Anthropic's Enterprise Frontier Safeguards (EFS), announced September 1, 2026, stores Claude conversation logs in the customer's own cloud infrastructure -- Amazon S3, Azure Blob Storage, or Google Cloud Storage -- while Anthropic's automated systems run cross-session misuse detection. No Anthropic employee reads the data. No extra charge. Rolling out fall 2026.
The compliance deadlock in enterprise AI has been real for years. Regulated enterprises -- banks, hospitals, law firms -- couldn't use frontier models because zero data retention meant no audit trail, and handing logs to Anthropic violated their own data governance policies. EFS is the architecture that breaks that stalemate: the customer holds the data, Anthropic operates the detection.
What Did Anthropic Actually Ship With EFS?
Enterprise Frontier Safeguards is a new data custody model for Claude API users. Instead of routing conversation logs through Anthropic's own storage, EFS writes them directly to the customer's cloud account -- S3, Azure Blob Storage, or GCS -- under the customer's encryption keys, access policies, and audit logging. Anthropic's automated detection systems connect to those buckets read-only, analyze a rolling 30-day window for misuse signals, and send any alerts to the customer's security team. No Anthropic employee reads the conversation content or the alerts.
All four controls are opt-in: customer-owned storage, customer-managed encryption keys, fully automated (zero human) review, and the 30-day cross-session detection window. None of this changes model behavior, API pricing, or rate limits.
Anthropic developed EFS in collaboration with more than 100 enterprise customers across financial services, healthcare, manufacturing, telecom, legal, retail, and the public sector. The co-developer group spans a quarter of the Fortune 100 and every US global systemically important bank. Named participants include Goldman Sachs, Morgan Stanley, Citi, Bank of America, Wells Fargo, Comcast, KPMG, Mastercard, Salesforce, Visa, and Stripe. The Analysis and Resilience Center -- a group whose membership includes the CISOs of the largest US financial institutions -- worked directly with Anthropic to define what compliant frontier AI deployment would require.
EFS is offered at no additional charge by Anthropic. Customers pay only standard cloud provider fees for storage, read/write operations, and data egress in their own S3, Azure Blob, or GCS bucket.
Why Zero Data Retention Was Blocking Enterprise AI Adoption
Anthropic's prior zero data retention (ZDR) approach deleted conversation data immediately after processing. That solved the data residency concern but created two different problems for regulated industries: no audit trail, and no cross-session visibility for misuse detection. Both matter enormously in enterprise deployments, and prior ZDR solved neither.
Regulated enterprises aren't just cautious about vendors holding their data -- they're often legally required to hold it themselves. A bank under regulatory examination must demonstrate what its AI systems were doing across weeks or months of usage. A hospital deploying Claude in clinical workflows needs HIPAA-compliant audit logs. A law firm processing client matters under AI needs those records under its own custody and retention schedule, not deleted by the vendor. Zero data retention was the wrong answer for these organizations -- not because retention is dangerous, but because the wrong entity was doing the deleting.
The misuse detection problem ran in parallel. The most sophisticated attacks on enterprise AI systems don't happen in a single session. They spread across multiple sessions, multiple accounts, and multiple days -- often using stolen or leaked enterprise credentials. A classifier that analyzes each interaction in isolation and then immediately discards it can't detect that pattern. Correlation across time requires a window. Prior ZDR made that window impossible.
Munish Kumar Sharma, Wells Fargo's CISO, put the EFS value proposition cleanly: "Our logs stay in a Wells-managed environment under Wells-managed keys. We keep custody of our data while Anthropic operates the detection." That sentence is the architecture in two clauses.
How the Split Custody Architecture Actually Works
The EFS data flow is straightforward once you see the split. Claude processes API requests normally. Instead of logs routing to Anthropic's infrastructure, session data writes to the customer's configured cloud storage bucket -- S3, Azure Blob, or GCS -- using customer-managed encryption keys (CMEK). The customer controls the access policies on that bucket. Anthropic's automated classifiers have read-only access, analyze the 30-day rolling window, and push detected signals back to the customer's security team. The underlying conversation content never passes through Anthropic's hands or any Anthropic employee's review queue.
The "split" in split custody means neither party can operate the full system alone. Anthropic can't access the logs -- the customer controls the encryption keys. The customer can't run the cross-session misuse detection -- Anthropic built and operates the classifiers. Each party contributes something the other can't replicate, and the result is a system where compliance and detection coexist without either party holding everything.
All monitoring is fully automated and explicitly designed with no Anthropic human review path. When the classifiers flag something, the alert goes to the customer's security team -- not Anthropic's. The customer decides what to do with it. This keeps detection expertise with Anthropic (who built the models and understands their misuse patterns) while keeping response authority with the customer (who owns the infrastructure and the affected accounts).
Anthropic holds ISO 27001:2022 certification for information security management and ISO/IEC 42001:2023 for AI management -- two certifications that procurement teams in financial services and healthcare require as baseline vendor qualifications. EFS is designed to work alongside these certifications and the compliance postures built around them.
Get the AI Agent Briefing
One email per week. The best AI agent news, tutorials, and tools -- written by someone who actually builds with them.
Subscribe Free
Which Platforms and Enterprises Does EFS Cover?
