Anthropic announced Enterprise Frontier Safeguards on September 1, 2026, an enterprise-oriented mechanism that combines zero data retention (ZDR) privacy with advanced automated monitoring for detecting model abuse. Rather than storing activity data in infrastructure controlled by Anthropic, the mechanism allows it to be stored within the customer's cloud account, using encryption keys, access policies, and audit logs managed by the organization itself.
The move responds to a practical problem facing companies that want to run more capable models and agents able to perform multiple tasks, but operate in sectors subject to strict regulatory constraints. Anthropic says advanced abuse, including fraud, cyberattacks, and attempts to use stolen credentials, may span multiple sessions and accounts, making it insufficient to analyze each individual interaction and then immediately delete it to detect patterns extending over time.
Data Within the Customer Environment
EFS is designed so that activity data used for monitoring can be stored in the customer's account on Amazon S3, Azure Blob Storage, or Google Cloud Storage. The data remains under customer-managed keys, while Anthropic's automated systems analyze a moving window of usage activity for indicators of serious abuse.
When a pattern requiring attention is detected, the signal is sent to the customer so that its team can conduct a review and take appropriate action. Anthropic emphasizes that human review by its employees is not required under this design; the organization can determine who has permission to view the data and reports, which is important in cases involving protected legal information, nonpublic data, or sensitive reports in regulated sectors.
What Changes in Practice?
EFS offers a middle ground between two options that previously forced a trade-off for organizations: zero data retention, which limits the ability of safety systems to link events over time, and retaining data with the model provider, which may conflict with regulatory requirements or customer policies. Under the new model, the organization retains the data and determines access and review mechanisms, while Anthropic provides the automated detection layer.
The company says the design resulted from collaboration with more than 100 customers from the financial services, healthcare, manufacturing, telecommunications, legal, retail, and public sectors, in addition to Amazon Web Services, Google Cloud, and Microsoft Azure. The consultations included members of the Analysis and Resilience Center for Systemic Risk, including chief information security officers at major U.S. banks, as well as companies such as Comcast, KPMG, Mastercard, Salesforce, and Visa.
Availability and Supported Platforms
Enterprise Frontier Safeguards will support Claude Code, Claude Enterprise, and Claude Platform, as well as Amazon Bedrock, Claude Platform on AWS, Google's Agent Platform, and Microsoft Foundry. Environments accessing the service through AWS, Google Cloud, or Microsoft Azure will receive equivalent controls, while activity data remains in the customer's own cloud account.
The rollout will begin in phases with the goal of broad availability later in fall 2026. Until EFS is ready, eligible customers will receive ZDR in Fable 5 and Fable 5.1. Anthropic explains that customer-owned storage, customer-managed encryption keys, and fully automated review are optional, and do not change model behavior, API pricing, or rate limits. The company does not charge fees for EFS, but the customer bears the storage, read and write, and data egress costs imposed by the cloud provider.
certi.news Analysis
The importance of the announcement lies not in adding a new model, but in moving a key part of the protection architecture into the enterprise's own environment. This could remove an obstacle to using advanced models in banks and regulated sectors, but it does not eliminate the customer's responsibility for configuring access and review policies or for cloud storage costs. The source also does not specify the details of detection indicators, exact retention periods, or methods for measuring false-positive rates—points that organizations will need to assess before adopting the service broadly.