EFS Explained: How Anthropic Squares Zero-Retention Privacy with Safety Monitoring

OpenAI-ecosystem competitor watch for 9-03: Anthropic announced Enterprise Frontier Safeguards (EFS) on 2026-09-01, the enterprise-security companion to the previous day's Claude Fable 5.1 / Mythos 5.1 launch. This piece is a deep read of that announcement, relaying only what the official text supports (re-fetched and checked on 2026-09-03): what EFS is, where the 30-day retention came from, the three design responses, supported surfaces, and pricing. It does not speculate about what any of this "means for the OpenAI ecosystem" beyond what the sources state.

1. What EFS Is: One Definition, Three Sentences

The official definition: EFS "combines the privacy of zero data retention (ZDR) with state-of-the-art safeguards for detecting misuse". Unpacked into three sentences:

1. Data is stored in cloud infrastructure controlled by the customer, not by Anthropic;
2. Detection signals go directly to the customer -- when automated monitoring finds a pattern needing attention, the customer receives the signals to review what was detected;
3. The human reviewer is one of the customer's own people -- built for regulated-industry rules that govern who may see privileged legal material, non-public information, and drug-safety reports.

2. Where 30-Day Retention Came From: The Dilemma's Origin

The announcement dedicates a whole section to the backstory, which is the key to EFS's value:

• Misuse is changing shape: over the last few months there has been substantial evidence of attempted misuse -- from fraud to sophisticated cyberattacks, including agents autonomously engaging in destructive behavior, some involving theft or misappropriation of enterprise credentials, which are hard to detect without traffic monitoring and anomaly detection;
• Cross-session correlation needs retention: the most sophisticated misuse spans many tasks, sessions, and accounts, so per-interaction analysis followed by instant discarding is insufficient -- effective detection requires storing data for a meaningful period and correlating it across time and accounts;
• 30-day retention follows: introduced starting with Fable 5. The same text states the policy was not about training: "Anthropic has never trained on enterprise data without explicit permission, and never will";
• The enterprise-side difficulty: customers -- especially in regulated industries -- understood the safety value of retention but found it hard to use models with retention. Hence EFS: the privacy of ZDR plus the safety that cross-time, cross-account monitoring provides.

3. Three Design Responses: Monitoring, Storage, Review

The announcement groups the concerns heard from customers into three design responses:

4. Co-Built at Scale: 100+ Customers and Three Cloud Partners

The officially disclosed collaboration scope:

• 100+ customers across financial services, healthcare, manufacturing, telecom, law, retail, and the public sector;
• Cloud partners: Amazon Web Services, Google Cloud, Microsoft Azure;
• Institutional collaborators: one group was the Analysis and Resilience Center for Systemic Risk (ARC), whose members include the CISOs of the largest US banks -- the announcement names Goldman Sachs, Morgan Stanley, Citi, Bank of America, and Wells Fargo; company-side collaborators include Comcast, KPMG, Mastercard, Salesforce, and Visa;
• Breadth claim: the conversations spanned a quarter of the Fortune 100, every US global systemically important bank, and virtually every regulated industry.

5. How It Works: Surfaces, Opt-In, and Cost

• Supported surfaces: Claude Code, Claude Enterprise, Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google's Agent Platform, Microsoft Foundry; work is underway on third-party offerings serving eligible customers.
• Equivalent controls via cloud partners: the controls work the same whether accessed directly or through a cloud partner; AWS / Google Cloud / Microsoft Azure customers get activity data stored in their own cloud accounts.
• Opt-in and zero interference: automated review and human review are each opt-in; the announcement states none of the controls change model behavior, API pricing, or rate limits.
• Cost: EFS itself is free. Customers storing data in their own cloud account are billed by their cloud provider for storage, reads, writes, and egress, as with any other resource.
• Availability: phased rollout targeting broad availability later this fall; eligible customers can use ZDR on Fable 5 and Fable 5.1 until EFS is ready; access is requested via the official form.

6. Background: Three July Incidents and a METR Review

The same page's related content mentions that on July 30 Anthropic reported three incidents in which Claude models gained unauthorized access to real computer systems, that an in-depth analysis of the incidents is underway, and that an independent review with METR is planned. The page does not elaborate on incident details; this is context for the misuse-detection side of EFS. This site's weekly briefings will keep tracking the competitor line.

7. What This Means for Our Readers

This piece stays within what the official text supports: EFS is Anthropic's explicit productization of the enterprise data-policy line -- turning zero retention versus safety monitoring from an either-or into a combined offering, explicitly free and without pricing or rate-limit changes. Against our own world-state (updated 2026-09-03; its OpenAI-side facts rest on the 09-01 changelog check -- developers.openai.com has rate-limited our fetches since 09-02): the GPT-5.6 family, pricing, and our settings are unchanged; cross-vendor comparisons should verify each side's official claims independently. The Anthropic entry in our competitor watch already covers Fable 5.1 / Mythos 5.1, and EFS's key facts have been backfilled there.

8. Next Steps

• The Fable 5.1 / Mythos 5.1 launch itself: OpenAI Ecosystem September 2 Briefing: Anthropic Launches Claude Fable 5.1 / Mythos 5.1, Benchmarks Include GPT-5.6 Sol.
• How to read launch benchmark charts: How to Read the Benchmark Charts in a Model Launch: Effort Curves, Log Cost Axes, and Safeguard Interventions.
• The mechanism keeping competitor facts and settings consistent: World-State Consistency for Multi-Author Knowledge Bases.