Anthropic hands enterprise data control over to its customers on the road to its IPO

By PYMNTS · September 1, 2026.

Anthropic announced on Tuesday a change to its data retention policy for enterprise customers, following complaints from them. The new system, called Enterprise Frontier Safeguards, gives companies more control over how their data is reviewed, stored and managed, while Anthropic carries out the security monitoring it considers necessary without requiring human review on its part. The company will not charge for it.

In June, Anthropic had introduced a policy requiring 30 days of traffic to be retained on Claude Fable 5 and Mythos 5, two of its most advanced models, arguing that it needed to defend against cyberattacks it described as "complex and novel." The company said it would not use that data for purposes unrelated to security, such as training its models — but the retention itself raised concerns among some of its enterprise customers.

Kate Jensen, Anthropic's head of the Americas, told CNBC that the company spent hundreds of hours working with customers to design an alternative. The result allows automated monitoring without human review by Anthropic, is available both through direct access and via cloud providers, and will be rolled out in phases with the goal of reaching broad availability this fall. The June policy will continue to apply to non-enterprise users of the Mythos models.

The adjustment is no accident. Enterprises generate most of Anthropic's revenue, according to CNBC, and the change comes as the company prepares for an anticipated IPO and competes with rivals in the advanced model market. CEO Dario Amodei himself previewed the system the previous week in an interview with CNBC, explaining that it aims to keep Salesforce customer data private without giving up the safeguards around its models.

Here is the heart of the matter, and it is more interesting than the announcement itself. The more powerful -and potentially dangerous- frontier models become, the more their creators need to monitor their use to detect abuse; but the more monitoring they impose, the more they clash with the confidentiality obligations their customers already have in banking, healthcare or public administration. Anthropic first bet on retention by default and is now shifting toward distributed control: security without direct custody of the data. It is a compromise, not a solved problem, and it will not be the last time we see that pendulum swing.

As sector context, large enterprise AI buyers have long been demanding minimal retention or restricted access to their prompts as a condition for signing contracts; offering less than that, above a certain customer size, means losing the contract to a rival willing to cede that control. That Anthropic is reversing free of charge, as it prepares to go public, a policy it had defended as necessary for security says as much about the real weight of its corporate customers as about how fragile "selling security" is as an argument when whoever pays the bill is not willing to give ground on privacy.

At Zendoric we have flagged this underlying tension before: as the capabilities of frontier models grow, monitoring these systems is no longer a regulatory whim but a real necessity. The challenge -how to monitor without intruding, how to protect without slowing enterprise adoption- is not solved by a press release; it is the kind of friction that will accompany agentic AI for years, while it finishes earning the trust needed to operate in the economy's critical functions. In the short term, this is a policy adjustment with an obvious commercial motive. In the long term, it is exactly the kind of governance detail -boring, technical, unglamorous- on which it depends whether AI becomes trusted infrastructure rather than a risk companies prefer to keep at arm's length.