disclosed neither a production deployment nor an adoption count.
For enterprises choosing among Covered and ordinary models, the policy is therefore still conditional. An expressly approved Fable exception may restore ZDR during the transition, while other Covered Model use follows the relevant provider’s retention regime. Enterprise Frontier Safeguards would move custody and review into the customer’s environment after it launches, but Anthropic has not yet published the eligibility rules or production evidence needed to show that the planned compromise works as promised.
