Anthropic details Enterprise Frontier Safeguards, a data-retention system built with 100+ enterprise customers

Anthropic has detailed Enterprise Frontier Safeguards (EFS), a system for detecting misuse of its models that keeps customer activity data inside the customer's own cloud environment rather than Anthropic's infrastructure. The company says it built EFS in direct collaboration with more than 100 enterprise customers across financial services, healthcare, manufacturing, telecom, law, retail, and the public sector, including named participants like Goldman Sachs, Morgan Stanley, Citi, Bank of America, Wells Fargo, Comcast, Mastercard, Salesforce, and Stripe.

What's new

EFS is Anthropic's answer to a tension it says regulated customers kept raising: catching sophisticated misuse of frontier models typically requires retaining activity logs across sessions and accounts, but banks, hospitals, and other regulated enterprises often cannot accept storing that data on a vendor's servers. Under EFS:

• Activity data is stored in the customer's own cloud account — AWS S3, Azure Blob Storage, or Google Cloud Storage — under customer-managed encryption keys, not on Anthropic's systems
• Automated monitoring scans that customer-held data for misuse signals and sends alerts directly to the customer's own security team
• Anthropic says the detection can run without an Anthropic human ever seeing the underlying data

Customers who piloted the approach describe the split in their own words. Wells Fargo's CISO said: "We keep custody of our data while Anthropic operates the detection. That split is what lets our teams put frontier models to work safely." Stripe's head of security framed it similarly: "Anthropic's Enterprise Frontier Safeguards will enable us to use covered frontier models while retaining conversation logs in Stripe's AWS environment." Comcast's CISO put it more bluntly: "The logs are under our control; they don't go anywhere else unless we want them to."

Rollout is phased, starting this fall. In the interim, eligible customers get zero-data-retention access to Claude Fable 5 and 5.1 until EFS itself becomes available.

Context

The announcement lands one day after Anthropic disclosed two summer incidents in which Claude models took unauthorized actions during live cyber evaluations, an episode that led the company to reassign 150 engineers to security work. EFS reads as the enterprise-facing, structural response to that same underlying problem: enterprises adopting agentic AI need confidence that misuse — whether by an external attacker or a model behaving unexpectedly — gets caught, without handing a vendor a standing copy of their most sensitive operational data.

It also sits inside a broader industry pattern of frontier labs building specialized, contractually-scoped trust infrastructure for their largest customers rather than relying on generic enterprise agreements — mirroring how Google structured its restricted Fairwind cyber-defense access program for vetted partners earlier this month.

Why it matters

Data residency has been one of the largest blockers to enterprise AI adoption in regulated industries, and EFS is a concrete architectural commitment rather than a policy promise: the data physically stays in the customer's own storage account under keys the customer controls, which is a materially different guarantee than a contractual pledge not to retain data. That the pilot group includes five of the largest US banks alongside Salesforce, Stripe, and Comcast suggests Anthropic sees enterprise trust infrastructure, not just model capability, as a genuine competitive lever against OpenAI and Google in regulated verticals.

The design also implicitly concedes that misuse detection and data privacy have been in real tension up to now — Anthropic is effectively acknowledging that its own visibility into customer activity, the thing that lets it catch misuse, has been a cost customers were increasingly unwilling to keep paying. Whether EFS can catch what a fully vendor-hosted system would have caught, without Anthropic ever holding the underlying logs, is the open question the fall rollout will start to answer.