EFS covers the full range of ways enterprises run Claude in production. Supported deployments include Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google's Agent Platform, and Microsoft Foundry. That's all three major cloud providers and both the direct API and managed service paths. If your team is using Claude in production today through any of these surfaces, EFS applies to your deployment.
Access is by invitation during the initial rollout phase. For eligible customers waiting on EFS, Anthropic is offering Fable 5 and Fable 5.1 zero data retention as an interim solution until their EFS onboarding is ready. Broad availability is targeted for fall 2026, though Anthropic hasn't published a timeline for organizations outside the Fortune 100 co-developer group.
The industries represented in the co-development process -- financial services, healthcare, manufacturing, telecom, legal, retail, public sector -- are the clearest signal of who the initial invitations will prioritize. If your company is in a regulated vertical and has an existing Claude Enterprise relationship, the path to EFS access runs through your Anthropic account team.
What the Misuse Detection Is Actually Looking For
Anthropic's EFS classifiers target three specific categories of serious misuse: attempts to develop offensive cyber capabilities, attempts to develop biological weapons capabilities, and evidence of stolen or compromised enterprise credentials being used across multiple accounts or sessions. These aren't content policy violations -- they're threat intelligence categories, and they're what the 30-day cross-session window is specifically designed to catch.
The credential theft scenario is the clearest example of why session duration matters. When an attacker acquires leaked enterprise API credentials, they don't use them in a single session. They run a distributed series of queries -- probing capability boundaries, testing what the credentials unlock, spreading activity across user accounts to avoid per-session rate limits. A 30-day rolling window across an organization's accounts can detect that pattern. Single-session analysis that discards context immediately cannot.
Anthropic's documented observation from monitoring its own systems is that the most sophisticated misuse "spreads across many tasks, sessions, and accounts." EFS is explicitly architected around detecting that distributed pattern, not flagging individual bad prompts. Individual prompt filtering already runs at inference time. EFS adds the layer above it.
What EFS Doesn't Answer Yet
EFS is a real architectural advance, and the split custody model solves a problem that was genuinely blocking enterprise Claude adoption. That said, several operational questions aren't publicly documented yet, and I'd want answers to these before building a deployment plan around EFS.
The invitation process has no published criteria or waitlist for organizations outside the Fortune 100 co-developer group. "Later this fall" and "by invitation" is the entire public timeline. If you're a mid-market SaaS company that wants EFS, there's no clear path to get it right now beyond contacting your Anthropic account rep.
The operational integration story is thin. The 30-day retention window is described as a rolling window, but Anthropic hasn't published how customers configure its duration, what the minimum and maximum settings are, what alert format EFS signals arrive in, or how those alerts integrate with existing SIEM platforms and SOC workflows. Security teams building deployment plans need those details.
The Register noted a fair concern: Anthropic's zero data retention guarantee requires enterprises to verify it themselves, not just trust the vendor's word. The technical controls that let a customer confirm that conversation data stayed in their bucket and wasn't copied to Anthropic's infrastructure aren't documented publicly. Trust-but-verify is the right default here, and that verification capability needs documentation before security teams can sign off on EFS for their most sensitive workflows.
FAQ
What is Anthropic's Enterprise Frontier Safeguards (EFS)?
EFS is Anthropic's split custody data model for Claude enterprise deployments, announced September 1, 2026. Customers store conversation logs in their own cloud accounts -- Amazon S3, Azure Blob Storage, or Google Cloud Storage -- under their own encryption keys. Anthropic's fully automated systems analyze a 30-day rolling window of that data for cross-session misuse, including offensive cyber attempts, bioweapons research, and credential theft patterns. No Anthropic employee accesses the data. No extra charge from Anthropic. Rolling out fall 2026 by invitation.
How is EFS different from Anthropic's prior zero data retention (ZDR)?
Prior ZDR deleted conversation data immediately after each interaction -- no logs anywhere, no audit trail possible, and no cross-session misuse detection. EFS retains a 30-day rolling window but stores it in the customer's own cloud infrastructure under the customer's encryption keys, not Anthropic's. Regulated enterprises that need audit trails for compliance (HIPAA, banking regulators, legal records) but can't hand logs to a vendor now have a path to run Claude at frontier capability without violating data governance policies.
Which cloud platforms and Claude products support EFS?
EFS supports Amazon S3, Azure Blob Storage, and Google Cloud Storage for data residency. It covers Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google's Agent Platform, and Microsoft Foundry. All three major cloud providers and all major Claude enterprise deployment paths are included. Access during the fall 2026 rollout is invitation-based, with Fable 5 and 5.1 ZDR available as an interim option for eligible customers.
Does EFS cost extra?
No. Anthropic offers EFS at no additional charge. Customers pay only standard cloud provider fees for storage, read/write operations, and data egress in their own S3, Azure Blob, or GCS bucket. No changes to Claude model pricing, API rate limits, or model behavior come with EFS. The cost is your cloud storage bill, not an Anthropic line item.
Which industries is EFS designed for?
EFS was co-designed with 100+ enterprises across financial services, healthcare, manufacturing, telecom, legal, retail, and the public sector -- with specific input from every US global systemically important bank and a quarter of the Fortune 100. Any organization in a regulated industry that needs to run frontier AI while maintaining their own audit trail and data custody will benefit from EFS over prior zero data retention approaches.
Get the AI Agent Briefing
One email per week. The best AI agent news, tutorials, and tools -- written by someone who actually builds with them.
Subscribe Free