The engineering claim underneath this is about a correlation window. Anthropic's position is that the most sophisticated misuse spans many tasks across multiple sessions and accounts, so scoring each interaction on its own and discarding the data at once does not find it, and effective detection needs data held long enough to correlate across time and accounts [5]. That is the stated reason retention arrived at all [6]. What regulated buyers objected to was not the detection but the custody: taking on another trusted data vendor means notifying their own customers, reopening contracts and satisfying internal rules for storing and auditing sensitive material [11]. EFS answers the custody objection and leaves the correlation window intact.
The composition of the design group tells you which objection was binding. Anthropic says its conversations spanned a quarter of the Fortune 100 and every US global systemically important bank [12], which works out to roughly 25 of those hundred firms [1], and that the groups it worked with included ARC, whose membership includes the chief information security officers of Goldman Sachs, Morgan Stanley, Citi, Bank of America and Wells Fargo [9]. It also cites work with leaders at Comcast, KPMG, Mastercard, Salesforce and Visa [16]. A design fitted to the buyers with the heaviest vendor-inventory duties becomes the reference design for everyone carrying lighter ones.
What actually changes is narrower than a retention rollback. The data is still kept for a meaningful period and still monitored; what moves is the location of the store and the reviewer of the output, with flagged patterns routed to the customer for review [2][10]. On the adjacent question, Anthropic states that retention was never motivated by a wish to train on enterprise data and that it has never trained on enterprise data without explicit permission [7]. That is the company's own account, and this announcement is the whole of the record for it.
A privacy blocker has been cleared, but that is not where the work ends. Once the store sits in the customer's cloud, the completeness of misuse detection depends on the customer's own retention and logging discipline, and triage depends on a security team that already runs an insider-risk queue [10]. Anthropic reports seeing credential theft and agents autonomously engaging in destructive behaviour among recent misuse attempts [14], and the buyer is now the party that sees those signals first.
The record does not settle the operational terms. The announcement does not say how long data must persist in customer-controlled storage, who holds the keys, what happens when a customer receives a signal and declines to act, or what EFS costs [3]. Rollout is phased from later this fall, and eligible customers keep zero data retention on Fable 5 and Fable 5.1 until it is ready [3][4], so the interim arrangement hands those buyers privacy without the cross-account correlation Anthropic argues detection requires [2]. A firm that takes the bridge this quarter is also accepting a migration next quarter into storage it has to provision, audit and staff, and that cost sits with the security organisation rather than with the vendor.
