More visibility and control over Claude Enterprise for security teams

Anthropic released an enterprise governance update designed to provide security teams with administrative visibility across artificial intelligence workflows. According to Claude Blog's report, the release introduces a centralized security center, generally available inference hooks, multi-organization management, endpoint detection and response integration, and new access controls across Claude Enterprise and Claude Platform accounts. These tools allow corporate administrators to compare configurations against recommended baselines, apply data loss prevention rules, and monitor local sandboxes directly from existing enterprise consoles.

This report details the configuration tools, enforcement hooks, identity structures, and virtualization protections delivered in the update. The features are built for enterprise security architects, compliance officers, and workspace administrators managing Claude across chat, Cowork, Claude Code, and desktop environments. By explaining how the security center audits settings, how inference hooks inspect model traffic, and how parent organizations coordinate access, this guide provides technical operators with the facts needed to secure and administer Claude within corporate networks.

More visibility and control over Claude: what actually changed

Anthropic introduced a dedicated security center within Organization settings that compares configurations against recommended baselines across chat, Cowork, and Claude Code. Administrators can view current values, recommended targets, and direct links to modify settings in one place. The security center evaluates configurations whenever the page loads, and organizations using the Compliance API can query identical results programmatically. Security teams can also conduct attestations to record control states, accept risks, and export reports for auditors that show which checks were automated and which were confirmed by an Owner.

Beyond configuration auditing, Anthropic moved inference hooks into general availability and launched new identity and endpoint restrictions. Inference hooks connect Claude to corporate data security tools, inspecting prompts and tool invocations before execution. Centralized management introduces an admin console at admin.claude.com to manage settings across Claude Enterprise and Claude Platform organizations. Administrators also gained self-serve IP allowlists, single-call session sign-out via the Admin API, and desktop controls that block outside accounts on managed computers. Additionally, an endpoint detection and response integration places telemetry sensors inside Cowork virtual machines.

More visibility and control over Claude: how it works

Inference hooks operate at two inspection junctures during user interactions to enforce enterprise data loss prevention policies without requiring rewritten rules. Prompt hooks route incoming prompts from supported Claude products to external security tools for evaluation before the model processes text. Pre-tool-call hooks inspect tool calls within supported products before commands run, ensuring that automated actions comply with corporate standards. Because these hooks interface directly with existing security vendors, enterprise security teams apply their pre-existing policies directly to Claude interactions, allowing vendor tools to enforce rules before prompts or tool calls continue.

Administrative controls and endpoint protections operate through dedicated APIs and local sandbox isolation. The Admin API allows security teams to sign out a user across all browsers, devices, and applications like Claude Code with one call, without removing the account or stopping API keys, background routines, or cloud sessions. On workstations, Cowork and unified Claude execute code in an isolated local virtual machine on macOS and Windows. Administrators deploy endpoint detection and response sensors into the virtual machine via device management. Claude verifies the vendor signature before installation, and sensors run in detection mode without blocking activity.

More visibility and control over Claude: why it matters

Enterprise adoption of artificial intelligence tools often creates compliance hurdles and governance blind spots across corporate infrastructure. Without consolidated dashboards, security administrators must inspect settings across disparate menus without knowing whether configurations meet vendor recommendations. The new security center eliminates manual auditing by displaying live comparisons, flagging recently modified controls, and producing verifiable attestation reports for compliance stakeholders. While Anthropic notes that the security center offers guidance rather than a legal compliance certification, the automated checks and attestation exports simplify formal security reviews by providing documented records of accepted risks and verified technical controls.

Real-time inspection and endpoint instrumentation address data protection and visibility risks on managed laptops. Unmonitored prompt submissions and tool invocations risk exposing proprietary information, but general availability for prompt and pre-tool-call hooks lets existing data loss prevention platforms intercept sensitive data before execution. On workstations, executing local code inside an isolated virtual machine protects the host operating system, but historically hid process activity from corporate monitoring consoles. Running endpoint detection and response sensors inside the virtual machine eliminates that visibility gap, providing security operations centers with continuous telemetry matching their host-level monitoring.

More visibility and control over Claude: who is affected

The governance updates apply primarily to Claude Enterprise organizations, with multi-organization management introducing new operational tiers. Companies operating across multiple Claude Enterprise and Claude Platform accounts can establish a parent organization at admin.claude.com. Parent organization owners manage all parent-level configurations, including assigning owner and admin roles, while parent organization admins manage company-wide policies without modifying role assignments. Companies can attach multiple identity providers with distinct single sign-on and SCIM directory sync setups. Subsidiary organizations maintain their own members, session duration rules, and tenant restrictions, while Compliance API keys created at the parent level read data across all child organizations.

Individual employees, developers, and endpoint managers experience specific operational changes across their Claude tools. Device administrators can set common Claude Desktop preferences from Organization settings without pushing updates through mobile device management software, while owners can restrict accessible workspace folders and block personal accounts on corporate machines. Developers using Claude Code will need to re-authenticate if an admin executes an API sign-out call. Additionally, the endpoint detection and response beta extends beyond Claude Enterprise to Claude Team accounts, as well as desktop deployments on Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry for supported vendors on macOS and Windows.

More visibility and control over Claude: what to watch

Security teams planning implementation must account for distinct feature availability states and setup prerequisites. The security center, parent organization management, Admin API sign-out, self-serve IP allowlists, and desktop account controls are currently in beta, while inference hooks are generally available. Setting up centralized management requires sequential approvals: an admin from any organization under the parent proposes parent admins at admin.claude.com, and admins across every other organization must review and approve the roster. Because the final approval permanently finalizes parent organization creation and cannot be undone, organizations must designate their administrative hierarchy before initiating the approval workflow.

Organizations must also track technical boundaries across API services, background workloads, and virtualization telemetry. Revoking sessions via the Admin API disconnects interactive browsers and applications like Claude Code, but does not terminate developer API keys, ongoing cloud sessions, or scheduled routines. Furthermore, the endpoint detection and response integration only instruments the local virtual machine on macOS and Windows endpoints, leaving cloud-hosted Cowork sessions outside its sensor scope. Because sensors run strictly in detection mode rather than blocking operations, security operations teams must ensure external consoles actively monitor incoming telemetry to identify and investigate potential anomalies reported from inside the virtual machine.