Anthropic is relocating a data retention practice, not ending it. Enterprise customers who use Mythos, Fable, and future flagship models will keep the same 30 day security hold, but from this fall the storage moves onto infrastructure the customer already controls, rather than Anthropic’s.
The policy itself dates to June. Anthropic began storing all customer data from those models for 30 days specifically to detect new cyberattacks that abuse the technology, a scope broad enough to sweep in regulated industries handling sensitive records. The retention window is not shrinking. Anthropic’s own infrastructure simply stops being the destination once the fall rollout lands, with each customer’s cloud environment taking over as custodian instead.
What makes this notable is Anthropic’s own account of why. A report cited by Bloomberg quotes the company characterizing the rule itself as a source of friction with customers and, in blunter terms, a threat to its business, an unusually candid concession from a lab that built its brand on safety credibility. Companies rarely put that kind of liability language about their own policies on record. Building the replacement took months and involved more than 100 customers from regulated sectors, according to Bloomberg’s reporting relayed by The Decoder.
Anthropic developer Boris Cherny confirmed the plans publicly on X, adding a second, independent line of confirmation beyond the Bloomberg account.
The shift lands one day after OpenAI previewed Private Safety Processing, its own attempt to keep safety monitoring intact under zero data retention commitments. The two approaches differ: OpenAI is working the problem with Databricks and Microsoft on a technical layer that inspects activity without persisting it, while Anthropic is keeping its 30 day retention model but moving custody to the customer. Both labs are solving the identical tension between security monitoring and enterprise data control, and both did it within days of each other. That timing is the real signal. When two competing frontier labs ship structurally different fixes to the same enterprise objection in the same week, it means the objection was strong enough to move both roadmaps, not a coincidence of parallel engineering.
The unanswered question is technical: who administers the encryption and access controls once data sits in a customer’s own cloud rather than Anthropic’s. Neither the report nor Cherny’s confirmation addresses whether Anthropic retains any operational access to run its cyberattack detection once the data leaves its infrastructure, which is the mechanism that determines whether this change actually satisfies the regulated customers it is meant to win back.
For enterprise buyers in finance, healthcare, and government who paused or slow walked Claude deployments over the June policy, this fall’s rollout is the moment to re-run procurement review. For competitors, the lesson is now doubled: whatever compromise you offer between safety telemetry and customer data control had better hold up under the same regulated-industry scrutiny that made Anthropic call its own rule a business risk.
The Decoder (Maximilian Schreiner), published August 21, 2026, citing Bloomberg’s reporting and a public confirmation from Anthropic developer Boris Cherny on X.
