Anthropic Plans to Let Businesses Hold Its Required AI Logs

An AI model can be powerful enough to justify strict safeguards and still be unusable if those safeguards send a company’s sensitive work to another vendor. Anthropic plans to separate the two.

On Aug. 20, the company announced safeguards that would operate on business customers’ infrastructure, leaving them in control of where data lives and who can access it. The plan concerns monitoring and safety records, not a self-hosted version of Anthropic’s models.

Anthropic executive Boris Cherny described the bargain directly:

Boris Cherny @bcherny

𝕏

We've been working on this with customers for a while. Mythos-class models require additional safety measures and enterprises need to meet their own privacy and compliance rules. Customers can own and control their own data and Anthropic retains none. It’s coming this fall.

Another Anthropic executive said the system had been developed with roughly 100 companies and was headed for a broader rollout this fall.

The architecture and contractual terms are not yet public, so Anthropic’s central promise, that it will retain no copy, cannot yet be independently assessed.

The rule is staying

Today, prompts sent to Fable 5 and Mythos 5, and the answers they produce, must be retained for 30 days. Anthropic says that history helps its safety systems find campaigns and repeated attempts that can be missed when each request is judged alone.

The 30-day window would remain under the planned design. What changes is custody: the business would hold the record under its own access controls, without a second copy at Anthropic. This is not zero data retention; the content would still exist for the required period.

Cloud customers do not have that boundary today. Amazon Web Services says covered prompts and completions are shared with Anthropic and retained for up to 30 days. Google requires prompt-response sharing with Anthropic, while Microsoft warns that Anthropic stores the data under separate terms, outside the customer’s Microsoft agreement.

For the covered models, Anthropic’s zero-data-retention option is currently unavailable. A company whose privacy or compliance rules prevent Anthropic from holding sensitive prompts can therefore be locked out of Fable and Mythos even if it wants to use them. Customer custody could remove that barrier without eliminating the 30-day view Anthropic says it needs.

Two labs draw the line differently

Anthropic is not alone in trying to watch for abuse without taking possession of customer content. One day earlier, Onset reported that OpenAI had previewed Private Safety Processing, which examines patterns across related requests inside customer-controlled, encrypted infrastructure.

Both labs have reached the same basic safety conclusion: dangerous activity is easier to detect across a sequence than one request at a time. Their proposed privacy boundaries differ. Anthropic would preserve 30 days of prompts and outputs under customer custody; OpenAI says the content remains under zero-data-retention terms and only limited safety signals, such as an alert’s category and severity, leave the customer environment.

Anthropic’s proposed answer is to keep the 30 days, and let the customer keep them.