Claude Enterprise adds security controls, with rollout limits
Anthropic announced new Claude Enterprise security and administration controls on November 5, covering configuration reviews, policy checks and companywide account management. For IT teams, the useful change is a clearer set of places to inspect access, enforce rules and respond when something goes wrong. Claude release notes.
Availability varies. Prompt and pre-tool-call inference hooks are generally available. The security center, centralized management, member sign-out through the Admin API and self-service IP allowlists remain in beta. Anthropic also highlighted its recently announced endpoint detection and response integration, which has a narrower scope than the broader Enterprise controls. Anthropic’s announcement.
| Control | Status | Main boundary |
|---|---|---|
| Inference hooks | Generally available | Governed Enterprise requests |
| Security center | Beta | View-only review for Owners |
| Centralized management | Beta | Enrollment cannot be undone |
| Member sign-out API | Beta | Sessions end but membership remains |
| IP allowlisting | Public beta | External artifacts are excluded |
| EDR integration | Beta | Local VM detection only |
The practical question is which layer each feature covers. A dashboard can reveal configuration gaps, a hook can reject a request, and an endpoint sensor can report activity. Teams should map those jobs separately before deciding that a rollout is ready.
Security center makes settings easier to review
The security center compares Claude settings with Anthropic’s recommendations across chat, Cowork and Claude Code. It shows current values, recommendations where available, and links to the settings that need attention. However, the page is view-only. Owners and the Primary Owner can open it; ordinary Admins and users cannot. Security center documentation.
Some controls are checked automatically. Others, including settings managed outside Claude, need an Owner’s attestation. The downloadable report distinguishes those two sources of evidence. Neither that report nor a passing control is a security certification or compliance guarantee. Security center documentation.
A useful review should therefore record who checked each external control and what evidence they used. Otherwise, a completed checklist can hide unanswered questions about the actual device or network configuration.
Inference hooks need explicit coverage and failure choices
Inference hooks connect Enterprise conversations to a security service operated by the organization or its vendor. That service returns an allow or deny decision before a governed request proceeds. The approach supports policy enforcement without installing a separate hook client on every employee device. Inference hooks overview.
The technical documentation draws important boundaries. Hooks do not cover Claude Platform API organizations and are unavailable on Amazon Bedrock or Google Cloud. Attachments contribute extracted text rather than raw file or image bytes, so image-only content is not inspected. The verdict cannot rewrite or redact a prompt. Inference hooks technical documentation.
Coverage also has exceptions within Enterprise products, including some internal tool calls and model calls behind hosted Claude Security scans, Code Review and smart reports. Voice has additional tool-call gaps. If the security endpoint fails, the configured policy determines whether requests are blocked or proceed without inspection. Inference hooks technical documentation.
For a pilot, test an allowed request, a blocked request and an unavailable policy server. Record what employees see and how the security team receives the event. That is more informative than checking whether the feature is enabled.
Centralized management changes who controls identity
Centralized management brings shared identity settings into a parent console while child organizations retain their own users, roles and projects. Parent owners can assign parent owners and admins; parent admins cannot make those role assignments. Neither role automatically grants membership inside a child organization. Centralized management documentation.
Parent-level API keys deserve separate review because they work across the organizations beneath that parent, subject to their scopes. Keys can remain active after their creator leaves. This is a broader boundary than an administrator’s access to one organization. Centralized management documentation.
Enrollment cannot be undone and can finish immediately when the initiating owner administers every child organization. During beta, parents with a HIPAA-enabled child or only Team-plan children cannot enroll. Claude Platform organizations purchased through AWS or Azure Marketplace cannot join a parent. Centralized management documentation.
Treat enrollment as an identity migration with named owners and an agreed access model. A reversible trial of a dashboard and an irreversible change to account administration need different preparation.
Signing someone out does not finish incident response
The member sign-out feature ends sessions across devices and revokes trusted devices without deleting chats, projects or membership. Its effect also reaches other organizations and personal plans used by that account. The person can sign in again immediately unless access is separately removed. Scheduled Claude Code routines continue, and connected Cowork computers are not unlinked. Member management documentation.
Anthropic’s announcement also says the Admin API action does not terminate API keys or already running cloud sessions. An incident checklist should distinguish reauthentication from removing access and stopping work. Anthropic’s announcement.
IP allowlisting adds another boundary. Enterprise Owners and Claude Platform Admins can manage approved addresses themselves during the public beta. Externally shared artifacts are not restricted by the organization’s allowlist, so publication and sharing policies still need their own review. IP allowlisting documentation.
EDR visibility stops at the local virtual machine
The EDR integration runs an approved vendor’s sensor inside Cowork’s local virtual machine. It is in beta for Team and Enterprise, including supported Claude Desktop deployments through Bedrock, Vertex AI and Microsoft Foundry on macOS and Windows. Cloud Cowork sessions are outside its coverage. EDR integration documentation.
Anthropic describes the sensors as detection-only: they report activity rather than blocking it. Anthropic’s announcement.
The setup requires supported desktop versions and device policy deployment. Sensor data can include prompts, files and Claude responses, and telemetry can leave the virtual machine outside Cowork’s network allowlist. Review the EDR vendor’s handling of that data alongside its deployment requirements. EDR integration documentation.
Start with the boundaries your team actually uses
ByteForward has not independently tested these controls. The sensible next step is to inventory where employees use Claude, assign an owner to each policy and test the response to failure. Evaluate the resulting evidence rather than the number of switches available.
For the separate activation and access questions around unified chat and Cowork, see ByteForward’s guide to the Claude Enterprise rollout. It provides context for deciding which workflows these security controls need to cover.
