Contents
On September 1, Anthropic announced it is walking back the data retention policy it introduced in June for its most capable models, after pushback from enterprise customers about how their prompts and outputs were being stored and reviewed. The replacement, called Enterprise Frontier Safeguards, does not just loosen the old rule. It changes where the data physically lives, and that shift is the more interesting engineering story.
This matters beyond Anthropic’s own customer base because it is a concrete answer to a tradeoff every team shipping an LLM product with abuse-detection eventually hits: you want zero data retention for privacy and compliance, but you also want to catch misuse, jailbreak attempts, and the kind of “complex and novel” attacks that only show up when someone reviews traffic after the fact. Those two goals usually fight each other, because reviewing traffic means someone has to hold onto it. EFS is Anthropic’s attempt to get both without the tradeoff.
What actually changed
The old policy kept a 30-day copy on Anthropic’s side. Introduced alongside the Claude Fable 5 and Mythos 5 launches in June, it retained traffic to those models for 30 days specifically to help Anthropic defend against misuse, with a promise the data would not be used for training. Enterprise customers in regulated industries pushed back hard: a 30-day copy sitting with a third party, reviewable by that party’s staff, is a real compliance problem for banks, hospitals, and law firms regardless of what the data is used for.
EFS keeps the retention, moves the custody. The 30-day window for safety monitoring stays, but the data now sits in cloud infrastructure the customer controls, not Anthropic’s. Automated safety classifiers can still run against it and flag abuse, but without requiring a human at Anthropic to review the underlying content. The privacy posture of zero retention, combined with the abuse-detection value of the old policy, is the pitch.
It was built with the customers who complained, and it works across clouds. Anthropic says it developed EFS with more than 100 customers across financial services, healthcare, manufacturing, telecom, law, retail, and the public sector, alongside AWS, Google Cloud, and Microsoft Azure. The controls apply whether you call Claude directly through Anthropic’s API or through a cloud marketplace like Bedrock or Vertex. It is free, rolling out in phases, with broader availability targeted for this fall. Non-enterprise users on Mythos-class models stay under the original June policy for now.
Try it: where does your data actually sit
About the demo
This models the general shape of the three retention modes Anthropic has described, not their exact internal implementation. Pick an access path and a data mode to see who can see your traffic and for how long.Pick an access path and a mode.
Switch the mode to Enterprise Frontier Safeguards and cycle through the access paths. The custody line is the whole point: the same 30-day safety window as the standard policy, but it never crosses the trust boundary into Anthropic’s own infrastructure, on any of the four access paths.
What this means for you
If you are building on Claude, or evaluating any LLM vendor for a regulated workload, EFS is worth reading as a template rather than an Anthropic-specific footnote. The pattern, run the abuse-detection classifier inside the customer’s trust boundary instead of shipping raw traffic to the vendor, is generally applicable to any provider that wants to offer both privacy guarantees and misuse monitoring. If your own product does anything resembling content moderation or abuse detection on user data, ask the same question Anthropic’s enterprise customers asked: does the detection logic actually need to see the raw data centrally, or can it run where the data already lives? My next step is checking whether EFS is available yet for the workloads I run through Bedrock, since a free control that removes a compliance blocker without giving up abuse detection is not something to leave on the table once it ships.