Anthropic introduces zero-retention AI safety monitoring for enterprises

Anthropic is introducing a new framework aimed at helping enterprises monitor AI misuse without ceding control over sensitive data, as organizations struggle to balance security visibility with strict compliance requirements.

The company announced a new solution called Enterprise Frontier Safeguards (EFS), which “combines the privacy of zero data retention (ZDR) with state-of-the-art safeguards for detecting misuse,” the company said in a blog post.

Under the approach, activity data used for monitoring will be stored in cloud infrastructure “controlled by the customer, not Anthropic,” the company added.

The feature will be rolled out in phases later this fall and will be offered to eligible customers.

To bridge the transition, Anthropic said customers will receive zero data retention on its Fable 5 and newly launched Fable 5.1 models until EFS is rolled out.

“EFS will be supported on Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google’s Agent Platform, and Microsoft Foundry,” the post added.

This came days after OpenAI introduced a similar approach to limit data retention while detecting abuse.

Customer control, but shared responsibility

The company said EFS is designed to give enterprises control over how monitoring data is stored and reviewed, while still allowing Anthropic to detect misuse patterns.

“With EFS, customers control how data gets reviewed,” Anthropic said adding automated systems flag suspicious activity and send those signals to customer teams.

“EFS has automated safety monitoring, no Anthropic human review required,” the company added.

Jaishiv Prakash, director analyst at Gartner, said the shift could help address compliance concerns.

“The shift is important because it neutralizes the compliance barriers that have prevented highly regulated enterprises from adopting frontier AI models,” Prakash said. “By shifting data custody away from centralized vendor environments, organizations maintain definitive governance over sensitive information.”

Sanchit Vir Gogia, chief analyst at Greyhound Research, said enterprises should not assume that data control means full visibility.

“Custody should not be confused with visibility,” Gogia said. “Anthropic still defines the safety framework and operates the detector.”

“Every buyer must therefore ask who holds the data, who holds the key, who operates the detector, who decides what an alert means, and who enforces the consequence,” he said.

Operational burden shifts to enterprises

Anthropic said its systems analyze activity across sessions and accounts for signals such as attempts to develop “offensive cyber or biological capabilities” or signs of “stolen or leaked credentials.” Alerts are sent to customer teams for review.

“Detection is automated under this model while accountability is not,” Gogia said, adding that the system creates an internal workflow for reviewing alerts without disclosing volumes or false-positive rates.

Prakash said the same shift increases operational demands.

“The model fundamentally shifts the operational burden of alert triage and incident response directly to the enterprise,” he said. “Security and risk leaders must proactively invest in AI-specific runbooks and adequately staff their operations centers.”

Reframing the data retention trade-off

Anthropic said it has seen “substantial evidence of attempted misuse… from typical forms of abuse, such as fraud, to sophisticated cyberattacks,” including cases involving “theft or misappropriation of enterprise customers’ credentials.”

“Effective detection requires storing data for a meaningful period of time so that it can be correlated across time and accounts,” the company said in the blog post.

Gogia said EFS does not remove that requirement.

“Enterprise Frontier Safeguards relocates the retained evidence rather than removing the need for it,” he said, describing the model as “provider-side zero data retention alongside customer-custodied retention.”

“Two frontier labs converging on one architecture within a fortnight makes the design a pattern rather than a differentiator,” Gogia said.

Fits alongside existing enterprise controls

Anthropic said EFS allows customers to store activity data in their own cloud environments, including Amazon S3, Azure Blob Storage, or Google Cloud Storage, using their own encryption keys and access controls.

“Customers want the ability to have their data live in infrastructure they control,” the company said. The company added that these features are optional and do not change model behavior or pricing.

Prakash said existing enterprise tools still play a role.

“Existing security architectures provide some oversight but generally lack the model-specific visibility needed to detect patterns of misuse across prompts, sessions and accounts,” he said.

Gogia said most enterprises already use gateways, data loss prevention tools, and SIEM platforms, and described EFS as “a provider-native sensor rather than a control plane.”

Emerging pattern, not yet a standard

Anthropic said it developed EFS with input from more than 100 organizations across sectors.

Gogia said the approach could become more widely adopted, but it is still early.

“EFS has a credible chance of becoming a procurement norm or a reference architecture for sensitive enterprise AI,” he said. “It is far too early to call it a standard.”

Prakash pointed out customer-controlled monitoring “could become an option for regulated enterprises,” but added that it will not replace private deployments.

“Regulated buyers should demand testable artefacts before they accept secure adjectives,” Gogia said, including “a published data-flow model, detector precision and recall, and immutable evidence with version provenance.” Anthropic said it will not charge for EFS, though customers will incur standard cloud infrastructure costs.