Anthropic plans to let enterprises host their own 30-day AI safety-retention data

Anthropic reportedly plans to let enterprise customers of its most capable models store required 30-day safety-retention data on their own cloud infrastructure rather than on Anthropic’s servers. A new safety system expected later in 2026 will still mandate the 30-day window but shift where the logs live.

The change would soften a policy that unsettled corporate buyers. Anthropic announced in June 2026 that it would retain 30 days of all traffic on covered models — its Fable and Mythos families and future frontier systems — across every platform to help detect cyberattacks. For companies with strict data-handling rules, blanket retention on a vendor’s servers was a hard sell.

Under the revised approach, that data would sit on infrastructure the customer controls. Anthropic has reportedly coordinated with more than 100 enterprise customers, including Salesforce, on the design, which has been in development for months.

Anthropic’s data-retention documentation for covered models states that retained data is inaccessible to its personnel by default, except through a logged and flagged safety review, and is automatically deleted after 30 days unless a legal or safety hold applies. The reported change would keep those handling rules but move the storage location, addressing where the data sits rather than how long it is kept.

The specifics remain unconfirmed by Anthropic, which has not published a first-party announcement of the change; the rollout is expected later in 2026. The open question is whether customer-hosted logs give Anthropic enough visibility to run the cyberattack detection the original policy was built to enable.