Anthropic's Enterprise Frontier Safeguards: abuse detection for agents, data stays in the customer's cloud
Announced with Fable 5.1, EFS pairs zero data retention with abuse detection for agents. Activity data stays encrypted in the customer's AWS, Google Cloud or Azure storage; Anthropic's systems flag risky patterns.
Anthropic has detailed Enterprise Frontier Safeguards (EFS), published on 1 September alongside Fable 5.1, which it says combines zero data retention with the abuse detection needed to monitor AI agents. Alex Albert of Anthropic called it "ZDR++", zero data retention built for a world of agents. Our Fable 5.1 story mentioned EFS in one sentence; this is the fuller picture.
The problem Anthropic says it is solving
According to the blog post, sophisticated misuse now spreads across sessions and accounts, so inspecting each interaction on its own is no longer enough. With traditional zero data retention, nothing was kept, so patterns across sessions could not be seen at all. Albert framed the same tension from the customer side, writing that companies are giving agents ever more access to internal systems while traditional ZDR gave them no view of what those agents were doing across sessions.
How the data flow works
Under EFS, activity data stays in the customer's own cloud infrastructure, in Amazon S3, Azure Blob Storage or Google Cloud Storage, encrypted with keys the customer manages. Anthropic's automated systems analyze that data and send signals about risky patterns directly to the customer; no Anthropic employee review is required, the company says. Anthropic charges nothing for the feature, while the cloud provider bills storage and processing at standard rates. The rollout is phased, with broad availability targeted by the end of autumn and a form to request access; in the meantime, ZDR remains available on Fable 5 and 5.1.
Albert described EFS as an observability and risk-reduction layer for agents and predicted it would become a standard requirement for enterprises running powerful AI inside their own systems. That is a forecast; the customer need Anthropic itself cites is simpler: the data stays under the customer's control and goes nowhere else. The direction is not Anthropic's alone; on the same day OpenAI announced a similar monitoring layer for Astra. Disclosure: GlobalFeed's own engine runs on Claude.
