Enterprise Frontier Safeguards: your bucket, your keys
Anthropic's Enterprise Frontier Safeguards keeps AI monitoring data in your own S3 or Blob Storage under your keys. It costs nothing. Your cloud bill isn't nothing.
Anthropic shipped Enterprise Frontier Safeguards (EFS), and the architecture is the story: misuse monitoring data lives in your cloud account, under your encryption keys, with your audit logging. Anthropic never takes custody of it. This replaces a data retention policy that regulated customers had been pushing back on hard, as CNBC reported.
What actually happened
The old bargain was blunt: you got safety monitoring, and Anthropic held the data to do it. Companies in healthcare, finance, and legal could not sign that. Zero data retention meant no monitoring; monitoring meant no zero data retention.
EFS unbundles the two. Activity data is written to a bucket you own — Amazon S3, Azure Blob Storage, or Google Cloud Storage — encrypted with keys you control. Automated systems analyze traffic across sessions and accounts on a rolling window, looking for credential theft, sophisticated attack patterns, and capability development attempts. When something trips, the signal goes to you. No Anthropic employee reviews it. Your team evaluates every flag.
Coverage spans Claude Code, Claude Enterprise, and the Claude Platform, plus Amazon Bedrock, Google's Agent Platform, and Microsoft Foundry. It targets Mythos-class models — Claude Fable 5.1 and newer.
Anthropic charges nothing for EFS. Your cloud provider still bills you at standard rates for storage, reads, writes, and egress, which is the line item nobody budgets for. Rollout is phased through fall 2026 and gated by a request form. Until your account is live, eligible customers get plain zero data retention on Fable 5 and Fable 5.1.
Why data custody matters for your business
Most small companies signed an AI vendor agreement without reading where the transcripts go. Then a client asks, or an auditor asks, and the honest answer is "we don't know." EFS is the shape that answer should take: named bucket, named region, keys you rotate, access logs you can produce.
But read the tradeoff before you celebrate. You are now the security operations team. Anthropic sends signals; nobody triages them for you. A flag that lands in an unwatched S3 bucket is worse than no flag, because you can no longer claim you did not know. Before you submit the form, decide who reads the alerts, on what schedule, and what happens when one fires at 2am.
The pattern generalizes past Anthropic. Vendor-held telemetry is a liability you inherit at renewal. Customer-held telemetry is an operational job you have to staff. Pick deliberately, write it into the contract, and put a name next to the alert queue.
Key takeaways
- EFS stores AI misuse-monitoring data in your own S3, Azure Blob, or GCS bucket under your encryption keys — never on Anthropic infrastructure
- Automated analysis runs across sessions and accounts on a rolling window; detected signals are sent to you, with no Anthropic human review
- Anthropic charges nothing for EFS; your cloud provider bills standard storage, read, write, and egress rates
- Covers Claude Code, Claude Enterprise, Claude Platform, Bedrock, Google's Agent Platform, and Microsoft Foundry on Fable 5.1 and newer
- Phased rollout through fall 2026 by request form; Fable 5 and 5.1 get interim zero data retention for eligible customers
- The catch: you own alert triage now — assign an owner before you enable it
Data custody is an architecture decision, not a checkbox. We map where your AI traffic actually lands, wire the alert path to a human who will read it, and build systems you can hand an auditor. Talk to us about AI governance or see what we build.
Sources: Anthropic: Developing Enterprise Frontier Safeguards with our customers, CNBC.
- #anthropic
- #data-retention
- #compliance
- #claude
- #governance
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
AI agents ran a full ransomware breach in under 10 hours
Unit 42 documented an AI-driven ransomware intrusion that took under 10 hours instead of two weeks, used 50+ ATT&CK techniques, and left an 80-page audit.
Read itStop Rogue AI Act wants a machine-readable agent inventory
The Stop Rogue AI Act directs NIST to set AI agent security standards — machine-readable inventory, tamper-proof action logs, and teeth for federal contractors.
Read it