Anthropic reverses enterprise data retention policy after client backlash

Anthropic said Tuesday it is unwinding the most contentious piece of a data-handling policy it imposed on business customers three months ago, after enterprise clients in banking, health care and law pushed back over having their Claude usage logs stored on Anthropic's own servers. The company's new program, called Enterprise Frontier Safeguards, will let customers keep those logs inside cloud accounts they control, under encryption keys they hold, rather than handing them to Anthropic.

The shift does not eliminate data retention outright. It relocates it. Under the June policy, Anthropic held 30 days of prompts and outputs from its most capable models on its own infrastructure so it could scan for signs of large-scale cyberattacks or bioweapons research. Under the new arrangement, that same 30-day window of activity will instead sit in a customer's own Amazon S3, Google Cloud Storage or Azure Blob Storage account, and automated systems — not Anthropic staff — will do the scanning, flagging anything suspicious back to the customer rather than to Anthropic's internal review teams.

A policy that became a liability

Anthropic introduced the original retention rule in June alongside the launch of Claude Fable 5 and Claude Mythos 5, two "frontier" models the company judged capable enough to warrant closer monitoring for misuse. The company argued that spotting sophisticated attacks — chains of prompts spread across sessions and accounts — required holding onto data long enough to analyze patterns, something a strict zero-retention setup could not support. Anthropic said at the time it would not use the retained data to train models or for any purpose unrelated to safety.

Enterprise customers were not reassured. Companies in regulated industries said the requirement conflicted with contractual and regulatory obligations that bar sensitive data from leaving their own environments, even temporarily and even with usage restrictions attached. According to reporting on internal discussions at the company, Anthropic came to see the rule as "a business risk" as much as a policy debate, particularly as rivals courted the same customers with looser terms.

How the new safeguards work

Anthropic said it built Enterprise Frontier Safeguards over several months with more than 100 customers spanning financial services, health care, manufacturing, telecommunications, law, retail and the public sector, including the Analysis and Resilience Center for Systemic Risk, a group representing major U.S. banks. The program was developed jointly with Amazon Web Services, Google Cloud and Microsoft Azure, and covers Claude Fable 5, Fable 5.1 and Mythos-class models across Claude Enterprise, Claude Code, the Claude Platform, Amazon Bedrock, Google's Agent Platform and Microsoft Foundry.

Anthropic and AWS confirmed the mechanics in a joint AWS announcement tied to the general release of Claude Fable 5.1: eligible customers can now run Fable models with zero data retention on Bedrock and the Claude Platform on AWS, with prompts and outputs staying inside the customer's own AWS account, access policies and audit logs rather than passing through Anthropic's systems. Anthropic said it will not charge for the program itself, though customers will pay their cloud provider's standard rates for the storage, reads, writes and data transfer the monitoring requires.

Mixed reception

The customers who helped design the system were broadly positive. Philip Martin, chief information security officer at Coinbase, said in Anthropic's announcement that the company had "clearly" listened, adding that the new design put customers "in the driver's seat" over their own data rather than leaving it in Anthropic's hands. Mayank Upadhyay, chief security and trust officer at Snowflake, said the arrangement let his company run "Anthropic's most capable models" while keeping data under keys Snowflake itself controls.

"Data stays in the customer's environment, under keys they control, running Anthropic's most capable models," Upadhyay said.

Not every account of the announcement was as favorable. Trade press coverage pointed out that the new setup shifts real work onto customers, who must now review the automated misuse alerts that Anthropic's own staff previously handled, and noted that broad zero-retention access to Fable models is arriving "by invitation" rather than as a default for every enterprise account, while Mythos users for now remain on the original June terms, according to a technical breakdown of the rollout. Kate Jensen, Anthropic's head of Americas, told reporters the company had put in "hundreds of hours" working directly with customers to reach the new design after what she described as extensive feedback on the original policy.

Who is affected, and what comes next

The change applies to Anthropic's enterprise and platform customers rather than to people using Claude's free, Pro or Max consumer plans, which were never subject to the June retention rule in the same way. It lands as Anthropic leans harder on enterprise revenue, which now makes up the bulk of its business, and as the company is widely reported to be preparing for an eventual public listing. Rivals have faced similar pressure: coverage of the announcement noted OpenAI has been pursuing comparable data-sovereignty arrangements with Microsoft and Databricks, underscoring an industry-wide tension between retaining enough data to catch misuse of increasingly powerful models and satisfying corporate customers who do not want that data leaving their own infrastructure at all.

Enterprise Frontier Safeguards is rolling out in phases starting this fall, with Anthropic aiming for broader availability across eligible customers by the end of the season. Until the full rollout lands, Anthropic said Fable 5 and 5.1 traffic for invited customers will run under temporary zero-data-retention terms. Whether the new system satisfies the regulated industries that objected loudest to the original policy, or simply relocates the same debate to a new set of technical and contractual details, is likely to become clearer only once auditors and compliance teams at banks, hospitals and law firms have had time to review how the automated monitoring performs in practice.