Anthropic said Tuesday it is rolling back the data-retention policy it introduced in June for its most advanced Claude models, after what the company described as extensive pushback from business customers uneasy about their data sitting on Anthropic's servers for 30 days by default.
The original policy, introduced alongside the June launch of Claude Fable 5 and Claude Mythos 5, required 30-day retention of all traffic sent to those models so Anthropic could screen for what it called complex and novel misuse and cyberattacks, even though the company pledged not to use the retained data to train its models. Enterprise customers, whose contracts increasingly hinge on strict data-handling guarantees, objected to Anthropic holding their prompts and outputs at all.
A Concession to Enterprise Customers
Under the new program, called Enterprise Frontier Safeguards, the 30-day retention window stays in place, but the data itself will sit within the customer's own cloud environment rather than on Anthropic's infrastructure. Businesses will also be able to run automated safety monitoring on that data without a human reviewer at Anthropic ever seeing it. Anthropic said it will not charge extra for the new safeguards, that they will be available whether customers access Claude directly or through cloud partners, and that it is targeting broader availability this fall.
Kate Jensen, Anthropic's head of Americas, said the company "spent hundreds of hours working with customers to develop an alternative" to the original policy. Chief executive Dario Amodei framed the change around a specific customer relationship, saying the new system is meant to help keep Salesforce customer data private while allowing Anthropic to maintain its safeguards.
The reversal comes as Anthropic leans more heavily on enterprise revenue and, according to people familiar with the company's plans, prepares for an eventual initial public offering — making it more sensitive to large corporate customers' objections than it might once have been. The original policy had drawn criticism as an aggressive stance relative to rivals including OpenAI, which offers shorter default retention windows for enterprise accounts. Anthropic has not said whether the concession will extend to individual or developer-tier users, who remain subject to the original 30-day policy.
The episode illustrates a recurring tension for AI labs as they add data retention meant to catch misuse of frontier models while courting corporate customers that increasingly treat data sovereignty as a non-negotiable procurement requirement. How Anthropic executes the shift, and whether competitors respond with similar customer-hosted arrangements, is likely to shape how enterprises weigh AI vendors as adoption of frontier models accelerates.