Claude Enterprise adds a security center, a parent admin console and new access controls

Anthropic added a security center to Organization settings in Claude Enterprise on 5 November 2026, in beta. It reviews Claude's security settings for chat, Cowork and Claude Code on one page and compares them with Anthropic's recommended configuration. Only Owners and the Primary Owner can open it.

For each control the page shows the current value, the recommended value where Anthropic gives one, and a link to the setting. Controls are grouped into Org, chats, Cowork and Claude Code. Each one is marked Passed, Failed, Not attested or Shown for information.

The page checks the configuration each time it loads, and it shows settings that are new or recently changed. It is view-only and changes nothing. With the Compliance API enabled, each status change sends an event to the organisation's activity feed.

Some controls sit outside Claude, such as device management settings, so the page can't read them. An Owner can attest to those instead. Attestations give admins a shared place to record the state of each control, accept a risk or flag an issue.

The report downloads as a PDF for auditors. It marks which controls were checked automatically and which an Owner confirmed. Organisations with the Compliance API can read the same results through it.

Anthropic calls the security center guidance, based on its own recommendations. Its article says the page is neither a certification nor a guarantee of compliance with any law or standard.

Our reading is that the report makes a good first page for an audit pack. An auditor can see which settings the system read and which rest on an Owner's word.

Our reading is that your auditors are likely to test the settings against your own policy and the standard you hold. Keep a map from each control on the page to that policy, so a Passed status becomes evidence for one line of it.

Only Owners can open the page. A security team without the Owner role will need an Owner to download the report, or will read the results through the Compliance API.

Centralized management gives a parent organisation its own admin console at admin.claude.com. It is in beta on Enterprise plans. The parent stands for your company and its commercial relationship with Anthropic, above your Claude Enterprise, Team and Claude Platform organisations.

The parent manages company-wide settings once. These are single sign-on (SSO), which lets staff sign in with their company login, and directory sync through SCIM. Directory sync provisions users and groups from your identity provider. Domain verification, groups and Compliance API keys move to the parent too.

Anthropic added two roles. Parent organization owners control everything at the parent, including who holds the owner and admin roles. Parent organization admins control everything except those role assignments.

The parent can hold more than one identity provider, each with its own SSO and SCIM. An acquired company can keep its provider, or move to a new one a domain at a time.

Each child organisation stays separate, with its own members, roles, projects and settings behind a hard data boundary. Each also keeps its own IP allowlisting, tenant restrictions and session length. A parent role gives no access inside an organisation by default, but Compliance API keys created at the parent can read data across organisations.

To start, an owner of any child organisation begins setup at admin.claude.com and proposes parent admins. Admins of the other organisations review and approve, and the last approval completes setup. If an organisation hasn't approved seven days after the last change, any child organisation owner can approve for it.

Anthropic says enrolment can't be undone, and it can complete as soon as someone starts setup. Its advice is to identify your parent organisation admins before you start.

A parent can't enrol if one of its child organisations has HIPAA enabled, or if its only child organisations are on Team plans. Claude Platform organisations bought through the AWS or Azure marketplace can't join. During the beta, new child organisations come only through your account team, and invoicing doesn't change. Until group availability ships, every Enterprise child admin can see every group under the parent.

Our reading is that this suits a group with several business units, or one that has bought companies. Sign-in and directory sync then sit in one place, and each unit's data stays in its own organisation.

Setup can't be undone, and it can finish as soon as someone starts it. The seven-day rule also lets any child organisation owner approve for an organisation that hasn't. Our reading is that the parent admins should be chosen and written down before anyone opens the console.

Compliance API keys created at the parent reach across every organisation. Treat them as the most sensitive keys in the set, and decide who may create them.

Until group availability ships, every Enterprise child admin can see every group under the parent. If group names carry client or project names, a business unit's admins will see the others' names.

Member sign-out, in beta, signs a member out with one Admin API call without removing them. It ends their sign-in sessions on every browser and device, and apps such as Claude Code must sign in again. It doesn't end API keys, cloud sessions already running or scheduled routines, and the member can sign in again.

Enterprise admins can set up and manage their own IP allowlist, in beta, without going through Anthropic. An IP allowlist is the list of network addresses, such as your offices or VPN, that are allowed to reach Claude. It covers most Claude products. Owners manage it on Enterprise, and Admins on Claude Platform.

Artifacts shared outside the organisation are not restricted by the allowlist. For API organisations in the Claude Console, the list takes up to 1,000 addresses or ranges, and requests to covered services from anywhere else are refused. Anthropic notes that this limits where a leaked API key can be used.

Admins can also set the most-used Claude Desktop settings without pushing them through device management, in beta. On managed computers they can block personal and other outside Claude accounts. Owners can limit which workspace folders people can use with Claude.

Inference hooks are now generally available. Prompt hooks send prompts from supported Claude products to your security tool for a check, and pre-tool-call hooks check tool calls before they run. Both work with security vendors, so your existing data loss prevention (DLP) and security policies can apply to Claude.

Our reading is that these close three gaps a security review raises before a wider rollout. They cut off a leaver's sessions, keep Claude on company networks and keep staff off personal Claude accounts on work laptops.

A leaver process that uses the sign-out call also has to revoke the person's API keys, stop cloud sessions already running, end scheduled routines and unlink their Cowork computers. It then disables the person in the identity provider, since automatic provisioning can add a removed member back at their next sign-in.

The allowlist doesn't limit artifacts shared outside the organisation. Decide whether staff may share them, and record that decision with the allowlist.

Our reading is that blocking outside accounts on managed computers keeps company work in the company's Claude organisation, where your own settings and audit records apply.

Endpoint detection and response (EDR) software watches company computers for suspicious activity and reports it to the security team. When Cowork and the unified Claude experience work locally, they run shell commands and code in an isolated virtual machine on the user's computer. The computer's existing EDR sensor can't see inside that virtual machine. In beta, admins can run the vendor's sensor inside it, and it reports to the same console as the rest of their computers.

Coverage is the virtual machine on the user's computer only. Cowork sessions in the cloud run on Anthropic's servers, where a sensor can't be installed.

Admins set it up through device management, the tool IT uses to push settings and software to company computers. On macOS that means MDM, and on Windows a registry policy through Group Policy or Intune. Each time the virtual machine starts, Claude checks the sensor package's signature from your EDR vendor before installing it.

Sensors run in detection mode, so they report activity to your security team and don't block it. Inside the virtual machine the sensor can see prompts, files and Claude's responses, and it sends that data to your EDR vendor. Anthropic doesn't receive or control it.

The beta covers macOS and Windows for Claude Team and Enterprise. It also covers the Claude desktop app on Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry, deployed through MDM for supported vendors. Anthropic doesn't charge extra during the beta, though the EDR vendor may charge for the sensor.

The supported vendors are CrowdStrike, Evoke Security, Neo, Palo Alto Networks, Pluto Security, SentinelOne and Tanium. Other vendors can apply through a form.

Our reading is that this covers the endpoint part of a security review. Commands run inside the virtual machine reach the same console as everything else on the computer. Because the sensor sends prompts and files to your EDR vendor, check that your contract with the vendor covers that data.

Detection mode means the sensor records and alerts, and a person decides what to do. Stopping an action before it runs has to come from other controls, such as pre-tool-call hooks or Cowork's own settings.

Cowork sessions in the cloud need a different control, because the sensor doesn't reach them. Later sensor updates also fall outside the signature check, which covers the package file only.