Anthropic to Let Enterprise Clients Store AI Safety Data On-Premises

Anthropic is preparing to give enterprise customers more control over sensitive data generated by its most powerful AI models, according to a person familiar with the company's plans.

The AI startup expects to launch a new safety system later this year that will still require business users to retain data for 30 days but will allow them to store it on their own cloud computing infrastructure rather than Anthropic's servers. The system has been under development for several months, Bloomberg reported.

A reversal from mandatory cloud retention

The change represents a significant shift from Anthropic's June announcement that it would retain all customer data from its advanced Mythos and Fable models—and future frontier systems—for 30 days on its own infrastructure. The company introduced that policy to detect and prevent novel cyberattacks exploiting its technology, promising not to use the data for AI training.

However, Anthropic acknowledged in a recent report that the mandatory retention policy would "be unpopular with customers who have come to expect zero retention, and pose real risks to our business success (especially if competitors do not follow)."

Why it matters

Data residency has become a critical issue for enterprises adopting advanced AI systems, particularly in regulated industries like healthcare, finance, and government. Allowing customers to retain safety monitoring data on their own infrastructure addresses compliance requirements and competitive pressure while preserving Anthropic's ability to monitor for misuse. The move suggests AI safety protocols may need to accommodate enterprise privacy expectations to achieve market adoption.

Competitive pressure mounts

The timing coincides with rival OpenAI's Wednesday announcement that it began testing "private safety processing," a method designed to avoid retaining customer data entirely. OpenAI is piloting the approach with companies including Databricks and investor Microsoft, with broader availability planned for September.

According to the person familiar with Anthropic's plans, the company began developing its on-premises option before customers voiced criticism and has been coordinating with more than 100 clients in highly regulated industries to build the system.

Anthropic declined to comment on the planned policy change.

The details were first reported by Bloomberg's Rachel Metz.