Anthropic announced new security and administration controls for Claude Enterprise on November 5. Businesses gain a center for reviewing settings, checks on requests sent to AI and a shared identity console. Inference hooks are generally available to Enterprise. The security center, centralized management and several access controls are in beta.
What the security team sees in the new center
The center groups controls for the organization, chats, Claude Cowork and Claude Code. Each control displays its current value, Anthropic's recommendation when available, and where the setting can be changed. The result distinguishes settings that match the recommendation, those that differ, controls that need manual confirmation and information-only items. Anthropic reserves the page for Owner and Primary Owner roles. The center is a review interface, with changes made on the pages for the individual settings.
For a policy enforced by a device management system, an owner can manually record the verified state and their observations. The downloadable report distinguishes automated checks from those confirmed by an Owner. Organizations using the Compliance API can retrieve the same results programmatically for their security tools.
Checking requests before execution
Inference hooks connect Claude's workflow to a checking service run by the organization. For covered requests, Claude asks for a verdict before generating a response. Tool calls are checked before the tool runs. The company's service applies its own rules and returns an allow or deny decision. Configuration takes place at the organization level and can cover chats, Cowork and Claude Code.
The material sent for inspection can include the conversation and text extracted from attachments. For attachments, metadata and extracted text are sent, while the documentation excludes raw file and image bytes. An image-only screenshot is outside this inspection mechanism. Claude Platform has a separate access scope, and the documentation places direct calls from Platform organizations outside the coverage of these Enterprise checks.
The organization also chooses how to handle a failure of its checking service. The setting determines whether a request is blocked or proceeds without inspection when the server fails to respond in time. The documentation specifies a default verdict timeout of five seconds. In observation mode, the service returns verdicts and requests continue. Other options let the organization inspect a selected percentage of requests and exclude particular custom roles.
Where the new controls apply
- Requests to AIChecks apply company policy to covered prompts and tool calls.
- Identity and accessThe shared console groups domains, directories and sign-in.
- The deviceDesktop policies control access. The Cowork sensor monitors the local virtual machine.
One console for multiple organizations
Centralized management brings common settings into the parent organization's console. It can manage single sign-on, directory synchronization through SCIM, verified domains, groups and Compliance API keys scoped to the parent. Associated organizations continue to manage their own members, roles and controls, such as session duration or approved IP addresses. Enrollment in the beta requires the associated organizations to approve the proposed roles for parent-level administration.
The console can associate multiple identity providers with the same domain. Each can have its own SSO configuration and directory. For users offered several providers, the sign-in page displays a choice after they enter their email address. Associated organizations use SCIM synchronization or invitation-only account provisioning. A domain with Microsoft Login enabled must turn off that option before adding a second provider.
Signing a member out and controlling network access
Sign-out through the Admin API is available in beta for Claude Enterprise. It ends the member's sign-in sessions and revokes their trusted devices. The account retains its role, chats and projects, and the person can sign in again. Claude Code and Claude Desktop require a new sign-in on the next use. The API call requires a key with the write:members permission. API keys, tasks already running in the cloud and scheduled routines retain their state.
The Admin API documentation also describes how the operation is recorded. A completed sign-out appears in the audit log export. Organizations with the Compliance API can also track it in the activity feed, with counts of the affected sessions and devices. If the operation stops before completion, the activity feed shows the incomplete state. This distinguishes the request from the actual revocation result.
The list of authorized IP addresses can now be managed directly in beta. Owner and Primary Owner roles manage it on Enterprise. Administrators manage it on Claude Platform. The list accepts up to 1,000 public addresses or ranges in IPv4 and IPv6. The policy applies to covered services for that organization. Artifacts users have shared outside the organization remain accessible according to their sharing settings.
Policies for Claude Desktop and Cowork
In Claude Desktop, owners can configure certain settings for their teams through the organization account, including workspace folder access and the startup of user-added local MCP servers. A policy on the computer takes precedence for the same option. Sign-in on managed devices can be restricted to approved organizations. The read-only setting applies to Claude's file tools. Commands and MCP servers in the Code tab can modify those files and access other folders.
The announcement also describes endpoint detection and response integration for Cowork. An approved vendor's sensor runs in the application's local virtual machine and reports activity to the company's security console. The application verifies the package signature when the virtual machine starts. The integration is in beta on macOS and Windows for Team and Enterprise plans. Its observation scope is the virtual environment on the computer, and the sensor provider receives the telemetry.
