Anthropic has introduced a new security framework called Enterprise Frontier Safeguards, designed to address data retention and privacy concerns for enterprise customers, particularly those in regulated industries. This framework allows organizations to maintain control over their Claude AI activity logs within their own cloud environments, using their own encryption keys and access policies.
The development of Enterprise Frontier Safeguards was influenced by feedback from over 100 customers, including a quarter of the Fortune 100 and all U.S. global systemically important banks, as well as companies like Comcast, KPMG, Mastercard, Salesforce, and Visa. Notably, eight members of the Analysis and Resilience Center for Systemic Risk, which includes CISOs from major financial institutions such as Goldman Sachs, Morgan Stanley, Citi, Bank of America, and Wells Fargo, collaborated with Anthropic on defining the requirements for running advanced frontier models securely within systemically important banks.
A key aspect of the new safeguards is that activity data, used for misuse detection, can reside in the customer's Amazon S3, Azure Blob Storage, or Google Cloud Storage account. This means customers control their encryption keys, access policies, and audit logging. When Anthropic's automated systems flag potential misuse, the alert is sent directly to the customer, and no Anthropic employees review the data.
This approach directly addresses a previous policy change by Anthropic. For its most capable models, starting with Fable 5, Anthropic moved away from "zero data retention" to a 30-day retention window for prompts and responses. The company stated this change was necessary for detecting sophisticated misuse patterns that spread across multiple tasks, sessions, and accounts, which could not be identified by analyzing and immediately discarding individual interactions. While Anthropic maintains it has never trained on enterprise data without explicit permission, the 30-day retention policy posed a significant hurdle for regulated enterprises due to compliance requirements regarding sensitive data storage and customer notifications.
Under Enterprise Frontier Safeguards, automated systems continuously analyze a rolling window of traffic for signals of serious misuse. This includes attempts to develop offensive cyber or biological capabilities and signs of stolen or leaked credentials. Anthropic highlights that credential theft, in particular, is difficult to detect without observing abnormal behavior over time, as individual requests from a stolen key might appear ordinary. The company also noted the risk of AI agents autonomously engaging in destructive behavior.
The customer objection was not to the review process itself, but to who performed the review. Many regulated firms have strict rules about who can access privileged legal material, non-public information, or drug safety reports, and their staff are already cleared and trained for such work. The new framework ensures that human review by Anthropic employees is not part of the loop.
Enterprise Frontier Safeguards will be supported across various Anthropic offerings, including Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google's Agent Platform, and Microsoft Foundry, with consistent controls whether customers purchase directly or through a cloud partner. Customer-owned storage, customer-managed encryption keys, and fully automated review are all opt-in features and do not affect model behavior, API pricing, or rate limits.
The rollout will occur in phases, with broad availability anticipated for later this fall. Eligible customers will receive zero data retention on Fable 5 and Fable 5.1 until the new safeguards are fully implemented for them. Anthropic will not charge for Enterprise Frontier Safeguards; customers will pay their cloud provider for storage, reads, writes, and egress, as they would for any other cloud resource.
