TL;DR

Anthropic has announced Enterprise Frontier Safeguards, a design that gives regulated customers zero-retention privacy without giving up misuse detection across their traffic. The trick is location: activity logs sit inside an account the buyer already owns, under encryption keys the buyer holds, and Anthropic’s automated systems reach in to analyse them. Phased rollout begins later this autumn.

The problem it exists to solve

Detecting serious abuse is not a per-request task. Credential theft and the more sophisticated attacks unfold over many accounts and sessions, so spotting them means correlating behaviour over time — which means keeping the data. Anthropic introduced a 30-day retention window alongside Fable 5 for exactly that reason, stressing that enterprise material has never fed its training without explicit permission.

Regulated buyers were unmoved. Adding another trusted data custodian means renegotiating contracts, notifying their own customers and satisfying internal audit rules, and for banks and hospitals that overhead often exceeded the benefit. The compromise architecture puts the logs in Amazon S3, Google Cloud Storage or Azure Blob Storage under the buyer’s control, with alerts routed straight to the buyer’s own security team rather than to a reviewer at the model provider. No Anthropic employee reads the flags.

Who shaped it

More than 100 customers were involved, alongside all three major cloud providers. The financial sector did much of the heavy lifting through ARC, a systemic-risk body whose membership includes the security chiefs of the largest American banks — Citi, Wells Fargo, Morgan Stanley and Goldman Sachs among them. Anthropic says the design talks reached roughly 25 of the Fortune 100.

The recurring demand was jurisdictional rather than technical: key custody, the limits of what automated review may see, and exactly when - if ever - a person gets to read any of it.

Looking forward

Two labs published safeguard architectures within hours of each other, and they diverge instructively. OpenAI’s Astra disclosure keeps detection and access-gating in-house. Anthropic’s hands the evidence, and the first review, to the buyer — which is closer to how UK financial and health institutions already run insider-risk monitoring, and a better fit for organisations answerable to a domestic regulator. It also relocates responsibility: once the flags arrive in your bucket, acting on them is your problem.